Search

Background: Lymph node (LN) involvement is a strong indicator of poor prognosis for breast cancer (BC), with adjuvant chemotherapy remaining fundamental to management of these patients. SWOG S8814 was a Phase III randomized trial of postmenopausal patients with pathologic LN-positive BC who were hormone receptor positive (HR+). The objectives of the clinical trial were to compare disease free survival (DFS) and overall survival (OS) of 1) these postoperative patients treated with a combination of cyclophosphamide, doxorubicin, fluorouracil (CAF) plus tamoxifen versus tamoxifen alone; and 2) patients treated with CAF followed by tamoxifen versus CAF plus concurrent tamoxifen. In this study we sought to evaluate the potential of applying computational image analysis on whole slide images (WSI) for predicting DFS and OS in SWOG S8814. Methods: A cohort of 135 patients (N=53 DFS event) diagnosed with HR+ & LN+ BC from clinical trial ECOG 2197 was utilized as training set D1. Validation set D2 comprised 630 patients (N=260 DFS event, N=195 death) with HR+& LN+ BC from SWOG S8814. Three deep learning models were employed to respectively detect nuclei, mitosis, and tubules in WSIs. Subsequently, a total of 1,810 features relating to nuclear morphology (e.g., spatial distribution, shape, texture, orientation), mitotic activity (e.g., mitosis hotspot, mitotic rates) and tubule formation (e.g., tubular nuclei distribution, ratio of tubule to non-tubule area) were extracted from each WSI. A lasso regularized Cox regression model (IbRiS) was trained on D1 to respectively identify four features from each of the feature categories (nuclei morphology, mitotic activity, and tubule formation) most strongly associated with DFS, a continuous risk score based on the selected features was then constructed. An optimal risk threshold was identified on D1 to dichotomize the risk scores into high vs. low risk of recurrence categories. Blinded validation of the machine learning model on SWOG S8814 using Cox regression was performed by SWOG to evaluate its performance in terms of DFS and OS. Results: In D2, patients identified as high risk of recurrence by IbRiS had a significantly worse prognosis in terms of DFS with hazard ratio=1.30 (p=0.039, 95% CI=1.01-1.66). IbRiS was also found to be significantly prognostic of OS with hazard ratio=1.38 (p=0.026, 95% CI=1.04-1.83). IbRiS was however, neither prognostic of DFS (HR = 1.20; 95% CI 0.93-1.54) nor OS (HR = 1.28; 95% CI 0.96-1.71) in multivariable analysis adjusting for treatment, tumor size, and number of positive nodes. IbRiS was also not a significant predictor of chemotherapy benefit (DFS p=0.45; OS p=0.25). Conclusion: We developed a prognostic model (IbRiS) based on the combined features of nuclear morphology, mitosis count, and tubule formation that can help further risk stratify HR+ & LN+ BC patients by only using H&E slides. Citation Format: Yuli Chen, William E. Barlow, Haojia Li, Cheng Lu, Andrew Janowczyk, German Corredor, Shridar Ganesan, Michael Feldman, Pingfu Fu, Hannah Gilmore, Kathy S. Albain, Lajos Pusztai, James Rae, Daniel Hayes, Andrew K. Godwin, Alastair M. Thompson, Anant Madabhushi. Computerized Measurements of Nuclear Morphology Features, Mitosis Rate, and Tubule Formation from H&E Images Predicts Disease-Free Survival in Patients with HR+ & LN+ Invasive Breast Cancer from SWOG S8814 [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P2-11-16.

Background: Adjuvant chemotherapy is a common treatment for breast cancer patients with aggressive tumors, but imposes severe side effects. Although lymph node involvement is associated with higher likelihood of mortality, a subset of LN+ patients can survive without relapse, even without adjuvant chemotherapy. SWOG-8814 (S8814) was a randomized clinical trial of hormone receptor positive, LN+ post-menopausal women (pathologic stage T1-3N1-2) to compare patient outcome between endocrine therapy alone (tamoxifen), chemotherapy (cyclophosphamide, doxorubicin, and 5-fluorouracil) followed by 5 years of tamoxifen, and chemotherapy with concurrent tamoxifen. In this work, we evaluated the ability of a machine learning approach called multiple instance learning (MIL) to predict long-term overall (OS) and disease-free survival (DFS) in patients from the S8814 clinical trial via analysis of digitized H&E tissue slides. Methods: The training set (St, n=121) consisted of digitized H&E Whole Slide Images (WSIs) of ER+ LN+ patients from ECOG-2197 - a clinical trial to compare patient outcome under chemotherapy with doxorubicin/docetaxel vs. doxorubicin/cyclophosphamide. For each WSI, the nuclei of ten tumor patches (2000 × 2000 pixels) were automatically segmented by a pretrained machine learning algorithm called Convolutional Neural Networks (CNN). The results of the CNN’s segmentations were used to obtain ten corresponding feature vectors of 3963 features relating to nuclear morphology and spatial arrangement. Five top discriminative features (Haralick texture, Delaunay triangulation, orientation entropy, perimeter kurtosis, architecture) were identified by reliefF-MI feature filtering and forward selection on St. A Normalized Set Kernel Support Vector Machine classifier was trained on St to construct a MIL binary classifier (M+), which generates a continuous survival score to predict which patients would have >10-years of DFS and OS. The score cutoff thatmaximized F1 score on St was used to generate binary predictions of patient outcome. Blinded validation of the risk scores on S8814 was performed by SWOG. Results: M+ was significantly prognostic of OS in univariate analysis (HR=0.72, p=0.020, 95% CI= 0.55 – 0.95) on S8814. M+ stratifications remained statistically significant after multivariable analysis controlling for treatment, number of positive lymph nodes, and tumor size (HR = 0.71, p=0.017, 95% CI=0.54-0.94). M+ predictions did not achieve statistical significance for DFS in univariate analysis (HR=0.81, p=0.081, 95% CI 0.64-1.03) nor multivariable analysis (HR=0.81, p=0.080, 95% CI 0.63-1.03). M+ was not predictive of chemotherapy benefit for OS (p=0.88) and DFS (p=0.65). There was no correlation of the generated continuous score with Recurrence Score (r=-0.05), previously shown to be a predictive factor for chemotherapy benefit. Conclusion: Our findings demonstrate that nuclear morphology provides insight into overall survival in ER+ LN+ breast cancer patients. Future work will involve combining the MIL based morphologic predictor with other clinicopathologic factors to improve DFS risk stratification. Funding: Funding NIH/NCI grants U10CA180888, U10CA180819, U24CA196175; and in part by Genomic Health, INC. (now Exact Sciences Corp.) Citation Format: Daniel Shao, William E. Barlow, Haojia Li, Cheng Lu, Kathy S. Albain, James Rae, Daniel F. Hayes, Andrew K. Godwin, Alastair M. Thompson, Anant Madabhushi, Lajos Pusztai. Computer analysis of nuclear morphology with Multiple Instance Learning Predicts Overall Survival for Node Positive Breast Cancer Patients from SWOG S8814: A Blinded Validation Study [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P2-11-11.

Supplementary methods, notes, figures, tables, video captions, and references

Supplemental Tables

Supplemental Figures

No predictive biomarkers can robustly identify patients with non–small cell lung cancer (NSCLC) who will benefit from immune checkpoint inhibitor (ICI) therapies. Here, in a machine learning setting, we compared changes (“delta”) in the radiomic texture (DelRADx) of CT patterns both within and outside tumor nodules before and after two to three cycles of ICI therapy. We found that DelRADx patterns could predict response to ICI therapy and overall survival (OS) for patients with NSCLC. We retrospectively analyzed data acquired from 139 patients with NSCLC at two institutions, who were divided into a discovery set (D1 = 50) and two independent validation sets (D2 = 62, D3 = 27). Intranodular and perinodular texture descriptors were extracted, and the relative differences were computed. A linear discriminant analysis (LDA) classifier was trained with 8 DelRADx features to predict RECIST-derived response. Association of delta-radiomic risk score (DRS) with OS was determined. The association of DelRADx features with tumor-infiltrating lymphocyte (TIL) density on the diagnostic biopsies (n = 36) was also evaluated. The LDA classifier yielded an AUC of 0.88 ± 0.08 in distinguishing responders from nonresponders in D1, and 0.85 and 0.81 in D2 and D3. DRS was associated with OS [HR: 1.64; 95% confidence interval (CI), 1.22–2.21; P = 0.0011; C-index = 0.72). Peritumoral Gabor features were associated with the density of TILs on diagnostic biopsy samples. Our results show that DelRADx could be used to identify early functional responses in patients with NSCLC.

Supplementary Data from Novel Radiomic Measurements of Tumor-Associated Vasculature Morphology on Clinical Imaging as a Biomarker of Treatment Response in Multiple Cancers

Purpose:The tumor-associated vasculature (TAV) differs from healthy blood vessels by its convolutedness, leakiness, and chaotic architecture, and these attributes facilitate the creation of a treatment-resistant tumor microenvironment. Measurable differences in these attributes might also help stratify patients by likely benefit of systemic therapy (e.g., chemotherapy). In this work, we present a new category of computational image-based biomarkers called quantitative tumor-associated vasculature (QuanTAV) features, and demonstrate their ability to predict response and survival across multiple cancer types, imaging modalities, and treatment regimens involving chemotherapy.Experimental Design:We isolated tumor vasculature and extracted mathematical measurements of twistedness and organization from routine pretreatment radiology (CT or contrast-enhanced MRI) of a total of 558 patients, who received one of four first-line chemotherapy-based therapeutic intervention strategies for breast (n = 371) or non–small cell lung cancer (NSCLC, n = 187).Results:Across four chemotherapy-based treatment strategies, classifiers of QuanTAV measurements significantly (P < 0.05) predicted response in held out testing cohorts alone (AUC = 0.63–0.71) and increased AUC by 0.06–0.12 when added to models of significant clinical variables alone. Similarly, we derived QuanTAV risk scores that were prognostic of recurrence-free survival in treatment cohorts who received surgery following chemotherapy for breast cancer [P = 0.0022; HR = 1.25; 95% confidence interval (CI), 1.08–1.44; concordance index (C-index) = 0.66] and chemoradiation for NSCLC (P = 0.039; HR = 1.28; 95% CI, 1.01–1.62; C-index = 0.66). From vessel-based risk scores, we further derived categorical QuanTAV high/low risk groups that were independently prognostic among all treatment groups, including patients with NSCLC who received chemotherapy only (P = 0.034; HR = 2.29; 95% CI, 1.07–4.94; C-index = 0.62). QuanTAV response and risk scores were independent of clinicopathologic risk factors and matched or exceeded models of clinical variables including posttreatment response.Conclusions:Across these domains, we observed an association of vascular morphology on CT and MRI—as captured by metrics of vessel curvature, torsion, and organizational heterogeneity—and treatment outcome. Our findings suggest the potential of shape and structure of the TAV in developing prognostic and predictive biomarkers for multiple cancers and different treatment strategies.

