Search

Supplementary Material and Methods

It is well-recognized that solid tumors are genomically, anatomically, and physiologically heterogeneous. In general, more heterogeneous tumors have poorer outcomes, likely due to the increased probability of harboring therapy-resistant cells and regions. It is hypothesized that the genomic and physiologic heterogeneity are related, because physiologically distinct regions will exert variable selection pressures leading to the outgrowth of clones with variable genomic/proteomic profiles. To investigate this, methods must be in place to interrogate and define, at the microscopic scale, the cytotypes that exist within physiologically distinct subregions (“habitats”) that are present at mesoscopic scales. MRI provides a noninvasive approach to interrogate physiologically distinct local environments, due to the biophysical principles that govern MRI signal generation. Here, we interrogate different physiologic parameters, such as perfusion, cell density, and edema, using multiparametric MRI (mpMRI). Signals from six different acquisition schema were combined voxel-by-voxel into four clusters identified using a Gaussian mixture model. These were compared with histologic and IHC characterizations of sections that were coregistered using MRI-guided 3D printed tumor molds. Specifically, we identified a specific set of MRI parameters to classify viable-normoxic, viable-hypoxic, nonviable-hypoxic, and nonviable-normoxic tissue types within orthotopic 4T1 and MDA-MB-231 breast tumors. This is the first coregistered study to show that mpMRI can be used to define physiologically distinct tumor habitats within breast tumor models.Significance:This study demonstrates that noninvasive imaging metrics can be used to distinguish subregions within heterogeneous tumors with histopathologic correlation.

Similarity index based on the weighted moving/sliding window was used to quantify the agreement for all incorrect classifications between the habitat-maps derived from MRI (6-parameters) and histology. The correct habitats from histological maps were compared with incorrect habitats from MRI maps. Thus, there were 9 incorrect classifications for maps containing V, H1, H2 and N, and 2 incorrect classifications for maps containing V, combined H1 + H2 and N to compare. The total average score for all wrong classifications for each sample is also shown in Supplementary Table S1.

SF1: Pimonidazole and carbonic anhydrase IX (CA-IX) staining patterns in 4T1 and MDA-MB-231 breast tumors; SF2: Breast tumor stained with hematoxylin & eosin (H&E), showing different stages of necrosis; SF3: Representative examples of the corresponding habitat maps from MRI and histology for the 4T1 and MDA-MB-231 tumor samples; SF4: Histograms showing the pixels values distribution of Area under the curve (AUC), slope and time to max (TTM) from DCE-MRI in each cluster of representative MRI habitat maps for 4T1 and MDA-MB-231 tumor samples; SF5: Representative examples showing that the DCE-MRI maps have visual similarity with pimonidazole staining in breast samples. SF6: Representative examples of MRI habitat maps, which were created by clustering five MRI-parameters [T2-map, T2*-map, Slope, Area under the curve (AUC) and time to max (TTM)] using a Gaussian Mixture Model (GMM) Expectation-Maximization algorithm. SF7: Comparison of mean values of each parameter from habitat maps generated by clustering five MRI-parameters [T2-map, T2*-map, Slope, Area under the curve (AUC) and time to max (TTM)]; SF8: Histograms showing the distribution of the Apparent diffusion coefficient (ADC) values in the cluster classified as non-viable tumor cells in the MRI habitat map. SF9: Representative images of the immunohistochemistry for perilipin-1 in breast tumor samples.

