Search

When an auto-segmentation model needs to be applied to a new segmentation task, multiple decisions should be made about the pre-processing steps and training hyperparameters. These decisions are cumbersome and require a high level of expertise. To remedy this problem, I developed self-configuring CapsNets (scCapsNets) that can scan the training data as well as the computational resources that are available, and then self-configure most of their design options. In this study, we developed a self-configuring capsule network that can configure its design options with minimal user input. We showed that our self-configuring capsule netwrok can segment brain tumor components, namely edema and enhancing core of brain tumors, with high accuracy. Out model outperforms UNet-based models in the absence of data augmentation, is faster to train, and is computationally more efficient compared to UNet-based models.

Deep learning (DL) models have demonstrated state-of-the-art performance in the classification of diagnostic imaging in oncology. However, DL models for medical images can be compromised by adversarial images, where pixel values of input images are manipulated to deceive the DL model. To address this limitation, our study investigates the detectability of adversarial images in oncology using multiple detection schemes. Experiments were conducted on thoracic computed tomography (CT) scans, mammography, and brain magnetic resonance imaging (MRI). For each dataset we trained a convolutional neural network to classify the presence or absence of malignancy. We trained five DL and machine learning (ML)-based detection models and tested their performance in detecting adversarial images. Adversarial images generated using projected gradient descent (PGD) with a perturbation size of 0.004 were detected by the ResNet detection model with an accuracy of 100% for CT, 100% for mammogram, and 90.0% for MRI. Overall, adversarial images were detected with high accuracy in settings where adversarial perturbation was above set thresholds. Adversarial detection should be considered alongside adversarial training as a defense technique to protect DL models for cancer imaging classification from the threat of adversarial images.

In this work we introduce CUTS (Contrastive and Unsupervised Training for Segmentation), the first fully unsupervised deep learning framework for medical image segmentation to better utilize the vast majority of imaging data that is not labeled or annotated. Segmenting medical images is a critical task for facilitating both patient diagnoses and quantitative research. A major limiting factor is the lack of labeled data, as obtaining expert annotations for each new set of imaging data or task can be expensive, labor intensive, and inconsistent across annotators. Thus, we utilize self-supervision from pixels and their local neighborhoods in the images themselves. Our unsupervised approach optimizes a training objective that leverages concepts from contrastive learning and autoencoding. In contrast to prior work, our framework segments medical images with a novel two-stage approach without relying on any labeled data at any stage. The first stage involves the creation of a "pixel-centered patch" that embeds every pixel along with its surrounding patch, using a vector representation in a high-dimensional latent embedding space. The second stage utilizes diffusion condensation, a multi-scale topological data analysis approach, to dynamically coarse-grain these embedding vectors at all levels of granularity. The final outcome is a series of coarse-to-fine segmentations that highlight image structures at various scales. We show successful multi-scale segmentation on natural images, retinal fundus images, and brain MRI images. Our framework delineates structures and patterns at different scales which may carry distinct information relevant to clinical interpretation. As we tackle the problem of segmenting medical images at multiple meaningful granularities without relying on any label, we demonstrate the possibility to circumvent tedious and repetitive manual annotations in future practice., Comment: Updated Table 1. Rerun supervised and STEGO after unified codebase. Correct bug in multi-class dice and hausdorff

This study aimed to assess if quantitative diffusion magnetic resonance imaging analysis would improve prognostication of individual patients with severe traumatic brain injury.We analyzed images of 30 healthy controls to extract normal fractional anisotropy ranges along 18 white-matter tracts. Then, we analyzed images of 33 patients, compared their fractional anisotropy values with normal ranges extracted from controls, and computed severity of injury to white-matter tracts. We also asked 2 neuroradiologists to rate severity of injury to different brain regions on fluid-attenuated inversion recovery and susceptibility-weighted imaging. Finally, we built 3 models: (1) fed with neuroradiologists' ratings, (2) fed with white-matter injury measures, and (3) fed with both input types.The 3 models respectively predicted survival at 1 year with accuracies of 70%, 73%, and 88%. The accuracy with both input types was significantly better (P0.05).Quantifying severity of injury to white-matter tracts complements qualitative imaging findings and improves outcome prediction in severe traumatic brain injury.