Purpose:Between 30%–40% of patients with prostate cancer experience disease recurrence following radical prostatectomy. Existing clinical models for recurrence risk prediction do not account for population-based variation in the tumor phenotype, despite recent evidence suggesting the presence of a unique, more aggressive prostate cancer phenotype in African American (AA) patients. We investigated the capacity of digitally measured, population-specific phenotypes of the intratumoral stroma to create improved models for prediction of recurrence following radical prostatectomy.Experimental Design:This study included 334 radical prostatectomy patients subdivided into training (VT, n = 127), validation 1 (V1, n = 62), and validation 2 (V2, n = 145). Hematoxylin and eosin–stained slides from resected prostates were digitized, and 242 quantitative descriptors of the intratumoral stroma were calculated using a computational algorithm. Machine learning and elastic net Cox regression models were constructed using VT to predict biochemical recurrence-free survival based on these features. Performance of these models was assessed using V1 and V2, both overall and in population-specific cohorts.Results:An AA-specific, automated stromal signature, AAstro, was prognostic of recurrence risk in both independent validation datasets [V1,AA: AUC = 0.87, HR = 4.71 (95% confidence interval (CI), 1.65–13.4), P = 0.003; V2,AA: AUC = 0.77, HR = 5.7 (95% CI, 1.48–21.90), P = 0.01]. AAstro outperformed clinical standard Kattan and CAPRA-S nomograms, and the underlying stromal descriptors were strongly associated with IHC measurements of specific tumor biomarker expression levels.Conclusions:Our results suggest that considering population-specific information and stromal morphology has the potential to substantially improve accuracy of prognosis and risk stratification in AA patients with prostate cancer.

Supplementary Sheet

Labels of supplementary figures

Figure S3. Boxplots corresponding to the top SpaTIL features. (a) D1, (b) D2, (c) D3, and (d) D4.

Table S1. List of all SpaTIL features. Table S2. List of the most discriminating spaTIL features for predicting recurrence in NSCLC. In this study, density is defined as the ratio of the number of cells within the cluster to the cluster pixel area, betweenness centrality is a measure of centrality in a graph based on shortest paths, and closeness centrality is a measure of centrality calculated as the sum of the length of the shortest paths between the node and all other nodes in the graph. Table S3: Summary of clinical and pathological features of the studied NSCLC cohorts. Table S4: Multivariable survival analysis on the test sets including SpaTIL. Table S5. Treatment information for the datasets included in this study.

Figure S1. Patient selection workflow for the datasets included in this study.

Description of a use case to illistrate workflow with Sedeen Viewer

Pathology Image Informatics Platform (PIIP) is an NCI/NIH sponsored project intended for managing, annotating, sharing, and quantitatively analyzing digital pathology imaging data. It expands on an existing, freely available pathology image viewer, Sedeen. The goal of this project is to develop and embed some commonly used image analysis applications into the Sedeen viewer to create a freely available resource for the digital pathology and cancer research communities. Thus far, new plugins have been developed and incorporated into the platform for out of focus detection, region of interest transformation, and IHC slide analysis. Our biomarker quantification and nuclear segmentation algorithms, written in MATLAB, have also been integrated into the viewer. This article describes the viewing software and the mechanism to extend functionality by plugins, brief descriptions of which are provided as examples, to guide users who want to use this platform. PIIP project materials, including a video describing its usage and applications, and links for the Sedeen Viewer, plug-ins, and user manuals are freely available through the project web page: http://pathiip.org. Cancer Res; 77(21); e83–86. ©2017 AACR.

Supplementary Document

Purpose:To (i) create a survival risk score using radiomic features from the tumor habitat on routine MRI to predict progression-free survival (PFS) in glioblastoma and (ii) obtain a biological basis for these prognostic radiomic features, by studying their radiogenomic associations with molecular signaling pathways.Experimental Design:Two hundred three patients with pretreatment Gd-T1w, T2w, T2w-FLAIR MRI were obtained from 3 cohorts: The Cancer Imaging Archive (TCIA; n = 130), Ivy GAP (n = 32), and Cleveland Clinic (n = 41). Gene-expression profiles of corresponding patients were obtained for TCIA cohort. For every study, following expert segmentation of tumor subcompartments (necrotic core, enhancing tumor, peritumoral edema), 936 3D radiomic features were extracted from each subcompartment across all MRI protocols. Using Cox regression model, radiomic risk score (RRS) was developed for every protocol to predict PFS on the training cohort (n = 130) and evaluated on the holdout cohort (n = 73). Further, Gene Ontology and single-sample gene set enrichment analysis were used to identify specific molecular signaling pathway networks associated with RRS features.Results:Twenty-five radiomic features from the tumor habitat yielded the RRS. A combination of RRS with clinical (age and gender) and molecular features (MGMT and IDH status) resulted in a concordance index of 0.81 (P < 0.0001) on training and 0.84 (P = 0.03) on the test set. Radiogenomic analysis revealed associations of RRS features with signaling pathways for cell differentiation, cell adhesion, and angiogenesis, which contribute to chemoresistance in GBM.Conclusions:Our findings suggest that prognostic radiomic features from routine Gd-T1w MRI may also be significantly associated with key biological processes that affect response to chemotherapy in GBM.

Supplementary Information

Prostate cancer treatment planning is largely dependent upon examination of core-needle biopsies. The microscopic architecture of the prostate glands forms the basis for prognostic grading by pathologists. Interpretation of these convoluted three-dimensional (3D) glandular structures via visual inspection of a limited number of two-dimensional (2D) histology sections is often unreliable, which contributes to the under- and overtreatment of patients. To improve risk assessment and treatment decisions, we have developed a workflow for nondestructive 3D pathology and computational analysis of whole prostate biopsies labeled with a rapid and inexpensive fluorescent analogue of standard hematoxylin and eosin (H&E) staining. This analysis is based on interpretable glandular features and is facilitated by the development of image translation–assisted segmentation in 3D (ITAS3D). ITAS3D is a generalizable deep learning–based strategy that enables tissue microstructures to be volumetrically segmented in an annotation-free and objective (biomarker-based) manner without requiring immunolabeling. As a preliminary demonstration of the translational value of a computational 3D versus a computational 2D pathology approach, we imaged 300 ex vivo biopsies extracted from 50 archived radical prostatectomy specimens, of which, 118 biopsies contained cancer. The 3D glandular features in cancer biopsies were superior to corresponding 2D features for risk stratification of patients with low- to intermediate-risk prostate cancer based on their clinical biochemical recurrence outcomes. The results of this study support the use of computational 3D pathology for guiding the clinical management of prostate cancer.Significance:An end-to-end pipeline for deep learning–assisted computational 3D histology analysis of whole prostate biopsies shows that nondestructive 3D pathology has the potential to enable superior prognostic stratification of patients with prostate cancer.

Background: CDK 4/6 inhibitors (CDKI) with Endocrine therapy (ET) is mainstay treatment for hormone receptor-positive (HR+) Her2- metastatic breast cancer (MBC). Despite excellent efficacy, most patients develop resistance to CDKI limiting its utility. CDK4/6 are thought to act by upregulating VEGF causing tortuous angiogenesis and promoting cancer progression. Recent studies have shown aggressive tumors possess a higher density and tortuous vasculature. In this study, we evaluated if vascular radiomic features computationally extracted from liver CT scans pre- and post CDKI treatment could predict patient survival and treatment response. Method: From a registry of 350 patients on treatment with CDKI at institution 1(S1), 51 pts with HR+, Her2-, MBC patients with evidence of liver mets and disease progression (PFS) data were identified. 30 pts discontinued treatment due to progression or death, with a median time to progression of 195 days. Pre-treatment and first post-treatment CT exams were analyzed from 25 and 34 patients, respectively. Median time between scans was 128 days. To validate the prognostic value of our signature, a cohort of 29 patients with available OS data was identified from institution 2 (S2). A publicly available pre-trained deep learning model was applied to isolate liver metastases and vessels. Next, fast marching algorithm was applied to reduce vessels to their centerlines and divide the vasculature into constituent branches. 7 quantitative metrics were computed measuring vascularity of metastases and 3-D shape of hepatic vessels. First, the number (f1) and percentage (f2) of hepatic vessels arising from the tumor were computed. Vessel tortuosity - measuring the degree of twisting across a vessel - was computed separately for each branch. The mean (f3), standard deviation (f4), maximum (f5), skewness (f6), and kurtosis (f7) tortuosity values were calculated to summarize these measurements at patient level. The features were individually assessed at pre- and post-treatment for association with PFS at S1 in univariable Cox proportional hazards models. Features found to be associated in S1 were evaluated for association with OS in S2. Results: On the initial post-treatment scan, features of both tumor vascularization (f6 - HR=1.115 [1.039-1.196]; f7 - HR=10.646 [2.539-44.641]), as well as two features of vessel tortuosity (f3 - HR=0.011 [0.001-0.199]; f4 - HR=0.545 [0.313-0.949]) were significantly associated with PFS. Both tortuosity features were also significantly associated with OS in S2 (f3 - HR=0.085 [0.009-0.780]; f4 - HR=0.331 [0.130-0.842]). In addition the percentage (f7 - HR=6.445 [2.001-20.753]]),of vessels feeding the lesions was also significant in S2 while the number (f6 - HR=1.070 [0.966-1.184]), of vessels was not. No vessel metrics from the pre-treatment baseline exam were significantly associated with OS. Conclusions: Radiomic analysis of tumor vascularity and vessel tortuosity on CT scans post-CDK treatment was associated with patient survival and treatment response. Table 1.Association of vessel features with PFS in Institution 1(S1)and OS in Institution 2(S2)S1- Pretreatment CT (n=25)S1 - Post-Treatment CT (n=34)S2 - Post-treatment CT - Validation (n=29)f1Tortuosity - Mean0.9912702470.852294696–f2Tortuosity - St. Dev0.7563942230.273115373–f3Tortuosity - Max0.1387051870.0022517550.02931473f4Tortuosity - Skewness0.254463710.0318270370.02022882f5Tortuosity - Kurtosis0.3058266760.108967601–f6Number of vessels feeding lesions0.419096650.0025747910.19306052f7Percentage of vessels feeding lesions0.0517942660.0012205950.00179103 Citation Format: Vidya Sankar Viswanathan, Nathaniel Braman, Priyanka Reddy, Siddharth Kunte, Jame Abraham, Alberto J Montero, Anant Madabhushi. Post-treatment vascularity and vessel shape are associated with survival and response to CDK4/6 inhibitors in hormone receptor-positive metastatic breast cancer (MBC) [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P5-13-27.

Prostate cancer treatment planning is largely dependent upon examination of core-needle biopsies. The microscopic architecture of the prostate glands forms the basis for prognostic grading by pathologists. Interpretation of these convoluted three-dimensional (3D) glandular structures via visual inspection of a limited number of two-dimensional (2D) histology sections is often unreliable, which contributes to the under- and overtreatment of patients. To improve risk assessment and treatment decisions, we have developed a workflow for nondestructive 3D pathology and computational analysis of whole prostate biopsies labeled with a rapid and inexpensive fluorescent analogue of standard hematoxylin and eosin (H&E) staining. This analysis is based on interpretable glandular features and is facilitated by the development of image translation–assisted segmentation in 3D (ITAS3D). ITAS3D is a generalizable deep learning–based strategy that enables tissue microstructures to be volumetrically segmented in an annotation-free and objective (biomarker-based) manner without requiring immunolabeling. As a preliminary demonstration of the translational value of a computational 3D versus a computational 2D pathology approach, we imaged 300 ex vivo biopsies extracted from 50 archived radical prostatectomy specimens, of which, 118 biopsies contained cancer. The 3D glandular features in cancer biopsies were superior to corresponding 2D features for risk stratification of patients with low- to intermediate-risk prostate cancer based on their clinical biochemical recurrence outcomes. The results of this study support the use of computational 3D pathology for guiding the clinical management of prostate cancer. Significance: An end-to-end pipeline for deep learning–assisted computational 3D histology analysis of whole prostate biopsies shows that nondestructive 3D pathology has the potential to enable superior prognostic stratification of patients with prostate cancer.