BACKGROUND. Higher categories of background parenchymal enhancement (BPE) increase breast cancer risk. However, current clinical BPE categorization is subjective. OBJECTIVE. Using a semiautomated segmentation algorithm, we calculated quantitative BPE measures and investigated the utility of individual features and feature pairs in significantly predicting subsequent breast cancer risk compared with radiologist-assigned BPE category. METHODS. In this retrospective case-control study, we identified 95 women at high risk of breast cancer but without a personal history of breast cancer who underwent breast MRI. Of these women, 19 subsequently developed breast cancer and were included as cases. Each case was age matched to four control patients (76 control patients total). Sociodemographic characteristics were compared between the cases and matched control patients using the Mann-Whitney U test. From each dynamic contrast-enhanced MRI examination, quantitative fibroglandular tissue and BPE measures were computed by averaging enhancing voxels above enhancement ratio thresholds (0–100%), totaling the enhancing volume above thresholds (BPE volume in cm(3)), and estimating the percentage of enhancing tissue above thresholds relative to total breast volume (BPE%) on each gadolinium-enhanced phase. For the 91 imaging features generated, we compared predictive performance using conditional logistic regression with 80:20 holdout cross validation and ROC curve analysis. ROC AUC was the figure of merit. Sensitivity, specificity, PPV, and NPV were also computed. All feature pairs were exhaustively searched to identify those with the highest AUC and Youden index. A DeLong test was used to compare predictive performance (AUCs). RESULTS. Women subsequently diagnosed with breast cancer were more likely to have mild, moderate, or marked BPE (odds ratio, 3.0; 95% CI, 0.9–10.0; p = .07). According to ROC curve analysis, a BPE category threshold greater than minimal resulted in a maximized AUC (0.62) in distinguishing cases from control patients. Compared with BPE category, the first gadolinium-enhanced (phase 1) BPE% at the 30% and 40% enhancement ratio thresholds yielded significantly higher AUC values of 0.85 (p = .0007) and 0.84 (p = .0004), respectively. Feature combinations showed similar AUC values with improved sensitivity. CONCLUSION. Preliminary data indicate that quantitative BPE measures may outperform radiologist-assigned category in breast cancer risk prediction. CLINICAL IMPACT. Future risk prediction models that incorporate quantitative measures warrant additional investigation.

Rationale: Hypoxic regions (habitats) within tumors are heterogeneously distributed and can be widely variant. Hypoxic habitats are generally pan-therapy resistant. For this reason, hypoxia-activated prodrugs (HAPs) have been developed to target these resistant volumes. The HAP evofosfamide (TH-302) has shown promise in preclinical and early clinical trials of sarcoma. However, in a phase III clinical trial of non-resectable soft tissue sarcomas, TH-302 did not improve survival in combination with doxorubicin (Dox), possibly due to a lack of patient stratification based on hypoxic status. Therefore, we used magnetic resonance imaging (MRI) to identify hypoxic habitats and non-invasively follow therapies response in sarcoma mouse models. Methods: We developed deep-learning (DL) models to identify hypoxia, using multiparametric MRI and co-registered histology, and monitored response to TH-302 in a patient-derived xenograft (PDX) of rhabdomyosarcoma and a syngeneic model of fibrosarcoma (radiation-induced fibrosarcoma, RIF-1). Results: A DL convolutional neural network showed strong correlations (>0.76) between the true hypoxia fraction in histology and the predicted hypoxia fraction in multiparametric MRI. TH-302 monotherapy or in combination with Dox delayed tumor growth and increased survival in the hypoxic PDX model (p

It is well-recognized that solid tumors are genomically, anatomically, and physiologically heterogeneous. In general, more heterogeneous tumors have poorer outcomes, likely due to the increased probability of harboring therapy-resistant cells and regions. It is hypothesized that the genomic and physiologic heterogeneity are related, because physiologically distinct regions will exert variable selection pressures leading to the outgrowth of clones with variable genomic/proteomic profiles. To investigate this, methods must be in place to interrogate and define, at the microscopic scale, the cytotypes that exist within physiologically distinct subregions (“habitats”) that are present at mesoscopic scales. MRI provides a noninvasive approach to interrogate physiologically distinct local environments, due to the biophysical principles that govern MRI signal generation. Here, we interrogate different physiologic parameters, such as perfusion, cell density, and edema, using multiparametric MRI (mpMRI). Signals from six different acquisition schema were combined voxel-by-voxel into four clusters identified using a Gaussian mixture model. These were compared with histologic and IHC characterizations of sections that were coregistered using MRI-guided 3D printed tumor molds. Specifically, we identified a specific set of MRI parameters to classify viable-normoxic, viable-hypoxic, nonviable-hypoxic, and nonviable-normoxic tissue types within orthotopic 4T1 and MDA-MB-231 breast tumors. This is the first coregistered study to show that mpMRI can be used to define physiologically distinct tumor habitats within breast tumor models. Significance: This study demonstrates that noninvasive imaging metrics can be used to distinguish subregions within heterogeneous tumors with histopathologic correlation.