Deep-learning methods for auto-segmenting brain images either segment one slice of the image (2D), five consecutive slices of the image (2.5D), or an entire volume of the image (3D). Whether one approach is superior for auto-segmenting brain images is not known.We compared these three approaches (3D, 2.5D, and 2D) across three auto-segmentation models (capsule networks, UNets, and nnUNets) to segment brain structures. We used 3430 brain MRIs, acquired in a multi-institutional study, to train and test our models. We used the following performance metrics: segmentation accuracy, performance with limited training data, required computational memory, and computational speed during training and deployment.3D, 2.5D, and 2D approaches respectively gave the highest to lowest Dice scores across all models. 3D models maintained higher Dice scores when the training set size was decreased from 3199 MRIs down to 60 MRIs. 3D models converged 20% to 40% faster during training and were 30% to 50% faster during deployment. However, 3D models require 20 times more computational memory compared to 2.5D or 2D models.This study showed that 3D models are more accurate, maintain better performance with limited training data, and are faster to train and deploy. However, 3D models require more computational memory compared to 2.5D or 2D models.

Recent studies demonstrate that low-level light therapy (LLLT) modulates recovery in patients with traumatic brain injury (TBI). However, the impact of LLLT on brain activity following TBI has not been well described. Here we use a randomized, double-blind, placebo-controlled design to investigate the effect of LLLT on resting-state connectivity at acute (within 1-week), subacute (2–3 weeks), and late-subacute (3-month) time-points following moderate TBI. A characteristic connectivity profile was observed during TBI recovery in both sham- (n = 21) and LLLT-treated patients (n = 17) compared to healthy controls, with increased resting-state connectivity between frontal and parietal cortices. Temporal comparisons between LLLT- and sham-treated patients showed that the acute-to-subacute changes in resting-state connectivity were significantly greater in LLLT-treated patients. These results demonstrate that LLLT increased resting-state connectivity in the presence of a regional hyperconnectivity response to moderate TBI, suggesting that LLLT can modulate activity in the injured brain and encouraging its further exploration as a therapy for TBI.

PURPOSE Translation of AI algorithms into clinical practice is significantly limited by lack of large individual hospital-based datasets with expert annotations. Current methods for generation of annotated imaging data are significantly limited due to inefficient imaging data transfer, complicated annotation software, and time required for experts to generate ground truth information. We incorporated AI tools for auto-segmentation of gliomas into PACS that is used at our institution for reading clinical studies and developed a workflow for annotation of images and development of volumetric segmentations in neuroradiology clinical workflow. Material: 1990 patients from Yale Radiation Oncology Registry (2012-2019) were identified. Segmentations were performed using a UNETR algorithm trained on BRaTS 2021 and an internal dataset of manually segmented tumors. Segmentations were validated by a board-certified neuro-radiologist and natively embedded PyRadiomics in PACS was used for feature extraction. RESULTS In 7 Months (05/2021 - 08/2021, 03/2022 - 05/2022) segmentations and annotations were performed in 835 patients (322 female, 467 male, 46 unknown, mean age 53 yrs). Dataset includes 275 Grade 4 Gliomas (54 Grade 3, 100 Grade 2, 31 Grade 1, 375 unknown). Molecular subtypes include IDH (113 mutated, 498 wildtype, 2 Equivocal, 222 unknown), 1p/19q (87 deleted or co-deleted, 122 intact, 626 unknown), MGMT promotor (182 methylated, 95 partially methylated, 275 unmethylated, 283 unknown), EGFR (76 amplified, 177 not amplified, 582 unknown), ATRX (40 mutated, 157 retained, 638 unknown), Ki-67 (616 known, 219 unknown) and p53 (549 known, 286 unknown). Classification of gliomas between grade 3/4 and grade 1/2, yielded AUC of 0.85. CONCLUSION We developed a method for incorporation of volumetric segmentation, feature extraction, and classification that is easily incorporated into neuroradiology workflow. These tools allowed us to annotate over 100 gliomas per month, thus establishing a proof of concept for rapid development of annotated imaging database for AI applications.