Background: CDK 4/6 inhibitors are currently standard of care for patients CDK 4/6 inhibitors but currently lack validated predictive markers. We evaluated whether changes in intra-lesional heterogeneity, extracted computationally from routine computed tomography (CT) of MBC could predict survival following treatment with CDK 4/6 inhibitors. Methods: 33 patients with ER+ breast cancer metastasized to the liver who received palbociclib in combination with anti-estrogen therapy were identified from a registry. For each patient, CT exams were acquired prior to treatment and following initiation of treatment (median time between scans of 106 +/- 48.6 days). Liver metastases were manually delineated on all scans with consultation of radiology reports. On each scan, gray level co-occurrence matrix (GLCM) entropy, a computational measure of intra-lesional heterogeneity, was calculated and averaged across all metastatic lesions and its change over treatment was computed. For comparison, response was also assessed via RECIST v1.1: 65% of patients had objective response/stable disease, while 35% had progressive disease on scan following initiation of treatment. Imaging metrics were assessed individually and in a multivariate comparison including clinical features for association with overall survival (OS) via a cox proportional hazards model. Results: Change in intra-lesion heterogeneity was associated with OS with a hazard ratio [HR] of 2.82 (p=0.019). An increase in entropy between the pre- and treatment scan was associated with poorer viability on treatment including CDK inhibitors, indicating that a favorable response can be distinguished by a decrease in intra-lesion heterogeneity. RECIST response was also associated with OS (HR=0.24, p Conclusions: The change in lesion heterogeneity on clinical imaging following the initiation of treatment was found to offer independent value in distinguishing patients with favorable survival following CDK 4/6 inhibitor treatment. Multivariate association of imaging features and clinical variables with overall survivalHazard ratio (HR)p-valueChange in Heterogeneity4.670.019RECIST response0.160.003PR status1.010.99HER2 status1.130.92Age1.040.26Pathology0.770.39Line of Treatment1.360.12 Citation Format: Nathaniel Braman, Siddharth Kunte, Kaustav Bera, Jame Abraham, Albert Montero, Anant Madabhushi. Change in intra-lesion heterogeneity on CT predicts long-term survival following treatment with CDK4/6 inhibitors in hormone receptor-positive metastatic breast cancer (MBC) [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS5-42.

Background: Breast cancer is the most common cancer worldwide. There has been emerging interest in studying differences in breast tumor phenotypes between South Asian and South American women given that racial/ethnic disparities in incidence and mortality have been demonstrated in multiple studies. South Asian women are more likely to be diagnosed with an advanced stage breast tumor despite lower incidence than in North American women. There is evidence that computer-extracted nuclear morphological features on H&E slide images may be associated with breast cancer aggressiveness, specifically recurrence and disease free survival. However, studies have mostly focused on North American women. In this work, we evaluated whether there is a difference in computer extracted features of nuclear morphology from H&E tissue slide images between South Asian (SA) and North American (NA) women and we also investigated how these differences could impact the development of population specific breast cancer prognostic models. Methods: H&E slides of breast tumors from patients who were diagnosed with ER+ early stage invasive breast cancer from Tata Memorial Centre, India (SA: 69 (20 recurrence)) and from University Hospitals Cleveland Medical Center (NA: 121 (20 recurrence)), along with outcome information were collected. All slides were digitized on either a Ventana DP 200 or a Roche Ventana iScan HT slide scanner. For each image, a conditional Generative Adversarial Network model was employed to segment the individual nuclei, which were used to generate 241 nuclear features including nuclear architecture, shape, orientation disorder, and texture features. Half of the patients (95) were randomly selected as the training set (Stra) with the remaining patients as the hold-out validation set (Stest). Three elastic net regularized Cox regression models (MSA, MNA, MSA+NA) were trained fitting between the nuclear features and disease free survival (DFS) respectively for SA subset, NA subset, and SA & NA set in Stra. The top five prognostic features were identified respectively from each of the three models (MSA, MNA, MSA+NA) which were further validated on Stest to evaluate their prognostic value in prediction of DFS. Results: We found that the prognostic features identified by MNA and MSA+NA were mostly shape features (three out of five for MNA, four out of five for MSA+NA), while the prognostic features identified by the model specifically trained with the SA population (MSA) were mostly texture features (three out of five), possibly reflecting chromatin patterns in the cell. MSA yielded a better performance (Hazard Ratio=4.99 (p=0.00928, CI=1.32-18.9) from log-rank test between model derived high and low risk categories) on SA population in Stest compared to MNA and MSA+NA. Conclusion: We found that nuclear histomorphometry features were different between breast cancer patients from South Asia and those from North America. The prognostic capability of the computational pathology-based models for South Asian women could be significantly improved by taking into account population-specific information. An additional independent validation set is needed to confirm the preliminary findings presented here. Citation Format: Haojia Li, Kaustav Bera, Paula Toro, Pingfu Fu, Vidya Rao, Shabina Siddique, Aparna Harbhajanka, Haley Sechrist, Zelin Zhang, Sangeeta Desai, Vani Parmar, Anant Madabhushi. Computerized image analysis of nuclear morphological features reveals differences in phenotype and prognosis of disease free survival of early stage ER+ breast cancers for South Asian and North American women [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS4-45.

Purpose: To (i) create a survival risk score using radiomic features from the tumor habitat on routine MRI to predict progression-free survival (PFS) in glioblastoma and (ii) obtain a biological basis for these prognostic radiomic features, by studying their radiogenomic associations with molecular signaling pathways. Experimental Design: Two hundred three patients with pretreatment Gd-T1w, T2w, T2w-FLAIR MRI were obtained from 3 cohorts: The Cancer Imaging Archive (TCIA; n = 130), Ivy GAP (n = 32), and Cleveland Clinic (n = 41). Gene-expression profiles of corresponding patients were obtained for TCIA cohort. For every study, following expert segmentation of tumor subcompartments (necrotic core, enhancing tumor, peritumoral edema), 936 3D radiomic features were extracted from each subcompartment across all MRI protocols. Using Cox regression model, radiomic risk score (RRS) was developed for every protocol to predict PFS on the training cohort (n = 130) and evaluated on the holdout cohort (n = 73). Further, Gene Ontology and single-sample gene set enrichment analysis were used to identify specific molecular signaling pathway networks associated with RRS features. Results: Twenty-five radiomic features from the tumor habitat yielded the RRS. A combination of RRS with clinical (age and gender) and molecular features (MGMT and IDH status) resulted in a concordance index of 0.81 (P < 0.0001) on training and 0.84 (P = 0.03) on the test set. Radiogenomic analysis revealed associations of RRS features with signaling pathways for cell differentiation, cell adhesion, and angiogenesis, which contribute to chemoresistance in GBM. Conclusions: Our findings suggest that prognostic radiomic features from routine Gd-T1w MRI may also be significantly associated with key biological processes that affect response to chemotherapy in GBM.

Background: Although the advent of targeted therapy has substantially improved outcomes for HER2+ breast cancer patients, many will still fail to achieve pathological complete response (pCR) following neoadjuvant chemotherapy (NAC). In order to reduce overtreatment among patients resistant to standard HER2-targeted NAC and identify candidates for alternative therapeutic interventions, there is a need for validated markers of anti-HER2 agent benefit. The computational analysis of pretreatment imaging has shown recent promise in identifying responsive breast cancers. However, previous applications have explored response prediction among cohorts of mixed subtype and therapeutic approach, thus limiting its relevance in informing specific therapeutic strategy. Methods: This study comprised retrospective contrast-enhanced MRI data from a total of 159 patients who received anti-HER2 therapy at 5 institutions. A deep learning (DL) model was trained and tuned using 100 HER2+ breast cancer patients who received neoadjuvant taxane (T), carboplatin (C), trastuzumab (H), and pertuzumab (P) at Institution A, of which 49 achieved pCR (ypT0/is). A convolutional neural network was designed to analyze pre- and post-contrast MRI images acquired before NAC and compute a patient's probability of achieving pCR. Institution A data was split randomly into a 85 patient training cohort, used to directly train the model, and a 15 patient internal validation cohort, used to monitor and improve training progress. Two external, held-out testing datasets were used to evaluate capability to predict response in HER2+: Testing Cohort 1, consisting of 28 patients (16 pCR, 12 non-pCR) treated with TCHP at Institution B, and Testing Cohort 2, consisting of 29 patients (10 pCR, 19 non-pCR) who received TCH at one of 3 other institutions as part of the BrUOG 211B trial. A multivariable clinical model (MCM) incorporating age, ER/PR status, stage, and size was evaluated separately and in combination with the neural network. Performance was assessed by area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Results: The neural network was able to strongly predict response to HER2-targeted NAC in the internal validation (AUC=.93) and testing cohorts (AUC=.84 and AUC=.77). This model offered superior performance compared to a MCM, which performed poorly across institutions. Strikingly, the higher accuracy of DL included correctly identifying responders within the ER+/PR+ subgroup of patients and non-responders within the ER-/PR- subgroup. Combining DL predictions with the clinical model improved performance to AUC of 0.89 in testing cohort 1, but did not improve AUC in cohort 2. Conclusions: DL analysis of breast DCE-MRI could be used to better identify benefit of HER2-targeted therapeutic approaches prior to administration. As the first exploration of automated response prediction from imaging with respect to a targeted NAC approach, this work uniquely has the potential to help guide therapeutic decision-making. Our approach was effective in predicting response to multiple HER2-targeted NAC regimens, with better performance in the cohort who received TCHP (as in the training cohort). The strong performance of this model across 5 institutions is a promising indicator of its robustness and ability to tailor therapy even within clinically-distinct HER2+ patient subpopulations. Deep learning (DL) and multivariable clinical model (MCM) pCR prediction by cohortCohortModelAUC (%)Sensitivity (%)Specificity (%)Accuracy (%)Validation (n=15, 53% pCR)DL93888687MCM66637173Testing 1 (n=28, 57% pCR)DL85889289MCM54388350DL+MCM89819286Testing 2 (n=29, 34% pCR)DL77708479MCM53208966DL+MCM76808483 Citation Format: Nathaniel Braman, Mohammed El Adoui, Manasa Vulchi, Paulette Turk, Maryam Etesami, Pingfu Fu, Stylianos Drisis, Vinay Varadan, Donna Plecha, Mohammed Benjelloun, Jame Abraham, Anant Madabhushi. Validation of neural network approach for the prediction of HER2-targeted neoadjuvant chemotherapy response from pretreatment MRI: A multi-site study [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P4-10-13.