This study evaluates breast MRI response of ER/PR+ HER2- breast tumors to pre-operative SABR with pathologic response correlation.Women enrolled in a phase 2 single institution trial of SABR for ER/PR+ HER2- breast cancer were retrospectively evaluated for radiologic-pathologic correlation of tumor response. These patients underwent baseline breast MRI, SABR (28.5 Gy in 3 fractions), follow-up MRI 5 to 6 weeks post-SABR, and lumpectomy. Tumor size and BI-RADS descriptors on pre and post-SABR breast MRIs were compared to determine correlation with surgical specimen % tumor cellularity (%TC). Reported MRI tumor dimensions were used to calculate percent cubic volume remaining (%VR). Partial MRI response was defined as a BI-RADs descriptor change or %VR ≤ 70%, while partial pathologic response (pPR) was defined as %TC ≤ 70%.Nineteen patients completed the trial, and %TC ranged 10% to 80%. For BI-RADS descriptor analysis, 12 of 19 (63%) showed change in lesion or kinetic enhancement descriptors post-SABR. This was associated with lower %TC (29% vs. 47%, P = .042). BI-RADS descriptor change analysis also demonstrated high PPV (100%) and specificity (100%) for predicting pPR to treatment (sensitivity 71%, accuracy 74%), but low NPV (29%). MRI %VR demonstrated strong linear correlation with %TC (R = 0.70, P.001, Pearson's Correlation) and high accuracy (89%) for predicting pPR (sensitivity 88%, specificity 100%, PPV 100%, and NPV 50%).Evaluating breast cancer response on MRI using %VR after pre-operative SABR treatment can help identify patients benefiting the most from neoadjuvant radiation treatment of their ER/PR+ HER2- tumors, a group in which pCR to neoadjuvant therapy is rare.

Pancreatic cancer (PaCa) is the third leading cause of cancer-related deaths in the United States. There is an unmet need to develop strategies to detect PaCa at an early, operable stage and prevent its progression. Intraductal papillary mucinous neoplasms (IPMNs) are cystic PaCa precursors that comprise nearly 50% of pancreatic cysts detected incidentally via cross-sectional imaging. Since IPMNs can progress from low- and moderate-grade dysplasia to high-grade dysplasia and invasion, the study of these lesions offers a prime opportunity to develop early detection and prevention strategies. Organoids are an ideal preclinical platform to study IPMNs, and the objective of the current investigation was to establish a living biobank of patient-derived organoids (PDO) from IPMNs. IPMN tumors and adjacent normal pancreatic tissues were successfully harvested from 15 patients with IPMNs undergoing pancreatic surgical resection at Moffitt Cancer Center & Research Institute (Tampa, FL) between May of 2017 and March of 2019. Organoid cultures were also generated from cryopreserved tissues. Organoid count and size were determined over time by both Image-Pro Premier 3D Version 9.1 digital platform and Matlab application of a Circular Hough Transform algorithm, and histologic and genomic characterization of a subset of the organoids was performed using immunohistochemistry and targeted sequencing, respectively. The success rates for organoid generation from IPMN tumor and adjacent normal pancreatic tissues were 81% and 87%, respectively. IPMN organoids derived from different epithelial subtypes showed different morphologiesin vitro, and organoids recapitulated histologic and genomic characteristics of the parental IPMN tumor. In summary, this pre-clinical model has the potential to provide new opportunities to unveil mechanisms of IPMN progression to invasion and to shed insight into novel biomarkers for early detection and targets for chemoprevention.

Hypoxic regions (habitats) within tumors are heterogeneously distributed and can be widely variant. Hypoxic habitats are generally pan-therapy resistant. For this reason, hypoxia-activated prodrugs (HAPs) have been developed to target these resistant volumes. The HAP evofosfamide (TH-302) has shown promise in preclinical and early clinical trials of sarcoma. However, in a phase III clinical trial, TH-302 did not improve survival in combination with doxorubicin (dox), most likely due to a lack of patient stratification based on hypoxic status. Herein, our goal was to develop deep-learning (DL) models to identify hypoxic habitats, using multiparametric (mp) MRI and co-registered histology, and to non-invasively monitor response to TH-302 in a patient-derived xenograft (PDX) of rhabdomyosarcoma and a syngeneic model of fibrosarcoma (RIF-1). A DL convolutional neural network showed strong correlations (>0.81) between the true hypoxic portion in histology and the predicted hypoxic portion in multiparametric MRI. TH-302 monotherapy or in combination with Dox delayed tumor growth and increased survival in the hypoxic PDX model (pOne Sentence SummaryDevelopment of non-invasive MR imaging to assess hypoxia is crucial in determining the effectiveness of TH-302 therapy and to follow response.