PURPOSECancer registries are important sources of real-world data (RWD) that reveal insights into practice patterns and cancer patient outcomes, but the prevalence of missing data can be high. Machine learning (ML) imputation methods can be applied to large RWD sets, but the performance of these approaches within cancer registries is unclear.METHODSWe identified non-small cell lung cancer (NSCLC) patients within the National Cancer Database diagnosed in 2014 with complete data in 19 variables of known clinical and prognostic significance. We generated synthetic missing data for each variable, then performed imputation using substitution (control) and five different ML approaches. Imputation efficacy was measured by normalized root-mean-square error (RMSE) for continuous variables and proportion of falsely classified entries (PFC) for categorical variables. We also measured algorithm runtimes and the impact of incorporating imputed values on survival modeling.RESULTS50,790 NSCLC patients were included for this study, with 81 features for each patient after data preprocessing. Among the tested ML methods, SoftImpute had the lowest RMSE (best performance) for continuous variables ranging from 0.071 to 0.080 for 10% to 50% missing data, and MissForest had the lowest PFC (best performance) for categorical variables ranging from 0.251 to 0.311 for 10 to 50% missing data. SoftImpute had a runtime of 3.28×10−4 seconds per patient record, and MissForest averaged 2.96×10−3 seconds per patient record. Deep learning imputation using a denoising autoencoder did not achieve improved performance despite higher algorithm runtimes. Cox models incorporating ML imputed data achieved similar C-index ranging from 0.787 to 0.801 for all ML methods tested.CONCLUSIONML imputation achieved promising performance for NSCLC patients within a large national cancer registry.

Background Echocardiography plays a central role in the diagnosis of infective endocarditis (IE). In recent years, additional imaging techniques have begun to challenge the conventional approach. We present a case where the use of transthoracic/transoesophageal echocardiography (TTE/TOE) in suspected IE failed to identify an extensive periannular abscess, later identified by 18F-flurodeoxyglucose–positron emission tomography (FDG–PET), requiring urgent intervention. Case summary A 69-year-old man with symptomatic Streptococcus sanguinis bacteraemia and a bicuspid aortic valve was found to have new-onset left bundle branch block that progressed to complete heart block. After starting on IV Penicillin G and having a temporary pacemaker inserted, his clinical condition improved. Transthoracic echocardiography and TOE showed no evidence of abscess. However, persistent first-degree atrioventricular block raised clinical suspicion of a possible extended infection. Subsequent FDG–PET revealed focal activity around the aortic root that extended inferiorly into the interatrial septum, consistent with active infection and possible abscess. Composite aortic root replacement with insertion of a mechanical prosthesis was carried out, revealing extensive IE and multiple periannular abscesses. Discussion As guidelines grapple with evolving understandings of how best to define the optimal imaging approach for the management of complicated IE, the results of this case clearly show the importance of heightened clinical suspicion and need for prompt operative intervention when faced with patients who present with predisposing conditions and concern for advanced conduction disease. Clinicians and researchers are encouraged to learn from the potential near-miss of an extensive periannular abscess to help guide guideline-development of imaging in complicated IE and prevent adverse outcomes in patients with similar presentations.