Background: Angiogenesis is crucial to a tumor's growth and an important factor in therapeutic outcome. Although quantitative analysis of tumors on dynamic contrast enhanced (DCE) MRI can provide indirect characterization of a tumor's vascularization, direct computational analysis of the tumor-associated vessel network remains a promising, but under-explored potential marker of therapeutic response. For instance, surrounding vasculature with a convoluted 3-dimensional shape and poor blood flow may indicate a more aggressive tumor and poorly facilitate delivery of therapeutic agents. In this work, we present a computational approach for the prediction of neoadjuvant chemotherapy response using quantitative imaging features describing the morphology and function of tumor associated vasculature on pretreatment MRI. Methods: 243 patients who received DCE-MRI scans prior to neo-adjuvant chemotherapy (NAC) at institution A [n=83], B [n=76], or one of nine other institutions as part of the ISPY1 Trial [n=84] were divided randomly into training (n=123) and testing (n=120) sets. 148 patients were HER2- and received neoadjuvant AC-T, while the 95 HER2+ patients were treated with TCHP (ISPY predates anti-HER2 therapy and HER2+ ISPY patients were excluded). 79 patients achieved pathological complete response [pCR, ypT0/is] following NAC. MRI exams were collected with a 1.5 or 3 Tesla scanner in the axial or sagittal plane. A baseline scan and 2-5 scans after injection of a gadolinium-based contrast agent with a median temporal resolution of 2.5 minutes were acquired. A portion of the tumor was manually delineated, then semi-automatically expanded to 3D. Vasculature was isolated from subtraction images with a specialized filtering approach to detect vessel-shaped objects. Features describing the 3D shape and architecture of the tumor-associated vessel network (e.g. curvature, torsion, and local orientation) and functional semi-quantitative pharmacokinetic (PK) measurements of temporal contrast enhancement changes (e.g. signal enhancement ratio, time to peak enhancement, and rates of uptake and washout) were calculated. Performance was assessed by area under the receiver operating characteristic curve (AUC), as well as the accuracy, sensitivity, and specificity at the operating point corresponding to the Youden Index. The most discriminating features were determined based on frequency of selection by the random forest classifier. Results: Within the training set, PK parameters of the vessels (AUC=.66) outperformed relative to the PK of tumor (AUC=.63) and the PK of peritumoral regions (AUC=.64); however, a combination of the three yielded best performance (AUC=.75). Vessel shape features alone achieved AUC=.67 in the training set. When multi-region PK features and tumor shape features were combined and applied to the 120-patient independent testing set, the random forest classifier achieved an AUC of 0.70 and identified 81% of patients who would achieve pCR. Non-pCR was best characterized by increased vessel curvature and PK parameters indicating poor perfusion, such as greater time to peak enhancement, slower uptake rate, and quicker washout. Conclusions: Our findings suggest that properties of the tumor-associated vessel network, such as its shape and enhancement profile, might provide value in identifying patients who will respond to NAC before administration of treatment. Accuracy (%)Sensitivity (%)Specificity (%)Full testing set (n=120)678160HER2+ (n=44)709548HER2- (n=76)646365HER2-, HR+ (n=51)678364Triple Negative (n=25)605067 Citation Format: Nathaniel Braman, Prateek Prasanna, Kaustav Bera, Mehdi Alilou, Manasa Vulchi, Maryam Etesami, Paulette Turk, Jame Abraham, Donna Plecha, Anant Madabhushi. Radiomic measurements of tumor-associated vasculature morphology and function on pretreatment dynamic MRI identifies responders to neoadjuvant chemotherapy [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P1-10-06.

No predictive biomarkers can robustly identify patients with non–small cell lung cancer (NSCLC) who will benefit from immune checkpoint inhibitor (ICI) therapies. Here, in a machine learning setting, we compared changes (“delta”) in the radiomic texture (DelRADx) of CT patterns both within and outside tumor nodules before and after two to three cycles of ICI therapy. We found that DelRADx patterns could predict response to ICI therapy and overall survival (OS) for patients with NSCLC. We retrospectively analyzed data acquired from 139 patients with NSCLC at two institutions, who were divided into a discovery set (D1 = 50) and two independent validation sets (D2 = 62, D3 = 27). Intranodular and perinodular texture descriptors were extracted, and the relative differences were computed. A linear discriminant analysis (LDA) classifier was trained with 8 DelRADx features to predict RECIST-derived response. Association of delta-radiomic risk score (DRS) with OS was determined. The association of DelRADx features with tumor-infiltrating lymphocyte (TIL) density on the diagnostic biopsies (n = 36) was also evaluated. The LDA classifier yielded an AUC of 0.88 ± 0.08 in distinguishing responders from nonresponders in D1, and 0.85 and 0.81 in D2 and D3. DRS was associated with OS [HR: 1.64; 95% confidence interval (CI), 1.22–2.21; P = 0.0011; C-index = 0.72). Peritumoral Gabor features were associated with the density of TILs on diagnostic biopsy samples. Our results show that DelRADx could be used to identify early functional responses in patients with NSCLC.

Purpose: The presence of a high degree of tumor-infiltrating lymphocytes (TIL) has been proven to be associated with outcome in patients with non–small cell lung cancer (NSCLC). However, recent evidence indicates that tissue architecture is also prognostic of disease-specific survival and recurrence. We show a set of descriptors (spatial TIL, SpaTIL) that capture density, and spatial colocalization of TILs and tumor cells across digital images that can predict likelihood of recurrence in early-stage NSCLC. Experimental Design: The association between recurrence in early-stage NSCLC and SpaTIL features was explored on 301 patients across four different cohorts. Cohort D1 (n = 70) was used to identify the most prognostic SpaTIL features and to train a classifier to predict the likelihood of recurrence. The classifier performance was evaluated in cohorts D2 (n = 119), D3 (n = 112), and D4 (n = 112). Two pathologists graded each sample of D1 and D2; intraobserver agreement and association between manual grading and likelihood of recurrence were analyzed. Results: SpaTIL was associated with likelihood of recurrence in all test sets (log-rank P < 0.02). A multivariate Cox proportional hazards analysis revealed an HR of 3.08 (95% confidence interval, 2.1–4.5, P = 7.3 × 10−5). In contrast, agreement among expert pathologists using tumor grade was moderate (Kappa = 0.5), and the manual TIL grading was only prognostic for one reader in D2 (P = 8.0 × 10−3). Conclusions: A set of features related to density and spatial architecture of TILs was found to be associated with a likelihood of recurrence of early-stage NSCLC. This information could potentially be used for helping in treatment planning and management of early-stage NSCLC. See related commentary by Peled et al., p. 1449

The rapid clinical expansion of immune checkpoint blockers (ICB) has transformed the therapeutic arsenal for non-small cell lung cancer (NSCLC). However, only a relatively small fraction of patients benefits from these treatments and those who respond can develop acquired resistance. This underlines the need to identify factors or biomarkers associated with treatment sensitivity and resistance to guide clinical decisions. It has been proposed that elevated levels of tumor infiltrating lymphocytes (TIL) in the tumor microenvironment could be associated with favorable clinical outcomes after treatment with ICB. However, few studies have explored this relationship and they used semi-quantitative scoring methods and/or analyzed small tumor areas. Here, we used multiplexed quantitative immunofluorescence (QIF) panels to study major TIL subsets (DAPI, CK, CD4, CD8 and CD20) and T-cell exhaustion markers (DAPI, CD3, LAG-3, PD-1, TIM-3) in full-face baseline whole-tumor slides from a large multi-institutional patient cohort (n=179). The analysis of numerous fields of view per case allowed us to examine the spatial distribution of immune cells and spatial immune heterogeneity. Our results demonstrate that CD8+ effector T-cells are the largest TIL subpopulation and that they are preferentially located in the stromal compartment. We also show that tertiary lymphoid structures have limited association with survival and their detection depends on the size of the tissue area analyzed. Additionally, higher levels of baseline CD8+ T-cells in the stroma were significantly associated with better survival in PD-L1 positive patients but not in PD-L1 negative. To validate these results, we used whole-exome sequencing to analyze the TCR-burden in an independent cohort of ICI-treated NSCLC patients. Increased expression of the exhaustion markers LAG-3 and TIM-3 in CD3 positive T-cells was associated with reduced survival. Finally, we used a novel multiparametric heterogeneity metric termed Rao’s quadratic entropy that enabled us to measure the spatial immune heterogeneity and was associated with worse survival. In summary, we quantitatively measured the density, functional properties, and spatial distribution of major TIL subsets in a large cohort of NSCLC patients treated with ICB. Our results highlight the prominent role of TILs in NSCLC and their potential role as a biomarker. These results could be easily translated into the clinic and used to guide optimal treatment options. Citation Format: Miguel Lopez de Rodas Gregorio, Venkata Nagineni, Arvind Ravi, Ila J. Datar, Mari Mino-Kenudson, German Corredor, Cristian Barrera, Lindsey Behlman, David L. Rimm, Roy S. Herbst, Anant Madabhushi, Jonathan W. Riess, Vamsidhar Velcheti, Matthew Hellmann, Justin F. Gainor, Kurt A. Schalper. Role of tumor infiltrating lymphocytes and spatial immune heterogeneity in sensitivity to PD-1 axis blockers in non-small cell lung cancer (NSCLC) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1722.

Introduction: The ODx test is a 21 gene assay that is currently employed for separating Estrogen Receptor positive (ER+) breast cancer patients into low (L) and high (H) risk of recurrence categories, helping clinicians decide if adjuvant chemotherapy is appropriate. In this study, we sought to explore whether computer extracted features pertaining to tumor grade (nuclear pleomorphism, tubule count, mitotic index) in conjunction with a machine learning classifier were predictive of the corresponding ODx risk category for ER+ breast cancer patients. Design: First, 2000x2000 pixel sub-regions of digitized H&E slides at 40x are processed to both identify and segment epithelial and stromal nuclei using a combination of watershed and deep learning (DL). 247 nuclear features consisting of architecture, shape, and texture features were extracted from these segmentations. Subsequently, the mitotic and tubule related features were extracted at each nuclei candidate using DL detectors. The input to this process was a binary mask computed by thresholding a blue ratio transformed image using Otsu's method. The identified regions were analyzed using DL to determine if a nucleus is a part of a tubule, and/or if it is mitotic. Finally, all of these features were combined, evaluated using Ranksum feature ranking, and then used to generate predictive models using four different supervised machine learning classifiers - random forest, support vector machine, linear discriminant analysis, and a neural network – via a 3-fold cross validation scheme. The classifiers were evaluated by their ability to distinguish between the four different classification tasks presented above using the area (AUC) under the Receiver Operating Characteristic (ROC) curve: 1) L ODx and L mBR grade vs. H ODx and H mBR grade (L-L vs. H-H), 2) L ODx vs. H ODx, 3) L ODx vs. Intermediate (T) and H ODx, 4) L and T ODx vs. H ODx. Results: The highest performing features were consistently mitosis, epithelial architectural, and tubule features. Classification accuracy ranged from 0.61 (L vs. T and H) to 0.97 (L-L vs. H-H) (Table 1). These features were able to provide the highest level of classification utility for the most distinct cases (L-L vs. H-H) and had less classification accuracy with classification problems involving more difficult T cases. Number top 10 Features in each categoryNumber top 10 Features in each categoryNumber top 10 Features in each categoryExperimentMax AUCMitosisTubuleEpithelial ArchitectureL-L vs. H-H (N=36)0.97315L vs. H (N=72)0.77505L vs. T and H (N=125)0.61208L and T vs. H (N=125)0.75505Table 1: Maximum AUC, and best features used to obtain those results. Conclusion: Computer derived features pertaining to nuclear architecture and mitotic index were predictive of ODx risk categories. Additional independent validation of these findings is needed in a separate test set. Citation Format: Whitney JR, Romeo-Bucheli D, Janowczyk A, Ganesan S, Feldman M, Gilmore H, Madabhushi A. Computer extracted features of tumor grade from H&E images predict oncotype DX risk categories for early stage ER+ breast cancer [abstract]. In: Proceedings of the 2017 San Antonio Breast Cancer Symposium; 2017 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2018;78(4 Suppl):Abstract nr P4-09-11.