Radiomic features are being increasingly studied for clinical applications. We aimed to assess the agreement among radiomic features when computed by several groups by using different software packages under very tightly controlled conditions, which included standardized feature definitions and common image data sets. Ten sites (9 from the NCI's Quantitative Imaging Network] positron emission tomography–computed tomography working group plus one site from outside that group) participated in this project. Nine common quantitative imaging features were selected for comparison including features that describe morphology, intensity, shape, and texture. The common image data sets were: three 3D digital reference objects (DROs) and 10 patient image scans from the Lung Image Database Consortium data set using a specific lesion in each scan. Each object (DRO or lesion) was accompanied by an already-defined volume of interest, from which the features were calculated. Feature values for each object (DRO or lesion) were reported. The coefficient of variation (CV), expressed as a percentage, was calculated across software packages for each feature on each object. Thirteen sets of results were obtained for the DROs and patient data sets. Five of the 9 features showed excellent agreement with CV &lt, 1%, 1 feature had moderate agreement (CV &lt, 10%), and 3 features had larger variations (CV ≥ 10%) even after attempts at harmonization of feature calculations. This work highlights the value of feature definition standardization as well as the need to further clarify definitions for some features.

Background parenchymal enhancement (BPE) on multi-parametric breast magnetic resonance imaging (mpMRI) reflects the volume and intensity of gadolinium uptake. We developed a semi-automated segmentation algorithm to extract and quantify measures of BPE and investigated the agreement of computed measures of BPE with radiologist-assigned categories. We retrospectively identified 123 high risk patients with breast mpMRI performed for screening indications. Pre- and post-gadolinium T1-weighted series with and without fat suppression were co-registered. Using Otsu’s method and an active contours method, the breast tissue was segmented from the chest wall and non-fat voxels were clustered to identify the amount of fibroglandular tissue. Median and inter-quartile ranges for the absolute volume of BPE (BPE cm3) and the percentage of BPE (BPE%) using a threshold PE of 30% were computed within the segmented FGT on the first post-contrast phase. Student’s t-test was used to evaluate BPE volume and BPE% by radiologist-assigned categories. Both BPE volume and BPE% differed significantly between minimal/mild and moderate/marked radiologist assigned categories (p=0.030 and 0.004, respectively). Using our newly developed semi-automated segmentation pipeline, we quantified fibroglandular tissue and BPE. The overlapping ranges of quantitative BPE corresponding to the radiologist-assigned BPE categories showcases the inter-reader variability with visual estimation of BPE.

International audience; Background Radiomic features may quantify characteristics present in medical imaging. However, the lack of standardized definitions and validated reference values have hampered clinical use. Purpose To standardize a set of 174 radiomic features. Materials and Methods Radiomic features were assessed in three phases. In phase I, 487 features were derived from the basic set of 174 features. Twenty-five research teams with unique radiomics software implementations computed feature values directly from a digital phantom, without any additional image processing. In phase II, 15 teams computed values for 1347 derived features using a CT image of a patient with lung cancer and predefined image processing configurations. In both phases, consensus among the teams on the validity of tentative reference values was measured through the frequency of the modal value and classified as follows: less than three matches, weak; three to five matches, moderate; six to nine matches, strong; 10 or more matches, very strong. In the final phase (phase III), a public data set of multimodality images (CT, fluorine 18 fluorodeoxyglucose PET, and T1-weighted MRI) from 51 patients with soft-tissue sarcoma was used to prospectively assess reproducibility of standardized features. Results Consensus on reference values was initially weak for 232 of 302 features (76.8%) at phase I and 703 of 1075 features (65.4%) at phase II. At the final iteration, weak consensus remained for only two of 487 features (0.4%) at phase I and 19 of 1347 features (1.4%) at phase II. Strong or better consensus was achieved for 463 of 487 features (95.1%) at phase I and 1220 of 1347 features (90.6%) at phase II. Overall, 169 of 174 features were standardized in the first two phases. In the final validation phase (phase III), most of the 169 standardized features could be excellently reproduced (166 with CT; 164 with PET; and 164 with MRI). Conclusion A set of 169 radiomics features was standardized, which enabled verification and calibration of different radiomics software. © RSNA, 2020 Online supplemental material is available for this article. See also the editorial by Kuhl and Truhn in this issue.