PURPOSE Monitoring metastatic disease to the brain is laborious and time-consuming, especially in the setting of multiple metastases and when performed manually. Response assessment in brain metastases based on maximal unidimensional diameter as per the RANO-BM guideline is commonly performed1, however, accurate volumetric lesion estimates can be crucial for clinical decision-making2 and enhance outcome prediction3. We propose a deep learning (DL)-based auto-segmentation approach with the potential for improvement of time-efficiency, reproducibility and robustness against inter-rater variability. Materials and METHODS We retrospectively retrieved 259 patients with a total number of 916 lesions from our institutional database from 2014 - 2021. Patients with prior history of local radiation therapy or surgery were excluded. Manually generated trainee segmentations were revised and adjusted by a board-certified radiologist and served as ground truth for evaluation of segmentation accuracy. Model performance was tested via dice-similarity-coefficient (DSC). Volumetric measurements were then obtained within our PACS-integrated workflow on Visage 7 (Visage Imaging, Inc., San Diego, CA) at the click of one button. RESULTS For model training and evaluation, a train-test split of 90:10 on patient-level was performed (n= 234:25 (Patients), n= 861:55 (Lesions). A DL-algorithm (nnUNet) was incrementally trained on 10 batches of 23 patients. The DSC of the U-Net gradually increased throughout the training process and heuristically reached a plateau of 0.85. The sensitivity of the algorithm was 83% with detection of 46 out of 55 lesions in the testing dataset. The lesions that were not detected by the algorithm were below 5 mm in size. The false positive rate was 8% (n=4/50). CONCLUSION Our study demonstrates the feasibility of PACS-based integration of automatized segmentation workflows of brain metastases. An incremental-training approach is recommended to adapt DL algorithms to specific hospital settings.

Key Points Question Is near-infrared low-level light therapy (LLLT) feasible and safe after moderate traumatic brain injury, and does LLLT affect the brain and exhibit neuroreactivity? Findings In this randomized clinical trial including 68 patients with moderate traumatic brain injury who were randomized to receive LLLT or sham therapy, 28 patients completed at least 1 LLLT session without any reported adverse events. In the late subacute stage, there were statistically significant differences in the magnetic resonance imaging–derived diffusion parameters of the white matter tracts between the sham- and light-treated groups, demonstrating neuroreactivity of LLLT. Meaning The results of this clinical trial show that transcranial LLLT is feasible, safe, and affects the brain in a measurable manner., Importance Preclinical studies have shown that transcranial near-infrared low-level light therapy (LLLT) administered after traumatic brain injury (TBI) confers a neuroprotective response. Objectives To assess the feasibility and safety of LLLT administered acutely after a moderate TBI and the neuroreactivity to LLLT through quantitative magnetic resonance imaging metrics and neurocognitive assessment. Design, Setting, and Participants A randomized, single-center, prospective, double-blind, placebo-controlled parallel-group trial was conducted from November 27, 2015, through July 11, 2019. Participants included 68 men and women with acute, nonpenetrating, moderate TBI who were randomized to LLLT or sham treatment. Analysis of the response-evaluable population was conducted. Interventions Transcranial LLLT was administered using a custom-built helmet starting within 72 hours after the trauma. Magnetic resonance imaging was performed in the acute (within 72 hours), early subacute (2-3 weeks), and late subacute (approximately 3 months) stages of recovery. Clinical assessments were performed concomitantly and at 6 months via the Rivermead Post-Concussion Questionnaire (RPQ), a 16-item questionnaire with each item assessed on a 5-point scale ranging from 0 (no problem) to 4 (severe problem). Main Outcomes and Measures The number of participants to successfully and safely complete LLLT without any adverse events within the first 7 days after the therapy was the primary outcome measure. Secondary outcomes were the differential effect of LLLT on MR brain diffusion parameters and RPQ scores compared with the sham group. Results Of the 68 patients who were randomized (33 to LLLT and 35 to sham therapy), 28 completed at least 1 LLLT session. No adverse events referable to LLLT were reported. Forty-three patients (22 men [51.2%]; mean [SD] age, 50.49 [17.44] years]) completed the study with at least 1 magnetic resonance imaging scan: 19 individuals in the LLLT group and 24 in the sham treatment group. Radial diffusivity (RD), mean diffusivity (MD), and fractional anisotropy (FA) showed significant time and treatment interaction at 3-month time point (RD: 0.013; 95% CI, 0.006 to 0.019; P, This randomized clinical trial evaluates the effect of near-infrared low-level light therapy in patients with traumatic brain injury from the time of injury until 3 months after the injury.