Background: Ductal Carcinoma in Situ (DCIS) is a pre-invasive stage of breast cancer, where malignant cells line the duct but have not spread into other parts of the breast. Oncotype DX (ODX) is a genomic test, which divide patients into three risk of recurrence categories (Low, Intermediate, and High) to help physicians determine if patients require adjuvant therapy. However, ODX is expensive, tissue destructive, and has a turnaround time of 7-10 days. There has been an interest in the use of image analysis of routine H&E histopathology slides to predict the course of the disease; the rationale being that the analytics are able to unearth subtle sub-visual cues regarding disease morphology that may escape visual examination. In this work, we evaluate the role of computer extracted features of nuclear morphology and the necrotic regions from surgically resected specimens to predict ODX categories in patients with DCIS. Methods: H&E slides from breast tissue of 37 patients who were diagnosed with DCIS and underwent a lumpectomy were acquired. Nine of the 37 had high ODX score (higher than 54), while the rest had a low score (lower than 39). All the slides were digitized on a Philips slide scanner. For each image, a watershed algorithm segmented the individual nuclei, which were used to generate 230 nuclei features including nuclear architecture, nuclear shape and nuclear texture features within each candidate breast duct. In addition, we captured the area of necrosis and empty lumen region inside breast ducts to generate features pertaining to tubule packing. The average feature values for each patient were calculated across all the breast ducts in each slide. A 3-fold cross validation scheme with 50 repetitions was used with the Support Vector Machine (SVM) classifier to predict the ODX risk category for each patient. We used a covariance algorithm to select the top 4 features that were independent of each other but relevant to the ODX class label. Results: The top ranked features included features from three categories: nuclei architectural features (standard deviation of triangle area in Delaunay graph, skewness of edge length in Cell Cluster Graph), nuclear texture (standard deviation of Haralick matrix intensity) possibly reflecting chromatin patterns in the cell, and the Tubule Packing Ratio, a measure of the ratio of necrosis area and empty lumen area inside the breast ducts compared to the whole breast duct area. The SVM in conjunction with these 4 features yielded a mean area under receive operator characteristic curve (AUC) of 0.95 in correctly predicting high and low ODX risk categories. Conclusion: We found that our histomorphometry features pertaining to nuclear arrangement, nuclear texture and necrosis could differentiate between DCIS patients with high and low ODX risk categories. Additional independent validation of the approach is needed to confirm the preliminary findings presented here. Citation Format: Li H, Whitney J, Thawani R, Gilmore H, Badve S, Madabhushi A. Quantitative image features of nuclear and tubule architecture distinguish high and low oncotype DX risk categories of ductal carcinoma in situ from H&E tissue images [abstract]. In: Proceedings of the 2017 San Antonio Breast Cancer Symposium; 2017 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2018;78(4 Suppl):Abstract nr P4-09-12.

Background: Recent research has shown that collagen fiber arrangement in the tumor microenvironment plays a key role determining breast cancer prognosis. In this work, we present a new quantitative histomorphometric approach to identify and evaluate whether Collagen Fiber Orientation Disorder from Tumor-associated Stroma (CFOD-TS) on digitized H&E slides are associated with disease free survival (DFS) in estrogen receptor positive (ER+) breast cancer patients. Method: 234 and 171 ER+ invasive breast cancer patient’s slides were acquired and digitized from E2197 (training set) and TCGA-BRCA (test set) respectively. Collagen fibers were first segmented out on the digitized H&E slides by a combination of Derivative of Gaussian filters. The CFOD-TS features were then extracted from tumor patches in the whole slide image. A multiple linear regression model (MLR) was then trained in conjunction with the CFOD-TS features distinguishing patients likely to recur from not (>15 year follow-up period) to predict DFS for the non-censored patients. Result: The CFOD-TS discriminated patients (p=0.01) with and without recurrence and was significantly associated with DFS on ECOG 2197. The univariate and multivariate analysis on the training set were summarized in table 1 and table 2. CFOD-TS was also independently prognostic on patients in the TCGA for stratifying patients into low- and high-risk groups with p=0.009 via a log rank test and a hazard ratio (HR) of 3.09 (95% CI =1.33~7.18, p=0.0088). HR between the two risk groups was 7.11 (n=104, 95% CI =1.53~32.94, p=0.01) for stage I / II tumors and 3.03 (n=39, 95% CI =0.79-8.67, p=0.11) for stage III / IV cancers. In a multivariate setting, CFOD-TS (HR -3.7) was more significant than T-stage (HR-2.88) in stratifying risk. (Table2) Table 1: Univariate DFS analysis on the training setP-valueHR(95% CI)Case numberCFDO-TS derived risk groups0.00272.04 (1.28 – 3.26)234low / intermediate Oncotype DX (ODx) vs High ODx0.940.96(0.36 – 2.56)68low ODx vs intermediate / High ODx0.00323.49 (1.52 – 8.01)68grade0.300.82 (0.56 – 1.19)209 Table 2: Multivariate DFS analysis on the training set (n=68) and test set (n = 171)VariableP-valueHR(95% CI)ECOG 2197low ODx vs intermediate / High ODx0.00144.00 (1.71 – 9.34)CFDO-TS derived risk groups: Low vs high-risk0.0192.53 (1.17 – 5.47)TCGAT-stage: T1/T2 vs T3/T40.032.88 (1.10 -7.50)N-stage: N0 vs N1/N2/N30.411.47 (0.59 - 3.70)CFOD-TS derived risk groups: Low vs high-risk0.0033.70 (1.57 - 8.67) Conclusion: Over-expression of CFOD-TS was found to be independently associated with lower likelihood of recurrence and could potentially serve as a prognostic marker of outcome for ER+ breast cancer. Citation Format: Haojia Li, Kaustav Bera, Hannah Gilmore, Nancy E Davidson, Lori J Goldstein, Anant Madabhushi. Histomorphometric measure of disorder of collagen fiber orientation is associated with risk of recurrence in ER+ breast cancers in ECOG-ACRIN E2197 and TCGA-BRCA [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P5-06-16.

Background: Ductal Carcinoma in Situ (DCIS) of the breast comprises a diverse group of cancerous lesions confined to the breast ducts. About 25% of all breast cancers in the US are DCIS and women with DCIS have a 10 times higher lifetime risk of progressing to invasive breast cancer (IBC). While DCIS is treated routinely with breast conserving therapy which includes a lumpectomy followed by adjacent radiation therapy (RT). However, studies have shown that RT can be omitted in women with a lower risk of recurrence/progression. Gene expression methods like Oncotype Dx can be used to risk stratify patients but are expensive, tissue-destructive and time consuming. Recently, there has been substantial interest in developing computerized approaches to quantitatively extract tumor histomorphometric (QH) features from digitized images of tissue slides to predict cancer outcome. In this work, we evaluate the role of QH features related to the architecture, shape, orientation disorder and texture of cancer nuclei from lumpectomy specimens of DCIS to predict disease-free survival (DFS). Methods: H&E whole slide images (WSI) and corresponding outcome information for 121 DCIS patients (61 with progression/recurrence and 60 without) from the no adjuvant treatment arm of the UK/ANZ DCIS trial were retrieved and collected from Queen Mary university of London. WSIs were scanned by 3DHistech scanner at 43x magnification. For each image, a deep learning based generative adversarial network model was employed to segment the cancer nuclei, following which 379 nuclear features from four feature families including architecture, shape, orientation disorder and Haralick texture features were extracted within each candidate breast duct in the WSI and averaged across the multiple cancerous breast ducts for each patient. The dataset was randomly split into training (D1, n=60) and a hold-out validation set (D2,n=61) with events split equally in the two sets. For each of the four feature families, a set of 6 top features was identified by an elastic Cox Model to predict DFS (time from diagnosis to recurrence/invasive breast cancer or death whichever is sooner) onD1. The feature sets giving significant prognostic value (assessed by log rank test between the feature derived risk groups) were combined, from which 6 top features were further identified by a Cox Model to predict DFS on D1. Results: The top ranked QH features included features from three feature families: 2 shape features, 1 orientation disorder feature and 3 texture features. The Elastic Cox model predicted DFS on D1 with hazard ratio (HR) of 7.29 (95% CI = 2.94 ~18.05, p=1.75e-05) and on D2 with HR of 2.07 (95% CI = 1.02 ~2.94, p=0.045). Conclusion: Computer extracted nuclear histomorphometric features pertaining to shape, orientation disorder and texture were associated with disease free survival for DCIS. Additional independent multi-site validation of the approach is planned to further confirm the preliminary findings presented here. Citation Format: Haojia Li, Kaustav Bera, Hannah Gilmore, Zelin Zhang, Jack Cuzick, Mangesh A Thorat, Anant Madabhushi. Computer extracted features of nuclear shape, orientation disorder and texture from H&E Whole slide images are associated with disease-free survival in ductal carcinoma in situ (DCIS) [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P5-06-15.

Background: 3D digital breast tomosynthesis (DBT) offers greater sensitivity than 2D mammography to the presence of architectural distortions (AD), a suspicious finding requiring biopsy. This higher sensitivity to AD is only beneficial if DBT can differentiate benign AD, such as radial scars (RS) from malignant AD (MAD). Automated analysis of clinical images using a neural network has potential to enhance the accuracy of breast cancer diagnosis, but such applications focus on screening and currently struggle with more challenging diagnoses (e.g. MAD vs. RS) where computer assistance would provide great clinical utility. In this work, we define a deep learning (DL) framework utilizing amalgamate predictions from a group of neural networks, specialized for multiple lesion regions and different acquisitional views to distinguish MAD from RS. Methods: A retrospective analysis was conducted of 69 patients screened at a single institution, each with AD visible on one or more DBT views. Stereotactic core biopsy determined 42 AD to be MAD and the remaining 27 to be RS. An attending breast radiologist manually identified a circular region of interest (ROI) containing the AD on a single slice of the mediolateral (ML, n=72) and/or craniocaudal (CC, n=68) views. 56 patients were utilized for training and tuning the DL networks in cross-validation, while 25 volumes from 13 patients (8 MAD, 5 RS) were held for independent testing. Volumetric samples of size 24x24x12 pixels were extracted randomly within each ROI for input to the networks. Two neural networks were trained to separately recognize patterns of MAD at the intra- and peri-lesional region. Pairs of networks were trained separately for analysis of the CC and the MLO view, resulting in 4 possible network variations. These networks were first assessed individually by cross-validation in the training set, then predictions from each were combined for diagnosis in the independent testing set. Performance was assessed by area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. Results: All networks individually achieved strong performance in distinguishing RS from MAD when assessed via cross-validation in the training set. On the independent test set, networks trained to recognize patterns of malignancy at the center of the distortion outperformed those focused on the lesion margin. The CC view was found to be most discriminating in neural network assessment. Independent Test PerformanceNetwork TypeAUCSensitivitySpecificityIntralesional, CC0.650.800.50Perilesional, CC0.510.400.63Intralesional, ML0.490.600.38Perilesional, ML0.550.600.50Combined Performance0.800.601.0 When predictions from these networks were combined for diagnosis on the independent validation set, performance increased significantly to an AUC of 0.8, with a sensitivity of 0.6 and a specificity of 1.0. Conclusions: In this investigation, we demonstrated that specialized deep learning classifiers, created to address the diagnosis of MAD vs RS, can assist in diagnostic settings without generating false negatives. Our approach combines neural networks to make a diagnosis based on various spatial regions from multiple DBT views. Our approach demonstrated very high specificity which may reduce the number of benign RS biopsies and surgical excisions for AD detected by DBT. Future work will further validate these findings on a larger, multi-institutional cohort and explore their influence on clinical decision-making. Citation Format: Tristan Maidment, Nathaniel Braman, Yijiang Chen, Farhad Mehrkhani, Uliyana Yankevich, Donna Plecha, Anant Madabhushi. A combination of intra- and peri-lesional deep learning classifiers from multiple views enables accurate diagnosis of architectural distortion malignancy with digital breast tomosynthesis [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr PD9-03.