Prostate cancer diagnosis and treatment continues to be a major public health challenge. The heterogeneity of the disease is one of the major factors leading to imprecise diagnosis and suboptimal disease management. The improved resolution of functional multi-parametric magnetic resonance imaging (mpMRI) has shown promise to improve detection and characterization of the disease. Regions that subdivide the tumor based on Dynamic Contrast Enhancement (DCE) of mpMRI are referred to as DCE-Habitats in this study. The DCE defined perfusion curve patterns on the identified tumor habitat region are used to assess clinical significance. These perfusion curves were systematically quantified using seven features in association with the patient biopsy outcome and classifier models were built to find the best discriminating characteristics between clinically significant and insignificant prostate lesions defined by Gleason score (GS). Multivariable analysis was performed independently on one institution and validated on the other, using a multi-parametric feature model, based on DCE characteristics and ADC features. The models had an intra institution Area under the Receiver Operating Characteristic (AUC) of 0.82. Trained on Institution I and validated on the cohort from Institution II, the AUC was also 0.82 (sensitivity 0.68, specificity 0.95).

Purpose Many radiomics features were originally developed for non-medical imaging applications and therefore original assumptions may need to be reexamined. In this study, we investigated the impact of slice thickness and pixel spacing (or pixel size) on radiomics features extracted from Computed Tomography (CT) phantom images acquired with different scanners as well as different acquisition and reconstruction parameters. The dependence of CT texture features on gray-level discretization was also evaluated. Methods and materials A texture phantom composed of 10 different cartridges of different materials was scanned on eight different CT scanners from three different manufacturers. The images were reconstructed for various slice thicknesses. For each slice thickness, the reconstruction Field Of View (FOV) was varied to render pixel sizes ranging from 0.39 to 0.98 mm. A fixed spherical region of interest (ROI) was contoured on the images of the shredded rubber cartridge and the 3D printed, 20% fill, acrylonitrile butadiene styrene plastic cartridge (ABS20) for all phantom imaging sets. Radiomic features were extracted from the ROIs using an in-house program. Features categories were: shape (10), intensity (16), GLCM (24), GLZSM (11), GLRLM (11), and NGTDM (5), fractal dimensions (8) and first-order wavelets (128), for a total of 213 features. Voxel-size resampling was performed to investigate the usefulness of extracting features using a suitably chosen voxel size. Acquired phantom image sets were resampled to a voxel size of 1 × 1 × 2 mm3 using linear interpolation. Image features were therefore extracted from resampled and original datasets and the absolute value of the percent coefficient of variation (%COV) for each feature was calculated. Based on the %COV values, features were classified in 3 groups: (1) features with large variations before and after resampling (%COV >50); (2) features with diminished variation (%COV