Introduction: Breast cancer (BCa) Patients with ER+ tumors that are lymph node negative (LN-) typically receive hormonal therapy. There is a need to identify ER+ LN- patients that will not benefit from adjuvant chemotherapy and will respond to hormonal therapy alone. Oncotype DX, a quantitative prognostic and predictive gene assay, provides a recurrence score that has been correlated with distant and early recurrence. In this work we present an approach that employ computer extracted features of nuclear architecture and morphology from routine H&E slides alone that can distinguish early and distant recurrence in ER+ breast cancers. By constructing graph networks within epithelium and stroma regions, built using nuclei as vertices and edge connections between proximal nuclei, local nuclear architecture can be quantitatively characterized. Hosoya index (HI) (originally introduced for analysis of chemical bonds) is a measure of a bond (in this context nuclei connections in a graph). In this work, we leverage HI to measure structural similarities of graphs across the populations that are indicative of recurrence in LN- ER+ breast cancer tissue microarray (TMA) images. Design: In this study we considered two tissue microarrays (TMAs) comprising 453 early-stage lymph-node negative (LN-) estrogen receptor positive (ER+) breast cancer (BCa) patients (diagnosed with invasive ductal carcinoma), with a total of N=90 patients experiencing lifetime distant recurrence and N=343 patients who did not. All TMA cores were digitized at 20x magnification (0.33 um/pixel spatial resolution) using a digital whole-slide scanner. Each nucleus was identified via an automated computerized image analysis algorithm developed by our group. Then, using a cluster cell graph that encodes a link between a pair of nodes based on proximity, a series of graphs are constructed for a TMA. A HI value was then assigned to each local graph. A support vector machine classifier was trained in conjunction with the distribution of HI values for the early and distant recurrence cases on the training TMA (n=243, 50 early recurrences). Independent validation of the SVM classifier was performed on the second TMA (n=210, 40 early recurrences). Results: For the LN- ER+ breast cancer dataset, our method was able to distinguish tumors with early and distant recurrence with an accuracy of 75.4%, a positive predictive value of 78.6% and a negative predictive value of 76.4%. The separation between the Kaplan-Meier curves for early and distant recurrence of LN-, ER+ breast cancers on the validation set was statistically significant (p < 0.00102). Conclusion: Based only on tiny H&E punches, a computer-aided morphometric classifier appears to identify lymph node negative, ER+ breast cancers with a low likelihood of recurrence. With further validation, this approach could be developed into an image based assay which could serve as a lower cost alternative to Oncotype DX. Citation Format: Ali S, Rimm D, Ganesan S, Madabhushi A. Local nuclear architecture features from H&E images predict early versus distant recurrence in lymph node negative, ER+ breast cancers. [abstract]. In: Proceedings of the Thirty-Eighth Annual CTRC-AACR San Antonio Breast Cancer Symposium: 2015 Dec 8-12; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2016;76(4 Suppl):Abstract nr P5-07-12.

Clinical management decisions in prostate cancer are often based on risk determination with major uncertainty since limited tools are available to understand the risk of disease progression and guide the treatment decision process. Computer-aided quantitative histomorphometric analysis has emerged as a powerful computing tool to identify, characterize, and quantitate histologic features of tissues beyond human visual capabilities. Several quantitative features can be assessed, such as precise numeric measurements pertaining to the spatial arrangement and architecture of nuclei, shapes of nests and nuclei, and nuclear texture. This technology has proven to be useful for the detection of cancer in tissue sections and also for predicting tumor biology and clinical outcome in cancer patients. Utilizing a combination of synergistic strategy of quantitative histomorphometry and biomarker expression of NF-κB/p65 from prostate tissue specimens, we sought to fuse structural and functional information from morphological and molecular markers to better characterize disease progression improving prediction of biochemical recurrence (BCR). Here we utilized radical prostatectomy specimens (n=23) for feature extraction from 15 patients without BCR and 8 patients who experienced BCR (PSA > 0.2 ng/ml) within two years of surgery. Digitized H&E slides were annotated for a representative cancerous region, glands were automatically segmented, and 216 features of gland architecture, shape, and orientation disorder were extracted. Nuclei were automatically segmented from NF-κB/p65 stained slides. Based on digitally calculated stain optical density, every nuclei pixel was classified as either negative or weakly, moderately, or strongly positive for NF-κB/p65 staining. We first used the H&E features alone in leave-one-out cross validation with a naive Bayes classifier, using the top two features by t-test in every fold, to obtain a recurrence probability for each patient. We repeated that experiment with the NF-κB/p65 features. We then multiplied together patients' posterior class probabilities from each model for a final probability. In every fold a probability threshold was identified from the training set to classify patients as BCR or non-BCR. Our results demonstrate that the three most predictive H&E features were all gland orientation disorder features. The top NF-κB/p65 feature was percentage of nuclei pixels positive for staining. Accuracy predicted was 78% with H&E features alone, 74% with NF-κB/p65 features alone, and 87% in the aggregate model. Taken together, our results demonstrate that fusing nuclei NF-κB/p65 and gland morphology information allows for functional and morphologic characterization of prostate cancer, potentially allowing for improved risk characterization and prognosis prediction in prostate cancer patients. Citation Format: Patrick Leo, Eswar Shankar, Robin Elliott, Andrew Janowczyk, Nafiseh Janaki, Gregory T. MacLennan, Anant Madabhushi, Sanjay Gupta. Combination of quantitative histomorphometry with NFκB/p65 nuclear localization is better predictor of biochemical recurrence in prostate cancer patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr LB-021.

Background: Applying the PAM50 classifier to targeted RNA-Sequencing data allows HER2+ tumors to be sub-categorized into intrinsic breast cancer subtypes. HER2+ breast cancers belonging to the HER2-enriched [HER2-E] subtype exhibit the highest rate of response to neoadjuvant therapy with combination of HER2-blockade and chemotherapy, as well as dual-HER2 blockade alone. A non-invasive predictor of PAM50 subtype from clinical dynamic contrast-enhanced MRI [DCE-MRI] could provide valuable clinical guidance in the treatment of HER2+ breast cancer. In this work, we identify a set of computer-extracted heterogeneity features computed within the lesion and its surrounding peritumoral region capable of distinguishing HER2-E from other HER2+ breast cancers [Non-HER2-E]. We then demonstrate that this imaging signature of HER2-E is also predictive of pathological complete response [pCR] in an independent HER2+ testing set, consistent with the HER2-E subtype's elevated response to HER2-targeted therapy. Methods: The training set consisted of 42 HER2+ patients with both 1.5 or 3 T DCE-MRI and targeted RNA sequencing collected prior to neoadjuvant treatment from a multicenter trial [BrUOG 211B, n=35] and The Cancer Genome Atlas-Breast Cancer project [TCGA-BRCA, n=7]. Intrinsic subtypes were assigned by unsupervised hierarchical clustering of the PAM50 gene set. 19 patients were determined to belong to the HER2-E subtype, while the remaining 23 represented non-HER2-E subtypes [19 HER2-Luminal, 4 HER2-basal]. Lesion boundaries were annotated by an expertly trained radiologist and expanded to 5 annular peritumoral regions in 3 mm increments out to a maximum radius of 15 mm. Computer-extracted heterogeneity features were computed voxelwise within intratumoral and peritumoral regions by first order statistics. A top HER2-E-associated feature from each region was identified by Wilcoxon feature selection and used to train a diagonal linear discriminant analysis [DLDA] classifier to predict HER2-E in a 3-fold cross-validation setting. This classifier was then applied to pCR prediction from DCE-MRI in a testing set of 28 HER2+ patients with available post neoadjuvant chemotherapy surgical specimens at one institution. 16 patients achieved pCR (ypT0/is), while the remainder had partial or no response (non-pCR). Results: A combination of heterogeneity features within the intratumoral region and annular peritumoral regions out to 12 mm from the tumor yielded optimal results within the training set, with an average HER2-E prediction AUC of .77 +/- .03. When applied to response prediction in an independent testing set, this HER2-E classifier was predictive of pCR (AUC = .72). Conclusions: Computer-extracted heterogeneity features calculated within the tumor and the surrounding peritumoral environment on DCE-MRI were able to distinguish the HER2-E PAM50 intrinsic subtype from other HER2+ breast cancers. HER2-E was characterized by elevated expression of intratumoral and peritumoral heterogeneity features, indicating a more disordered imaging phenotype within and around the tumor. Additional independent validation of these findings is needed. Citation Format: Braman N, Prasanna P, Singh S, Beig N, Gilmore H, Etesami M, Bates D, Gallagher K, Bloch BN, Somlo G, Sikov W, Harris L, Plecha D, Varadan V, Madabhushi A. Intratumoral and peritumoral MRI signatures of HER2-enriched subtype also predict pathological response to neoadjuvant chemotherapy in HER2+ breast cancers [abstract]. In: Proceedings of the 2017 San Antonio Breast Cancer Symposium; 2017 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2018;78(4 Suppl):Abstract nr P4-02-06.