Clinical breast cancers in which at least 10% of cells express the estrogen receptor are labeled as “ER positive.” First line therapy for these patients is typically continuous administration of anti-estrogen drugs at maximum tolerated dose (MTD) until progression. In the vast majority of patients, resistance to hormone therapy evolves in the breast cancer cells within 2 years leading to treatment failure and tumor progression. In prior studies, we have demonstrated continuous application of MTD chemotherapy results in evolutionary dynamics (termed “competitive release”) that accelerates proliferation of treatment-resistance populations. In contrast, evolution-informed application of treatment reduces drug administration to maintain substantial populations of therapy-sensitive cells to reduce proliferation of resistant phenotypes. Prior pre-clinical and clinical studies have shown this strategy can delay or prevent proliferation of resistant cells and prolong time to progression (TTP). We hypothesize that similar dynamics may be observed in hormonal therapy of ER+ breast cancers. Here we address two important dynamics. First, we consider a clinical scenario in which symptoms are sufficiently severe or life-threatening to require rapid and substantial tumor reduction. Can this be achieved while retaining evolutionary dynamics to subsequently delay proliferation of resistance? A second, related question is defining the cost of resistance to anti-estrogen therapy. Here, we investigated the evolutionary dynamics of resistance to anti-estrogen therapy using ER+ MCF-7 orthotropic xenografts treated with both continuous Tamoxifen as well as cycles in which estrogen stimulation is combined with estrogen suppression. As expected, continuous administration of anti-estrogen drugs successfully suppressed tumor growth. However we found that brief interruptions in drug administration permitted equal tumor control while administering up to 50% less drug and maintaining cell phenotypes that retained high levels of ER expression and lower levels of MDR1 expression. In follow-on experiments combining hormonal and chemo-therapies; we obtained similar tumor control to hormonal therapy alone but with more necrosis and significantly lower ER expression in the surviving population.

Preclinical mouse models of lung adenocarcinoma are invaluable for the discovery of molecular drivers of tumor formation, progression, and therapeutic resistance. Histological analyses of these preclinical models require significant investments of time and training to ensure accuracy and consistency. Analysis by a clinical pathologist is the gold standard in this approach, but may be difficult to obtain due to the cost and availability of their services. As an alternative we have developed a digital pathology tool to identify, segment, grade, and analyze tumors in mouse models of lung adenocarcinoma. This convolutional neural network (CNN) model, based on ResNet18, was trained to classify normal lung tissue, normal airways, and the different grades (1 – 4) of lung adenocarcinoma from 100,000 224 × 224 pixel image patches (~16,000 patches per class). Our training dataset was constructed from whole slide images of hematoxylin and eosin stained lung sections from 4 different mouse models of lung adenocarcinoma with oncogenic Kras (KrasG12D/+), in combination with oncogenic p53 mutations (KrasG12D/+; p53R172H/+ and KrasG12D/+;p53R270H/+), or with the loss of the tumor suppressive TAp73 (KrasG12D/+;TAp7fltd/fltd). Our CNN demonstrated a strong correspondence with human pathologists on our holdout dataset, achieving a micro-F1 score of 0.81 on a pixel-by-pixel basis. As a test of our CNN, we analyzed two mouse models to better understand the role of TAp73 in lung adenocarcinoma: KrasG12D/+ (“K”) and KrasG12D/+;TAp73fltd/fltd (“TK”). Both human raters and our CNN reported a significant increase in the tumor burden of the compound mutant “TK” mice compared to the single mutant “K” mice. According to our CNN, this increased tumor burden was driven primarily by an increase in tumor size and not an increased number of tumors in “TK” mice. Because our CNN can assign different grades to regions within the same image patch and tumor, we also uncovered a high degree of intratumor heterogeneity that was not reported by the human pathologists, who are trained to assign one grade to a single tumor with a bias for the highest grade present in a given tumor. The finer grading resolution allowed our CNN to uncover the increased tumor size observed in the “TK” mice was due to expansion of Grade 2 regions (characterized by enlarged nuclei without irregular shape) within tumors that would be considered a higher grade by pathologists. Our CNN also provides a detailed map of tumor grades overlaid on the H&E images used for analysis, allowing for precise targeting of regions within tumors with other assays. We are currently utilizing these outputs in conjunction with other assays, such as spatial transcriptomic analysis and immunohistochemistry, to investigate the molecular mechanisms that underlie the expansion of Grade 2 tumor regions in “TK” mice. Future work will expand this tool into a multidimensional digital pathology pipeline that can accelerate current investigations and reveal new therapeutic targets and prognostic markers. Citation Format: John H. Lockhart, Hayley D. Ackerman, Kyubum Lee, Mahmoud Abdalah, Andrew Davis, Nicole Montey, Theresa Boyle, James Saller, Ayensur Keske, Kay Hänggi, Brian Ruffell, Olya Stringfield, Aik Choon Tan, Elsa R. Flores. Automated tumor segmentation, grading, and analysis of tumor heterogeneity in preclinical models of lung adenocarcinoma [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-082.