Background: It has been previously shown that computer-extracted heterogeneity features calculated within the peritumoral region of a breast cancer from baseline dynamic contrast-enhanced MRI [DCE-MRI] can predict treatment outcome. This approach may due to microenvironmental signatures of elevated immune response, a predictor of favorable response to neoadjuvant chemotherapy [NAC] in HER2+ breast cancer. We assessed the correlation between peritumoral radiomic features and a tissue-derived genomic immune index [II] in HER2+ breast cancer and whether these immune-correlated imaging features are predictive of pathologic response. Methods: 33 HER2+ patients with both 1.5 or 3 T DCE-MRI imaging and targeted RNA sequencing of biopsy samples collected prior to NAC from a multicenter trial [BrUOG 211B, n=26] and The Cancer Genome Atlas-Breast Cancer project [TCGA-BRCA, n=7] were retrospectively analyzed. II was derived from expression of a 140-gene immune signature using the ESTIMATE algorithm. An attending breast radiologist annotated lesion boundaries on the DCE-MRI phase of peak contrast enhancement. Beyond this intratumoral region, 5 annular peritumoral regions in 3 mm increments out to a maximum radius of 15 mm were analyzed. Computer-extracted heterogeneity descriptors computed within the intratumoral and peritumoral regions were summarized by first order statistics. Redundancy was reduced by eliminating correlated imaging features (R2>.6). From the remaining features, the 5 features that were collectively best correlated with II were selected by feed forward, leave-one-out multilinear regression. The regression model was applied to an independent test set of 28 HER2+ patients with post NAC surgical specimens. The estimated II was assessed for its ability to differentiate patients who achieved a pathologic complete response in the breast [pCR, ypT0/is] (n=16) and those who did not (n=12) by 2-sided Wilcoxon rank sum test of median and area under the receiver operating characteristic curve (AUC). Results: The set of top features that significantly correlated (p Conclusions: From a set of quantitative features characterizing heterogeneity within the peritumoral region on DCE-MRI, we identified peritumoral imaging features correlated with a genomic index of immune response in HER2+ breast cancer and were predictive of pathologic response in an independent testing set. Our findings suggest that the predictive capability of peritumoral radiomics may be tied to a patient's immune response to the cancer. In addition to providing insight to the biological basis of peritumoral radiomics, imaging signatures of immune response themselves possess clinical value as a potential means for the non-invasive prediction of HER2+ cancer biology and treatment outcome. Additional independent validation is needed on a larger test set to confirm our preliminary findings. Citation Format: Braman N, Prasanna P, Singh S, Beig N, Gilmore H, Etesami M, Bates D, Gallagher K, Bloch BN, Somlo G, Sikov W, Harris L, Plecha D, Varadan V, Madabhushi A. Radiogenomic analysis of HER2+ breast cancer reveals MRI features correlated with genomic immune index are predictive of neoadjuvant chemotherapy response [abstract]. In: Proceedings of the 2017 San Antonio Breast Cancer Symposium; 2017 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2018;78(4 Suppl):Abstract nr P4-02-07.

Akt/protein kinase B and transcription factor NF-κB are important signaling molecules and key survival factors involved in the control of cell proliferation, apoptosis and oncogenesis. Although Akt and NF-κB are components of separate signaling pathways, Akt can induce cell survival signal through NF-κB activation. Previous studies from our laboratory have shown constitutive activation of NF-κB/p65 and PI3K-Akt in clinical prostate cancer specimens and in autochthonous mouse model of prostate cancer. In this study we investigated if these two signaling pathways converge and whether dual targeting is effective in blocking disease progression. Using tissue samples obtained during transurethral prostatic resection and paraffin-embedded sections of benign and malignant prostate tissue, we first analyzed nuclear levels of NF-κB/p65, native/activated Akt levels, and evaluated co-localization of these two proteins. Compared to benign tissue, cancer specimens exhibited constitutive Akt and NF-κB/p65 activation which was more pronounced in high-grade cancer (Gleason grade 7-10). Immunohistochemical analyses further confirmed a progressive increase in the activated form of Akt and NF-κB/p65 in the nucleus of cancer tissues compared to benign specimens, and exhibited co-localization of these proteins in a subset of aggressive cancer. Utilizing androgen-responsive prostate cancer LNCaP cells, with increased Akt activity, and androgen-refractory PC-3 cells harboring high levels of Akt and NF-κB. Individual treatment of cells with Akt Inhibitor VIII (0.075-1.8µM) and NF-κB inhibitor Parthenolide (0.32-60µM) for 24-72 h demonstrated partial suppressive effect in cell growth. Strikingly, concurrent blocking of Akt and NF-κB/p65 at 1:10 molar ratio resulted in potentiated toxicity with marked inhibition in proliferation and induction of cell cycle arrest. Interruption of Akt activation using dominant-negative approach resulted in significant inhibition in Akt-phosphorylation (Ser473) and induction of apoptosis in LNCaP cells, whereas ectopic expression of NF-κB/p65 increased doubling time and invasiveness in these cells, which was abrogated by combinational treatment. Furthermore, combination treatment decreased the expression of p-IKK (Ser176/180), p-NF-κB/p65 (Ser536), and p-Akt (Ser473) and downstream targets viz. cyclin D1, Bcl-2, VEGF and MMP9 in both cell lines. In vivo administration of Inhibitor VIII and Parthenolide at 1:10 ratio to PC-3 tumor xenograft for 8 weeks resulted in marked decrease in tumor volume and reduced expression of target genes, compared to individual treatments. These results suggest that nuclear co-localization of Akt and NF-κB/p65 may be developed as prognostic biomarker and concurrent targeting of these molecules by combination of pharmacological inhibitors is an effective approach and is more efficacious than use of single agent blocking prostate cancer progression. Citation Format: Eswar Shankar, Rajnee Kanwal, Aditi Goel, Xiaoping Yang, Sanjeev Shukla, Gregory T. MacLennan, Pingfu Fu, Anant Madabhushi, Parameswaran Ramakrishnan, Sanjay Gupta. Targeting the PI3K-Akt and NF-κB pathways as a combination therapy in blocking prostate cancer progression [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 1080. doi:10.1158/1538-7445.AM2017-1080

Oncotype DX (ODX) is a 21 panel gene-expression based assay for predicting whether patients with estrogen receptor-positive (ER+) breast cancer (BCa) are candidates for adjuvant chemotherapy. However, the time and expense associated with genomic assays suggests the need for a non-invasive, imaging-based, pre-therapeutic tool for assessment of disease risk and selection of an appropriate treatment regimen. The objective of this research was to determine whether (a) computer extracted image features on T2-weighted (T2w) MRI and H&E stained histopathology are independently able to distinguish ER+ BCa with low and high ODX recurrence scores (RS) and (b) to determine whether there is a correlation between MRI and histologic features identified as being predictive of low and high ODX risk categories. A total of 11 ER+ BCa patients were considered in this study, based on availability of in vivo 1.5 Tesla T2w MRI. For each study, the corresponding formalin-fixed paraffin-embedded H&E stained tissue specimens were digitized at 20x (0.5 μm/pixel) using a whole-slide scanner. Of the 11 patients, 8 were identified in the low ODX (RS < 18) and 3 in the high ODX (RS > 30) risk categories. Each dataset was accompanied by expert annotations of (a) the lesion ROI on MRI and (b) boundaries of epithelial nuclei from a representative field-of-view on the digitized histology slide. For each MRI study, a multi-scale, multi-orientation Gabor filter bank was convolved with the annotated lesion area providing a set of 192 texture features (FMRI). For each corresponding histology image, 471 features (FHIST) were extracted describing both nuclear morphology (NM) and Laws texture (LT) within the nuclear regions. Independent 2-sample t-tests were used to identify salient features in FMRI and FHIST that are able to distinguish low and high ODX risk categories. We found that, for the MRI dataset, Gabor texture features at several scales and orientations yielded salient features (p < 0.05) while on histopathology, nuclear texture and convexity (shape) features were identified as the top discriminative features (p < 0.01). Relationships between significant features were evaluated via Spearman's rank correlation test (see table), where high correlations were observed between lesion texture on T2w MRI and nuclear texture and shape on histology. Correlation of histologic and MRI features able to distinguish low and high ODX RSHistologic feature correlated with ODXMRI feature correlated with ODXCorrelation coefficient (ρ)p-valueLT: 70 Mean HSVGF: Scale 2: Orientation 3: min/max-0.85450.0008NM: ConvexityGF: Scale 5: Orientation 6: mean-0.85450.0008LT: 70 Mean HSVGF: Scale 2: Orientation 3: min/max-0.83640.0013LT: 70 Mean HSVGF: Scale 3: Orientation 8: mean-0.83640.0013LT: 70 Mean HSVGF: Scale 3: Orientation 2: mean-0.81820.0021 Our results suggest that quantitative features extracted on both T2w MRI and histopathology can independently distinguish between low and high risk ODX classes. Moreover, some of these MRI and histologic features appear to be significantly correlated, suggesting that information regarding tumor biology is reflected in both MRI and histologic image features. Citation Information: Cancer Res 2013;73(24 Suppl): Abstract nr P2-03-01.

There will be an anticipated 29,480 prostate cancer (PCa) deaths in 2014 in the US. Nearly 40% of PCa patients will undergo radical prostatectomy (RP), which reduces the risk of death from PCa. Roughly ∼15% of these PCa patients tend to recur and of these a portion will metastasize (∼40%). Hence, a more accurate and objective computer-assisted image analysis (CAIA) automated PCa grading system should be considered. The current standard for Gleason grading involves an expert pathologist visually scoring Gleason grade patterns based upon H&E histologic glandular architecture (size, shape and organization). Using the same tissue microarray (TMA) containing 80 radical prostatectomy (RP) cases stratified by Gleason scores, the authors selected a subset of up to fifty 20X Aperio whole slide scanned H&E images representing GG 3 and GG4 cases. Three separate approaches to measure tissue texture and nuclear structure were used: 1) adaptive active contour scheme (AdACM) segmentation which focuses on nuclear shape and texture features 2) Curvelet transform based histologic tissue texture method and 3) MediaCybernetics ImagePro 9.1 software that includes ∼22 nuclear structure features. Data analysis utilized area under the curve (AUC) for the receiver operating characteristics (ROC) with assessment of accuracy, sensitivity and specificity. Ali and Madabhushi at Case Western Reserve University utilized AdACM segmentation approach and three nuclear features to obtain an AUC = 0.88 yielding an accuracy of 86.1% with sensitivity = 85.2% and specificity = 87.1% applying a quadratic discriminant analysis (QDA) classifier to separate GG 3 from GG4. Next, the University of Pittsburgh's EEE group (Wen-Chyi Lin and C. C. Li) applied a curvelet based tissue feature extraction method and a tree-structured classifier consisting of three Gaussian-kernel support vector machines each with an embedded voting mechanism and discriminated Gleason grade (GG) 3 vs GG 4 generating a AUC = 0.98 yielding an accuracy = 95.7% with sensitivity = 94.5% and specificity = 96.9%. Finally, the Veltri laboratory differentiated GG3 and GG4 and obtained an AUC = 0.96 and yielding an accuracy = 85%, sensitivity = 82%, specificity = 90% to separate GG pattern 3 from 4. Therefore using CAIA with three different laboratories we were able to use PCa glandular histologic and cell nuclear quantitative measurements to accurately discriminate GG patterns 3 and 4. More work is required to apply these automated tools to confirm the optimal method and also to predict PCa outcomes of recurrence, metastasis and survival. Citation Format: Robert W. Veltri, Sahirzeeshan Ali, Wen-Chyi Lin, Guangjing Zhu, Jonathan I. Epstein, Ching-Chung Li, Anant Madabhushi. Cancer histologic and cell nucleus architecture differentiate prostate cancer Gleason patterns 3 from 4. [abstract]. In: Proceedings of the 106th Annual Meeting of the American Association for Cancer Research; 2015 Apr 18-22; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2015;75(15 Suppl):Abstract nr 4349. doi:10.1158/1538-7445.AM2015-4349

Introduction: The progression of prostate cancer (PCa) involves both tissue morphometric changes and critical molecular alterations (such as cancer-associated markers). We have studied 80 men that underwent radical prostatectomy (RP) to evaluate this dual approach to discriminate Gleason score and progression. Next, we applied our approach at the diagnostic biopsy of men that chose active surveillance (AS) as an option for the management of their PCa, to predict the likelihood for re-classification of their disease to immediate treatment. Materials & Methods: Two tissue microarrays (TMAs) with 80 RP PCa cases, stratified by Gleason scores were used first. Next, a total of 27 favorable and 24 unfavorable AS biopsy PCa cases were evaluated. In both series we applied multiplex tissue immunoblotting (MTI) & quantitative nuclear morphometry studies. Data were first analyzed alone and then in combination using multivariate logistic regression (MLR) to predict the aggressive RP cases or AS biopsy reclassification status of the cancer. MTI was used to simultaneously detect 5∼6 biomarkers ((-5,-7)ProPSA, PCNA, RBM3, Her2/neu, & CACNA1D, etc) on a single 5 micron section. Proteins on the slide were transferred onto a series of 5∼6 P-film membranes and each membrane was probed with different primary antibody. The quantified FITC fluorescence signal of the biomarker were normalized to CY5 labelled total protein. For the nuclear morphometry analysis, quantification of the nuclear features were achieved either using ImagePro Premier 9.1 Software for the TMAs or the adaptive active contour scheme (AdACM) that uses nuclear shape, architectural and textural features extracted from AS biopsies. RESULTS: Using MLR, to differentiate aggressive PCa (Gleason score 4+3 & > = 8) from less aggressive PCa (Gleason score 3+3 & 3+4) on the TMAs of RP cases, our biomarkers model generated an ROC-AUC of 0.8 with accuracy of 71.43%, while morphometry model generate an ROC-AUC of 0.92 with accuracy of 82.50 %. When combined, it improved to an ROC-AUC of 0.96 and accuracy of 87.01%. Additionally, MLR was used for differentiation unfavorable biopsy cases that requires reclassification due to upgraded Gleason score, increased tumor volume, and/or PSA/PSAD during monitoring and need definitive treatment. On the contrary, favorable cases are very low risk (VLR) PCa biopsies that are not reclassified and can continue monitoring. The biomarkers model produced an ROC-AUC of 0.71 with an accuracy of 73.91% while morphometry model produces an ROC-AUC of 0.84 and accuracy of 74.51%. When combined, the new model produces an ROC-AUC of 0.88 and accuracy of 80.43%. CONCLUSIONS: Our method combining tissue morphometry with biomarkers demonstrated its translational clinical relevance since it can predict PCa aggressiveness in men that have undergone RP. Also, this approach predicts AS PCa cases requiring reclassification and immediate treatment at biopsy. Citation Format: Guangjing Zhu, George Lee, Christine Davis, Luciane Tsukamoto Kagohara, Jonathan I. Epstein, Patricia Landis, H. Ballantine Carter, Anant Madabhushi, Robert W. Veltri. Prediction of prostate cancer progression with biomarkers and tissue morphometry changes. [abstract]. In: Proceedings of the 106th Annual Meeting of the American Association for Cancer Research; 2015 Apr 18-22; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2015;75(15 Suppl):Abstract nr 4352. doi:10.1158/1538-7445.AM2015-4352

Background: In this study we investigate the ability of computer extracted image features (nuclear morphology and texture) from digitized H&E tissue slides to stratify women with lymph node negative (LN-), estrogen receptor positive (ER+) breast cancer (BCa) as low or high risk as determined by Oncotype DX (ODX), a 21 gene-expression assay. Each year, over 120,000 women in the United States (1 million worldwide) are diagnosed with ER+ BCa. Treatment guidelines recommend hormone therapy (HT) plus chemotherapy (CT); however, up to 85% of ER+ BCa patients will not benefit from CT, yet will still suffer its side effects. ODX yields a numeric risk score (RS) ranging from 1-100; RS30 indicates need for adjuvant CT. Unfortunately, this test is expensive (>$4000), time-consuming, and involves destructive tissue testing. The goal of this study is to show that quantitative features calculated from H&E images can accurately predict risk stratification as determined by ODX in women with LN-, ER+ BCa, suggesting a histologic image based classifier could serve as a low-cost alternative. Methods: Digitized H&E-stained ER+ BCa tissue sampled from 111 patients (34 high and 77 low-risk as determined by ODX) were obtained from the University of Pennsylvania, the University of Medicine and Dentistry of NJ, and Case Western Reserve University. Regions of cancer were annotated manually by an expert pathologist, and representative fields of view (FOV) were chosen at 20x magnification (2000 by 2000 pixels) for each patient. A selection of nuclear boundaries was annotated manually in each FOV. For each nucleus, a set of 2343 features was extracted, including 21 morphological (size, shape, and boundary) and 2322 texture (Gabor, Local Binary Pattern, Greylevel, and Laws filter features). Using Minimum Redundancy Maximum Relevance (mRMR) feature selection, the 3 features best able to separate low and high ODX risk categories were identified and used to build a supervised Bayesian classifier. Classifier training employed a randomized 3-fold cross-validation scheme; in each trial, two-thirds of the dataset were randomly selected for training, and the remaining one-third employed for independent testing. Classifier performance was evaluated using area under the receiver operating characteristic curve (AUC), positive predictive value (PPV), and negative predictive value (NPV) with respect to low and high ODX risk categorization. Performance metrics were averaged over 100 trials of 3-fold cross-validation (see table). Results: The mRMR method selected one morphological feature (nuclear area) and two Laws-based texture features as being highly discriminating between risk categories. The Bayesian classifier trained with these 3 features yielded high AUC, PPV, and NPV measures with low variance in distinguishing ODX risk categories. The supervised classification results indicate that quantitative image features from H&E-stained histopathology are able to accurately discriminate between low and high risk patients as determined by ODX. Classification PerformancePerformance MetricAverage (100 Trials)Standard DeviationAUC0.870.018PPV0.810.039NPV0.880.017 Citation Information: Cancer Res 2013;73(24 Suppl): Abstract nr P4-03-04.

The Oncotype DX (ODX) is a 21 panel gene-expression based assay for identifying which Estrogen Receptor-positive (ER+) breast cancer (BCa) patients are candidates for adjuvant chemotherapy. The objective of this research was to identify whether computerized texture features on a staging DCE-MRI can distinguish ER+ BCa with low and high ODX recurrence scores (RS) (i.e. to distinguish which ER+ BCa patients are more likely to benefit from adjuvant hormonal therapy from those who require chemotherapy). This would provide a non-invasive, imaging based, pre-therapeutic assessment tool for predicting the appropriate treatment regimen. This work, to the best of our knowledge, is the first attempt to quantitatively correlate low versus high risk stratification via computer derived MRI measurements to corresponding risk stratification via the ODX assay. 52 ER+ BCa patient studies with high (>30, N = 28) and low ( Table 1Feature classAccuracy (μ±Δ)AUC (μ±Δ)DHoG87.07%±5.66%0.89±0.04DLBP85.86%±7.82%0.83±0.07EK82.36%±8.46%0.80±0.06PK81.14%±7.55%0.78±0.07TK75.93%±6.65%0.76±0.08IK76.43%±7.23%0.75±0.12Shape71.04%±6.81%0.70±0.06 Table 1 illustrates the mean and standard deviation in accuracy and AUC values over 200 runs of randomized cross validation. DHoG, DBLP and EK features yielded the highest classification accuracy and AUC. Although lesion shape has been shown to be important for discriminating benign and malignant lesions on MRI, shape appears to be less useful in distinguishing between ER+ BCa lesions with low and high ODX RS. Citation Information: Cancer Res 2013;73(24 Suppl): Abstract nr P2-02-12.

Introduction: We used two novel computer-assisted imaging applications that incorporate nuclear shape and texture analysis to differentiate Gleason grade patterns 3 and 4 of H&Ein a tissue microarray (TMA). The two approaches utilized a variational adaptive segmentation scheme (AdACM) (Rutgers) and the cardinal multiridgelet transform (CMRT) based texture analysis (University of Pittsburgh). Methods: The TMA include 0.6 mm cores from blocks that included Gleason patterns of CaP radical prostatectomy cases from 40 patients. The TMA comprised 13 Gleason patterns 6 (3+3), 8 pattern 7 (3+4), 7 pattern 8 (4+4) and 5 pattern 9 (4+5) Gleason grade (GG) patterns. The AdACM uses 200 × 200 pixel images extracted for nuclear & architectural textural features. All features were based on mathematical energy terms with a shape prior in a multi-level set formulation. The CMRT is an image transform that is translation- and rotation-invariant, and provides a means of displaying texture contents for selection of discriminative features. With a 3-scale decomposition, it gives 8 ridgelet packets and 16 statistical texture features in a 256 x 256 pixel image patch that are averaged over 9 non-overlapping patches in a given image for separation of GG patterns 3 and 4. Results: The CMRT study utilized 3 of 16 texture features extracted from images and trained a Gaussian kernel support vector machine (SVM), a leave-one-out cross-validation showed an accuracy of 94% and AUC of 0.97. For AdACM, nuclear features and architectural features were extracted. Linear Discriminant Analysis (LDA) was employed for feature selection and best features from all three classes could discriminate Gleason grade 3 vs. 4 with an accuracy of 86% via a SVM classifier. Conclusions: Unique quantitative imaging of tissue nuclear shape and texture features can accurately differentiate Gleason patterns 3 and 4. Our research will next evaluate CaP outcomes. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 103rd Annual Meeting of the American Association for Cancer Research; 2012 Mar 31-Apr 4; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2012;72(8 Suppl):Abstract nr 4061. doi:1538-7445.AM2012-4061

Background: Measurement of estrogen receptor (ER) expression is a routine part of clinical evaluation of individual breast cancers and is used to guide treatment. However not all ER+ breast cancers show equal benefit from hormonal or other treatment. The RT-PCR based Oncotype Dx assay, has been recently shown to robustly stratify early stage ER+ breast cancer and identify those tumors that will have low recurrence rates when treated with adjuvant hormonal therapy alone. Interestingly, standard pathologic grading, based on visual analysis of tumor morphology by trained pathologists, has a strong correlation with Oncotype Dx recurrence scores; low recurrence tumors are mostly low grade, and high recurrence tumors are mostly high grade. However, a major problem with use of pathologist-assigned histologic grade as a prognostic tool is the lack of reproducibility of histological grading between different pathologists. Computer aided image analysis and machine learning techniques offer a way to obtain highly reproducible image-based classification of ER+ breast cancer. These analyses can be performed on digital images of routinely obtained breast cancer histology and be incorporated into a prognostic assay.Materials and Methods: High resolution digital images were obtained for a series of ER+ breast cancers for which associated Oncotype Dx Assay results were available. Regions of invasive breast cancer were identified and then processed by computer-assisted image analysis methods. An automated nuclear detection scheme based on Expectation Maximization algorithm was used to identify cancer cell nuclei, followed by graph-based feature extraction using the Voronoi graph, Delaunay Triangulation and Minimum Spanning Trees, and dimensionality reduction using Graph Embedding with Support Vector Machine based classification. The final manifold generated by GE was unwrapped into a linear space and a Euclidean distance metric was used to generate a single score (Image Based Risk Score- (IbRiS)) for each sample. Correlations between IbRS, clinical features such as grade, and Oncotype Dx Recurrance Score were determined.Results: Unsupervised analysis of image-based features of high resolution digital images of ER+ breast cancer histology leads to natural separation of tumors, with low grade tumors separating from high grade tumors. There is also a robust separation of tumors with high Oncotype Dx Recurrance Scores from tumors with low Recurrance Scores.Discussion: Unsupervised analysis of high resolution image-based features can stratify ER+ breast cancers in a fashion that correlates well with a gene-expression based prognostic assay, Oncotype Dx. These data suggest that tumors with distinct gene-expression profiles also have distinct image-based features that can be measured by computer-aided image analysis and used to build prognostic and predictive assays. Citation Information: Cancer Res 2009;69(24 Suppl):Abstract nr 3046.