OBJECTIVES: Immune-checkpoint blockades have exhibited durable responses and improved long-term survival in a subset of advanced non-small-cell lung cancer (NSCLC) patients. However, highly predictive markers of positive and negative responses to immunotherapy are a significant unmet clinical need. The objective of this study was to identify clinical and computational image-based predictors of rapid disease progression phenotypes in NSCLC patients treated with immune-checkpoint blockades. MATERIALS AND METHODS: Using time-to-progression (TTP) and/or tumor growth rates, rapid disease progression phenotypes were developed including hyperprogressive disease. The pre-treatment baseline predictors that were used to identify these phenotypes included patient demographics, clinical data, driver mutations, hematology data, and computational image-based features (radiomics) that were extracted from pre-treatment computed tomography scans. Synthetic Minority Oversampling Technique (SMOTE) was used to subsample minority groups to eliminate classification bias. Patient-level probabilities were calculated from final clinical-radiomic models to subgroup patients by progression-free survival (PFS). RESULTS: Among 228 NSCLC patients treated with single agent or double agent immunotherapy, we identified parsimonious clinical-radiomic models with modest to high ability to predict progression phenotypes with area under the receiver-operator characteristics ranging from 0.812 to 0.843. Patients who experienced TTP < 2 months or hyperprogressive disease were classified with 73.4% and 82.3% accuracy after SMOTE subsampling, respectively. When the patient subgroups based on patient-level probabilities were analyzed for survival outcomes, patients with higher probability scores had significantly worse PFS. CONCLUSIONS: The models found in this study have potential important translational implications to identify highly vulnerable NSCLC patients treated with immunotherapy that experience rapid disease progression and survival poor outcomes.

PURPOSE/OBJECTIVE(S) Radiologic images have long been used clinically to assess tumor response to therapies, with a significant thrust in recent years to develop methods for predicting tumor response to therapy from radiologic images. Magnetic resonance imaging (MRI) of the apparent diffusion coefficient of water (ADC) in the tumor has been mechanistically linked to tumor cell density and viability. Post-radiation tumor ADC can arguably serve as an objective metric of tumor response to radiation treatment (RT). Here we hypothesize that the post-RT tumor ADC map can be predicted from spatially co-registered pre-RT multiparameter MRI (mpMRI) images and the delivered RT dose map. The overarching goal of this work is to provide a method to predict the RT dose map that will optimize local control. MATERIALS/METHODS We retrospectively examined 19 patients with breast cancer metastases to the brain (BCMB) who were treated with stereotactic radiation from 2013 to 2019. All patients had confirmed local recurrence. Subjects received a median dose of 21 Gy (range 15-30 Gy) in 1-5 fractions. T1-weighted unenhanced (T1W), T1-weighted contrast-enhanced (T1WCE), T2-weighted (T2W), Fluid-Attenuated Inversion Recovery (FLAIR) scans and ADC maps were acquired on all subjects before SRS, an average of 78 (15-158) days after SRS, and an average of 303 (54-831) days after SRS. Intensity-calibrated images from 11 subjects were used to train a forward model of the general form: ADCpostRT = f(RT Dose Map, T1WpreRT, T1WCEpreRT, T2WpreRT, FLAIRpreRT, ADCpreRT). RESULTS A significant univariable correlation was observed between the pixelwise change in tumor ADC (first post-SRS scan vs. pre-SRS scan) and the pixelwise RT dose (Table). A multivariable forward model with 13 fitted parameters that related ADCpostRT to an exponential function of the RT dose map with additional exponential terms involving ADCpreRT, T1WpreRT, T1W-CEpreRT, T2WpreRT, and FLAIRpreRT could predict ADCpostRT maps that were visually similar to the corresponding measured ADCpostRT maps (mean structural similarity index measure = 0.96). CONCLUSION We have demonstrated the feasibility of predicting post-RT ADC maps from pre-RT mpMRI images and the delivered RT dose maps. Further studies are underway to improve the forward model as well as to train a conditional Generative Adversarial Network (cGAN)-based inverse model. The ultimate study objective is to investigate whether the inverse model can predict the RT dose map that will yield a target post-RT ADC map in BCMB.