Jay J. Yoo

Jay J. Yoo, Khashayar Namdar, Matthias W. Wagner et al. · utoronto

Segmentation of regions of interest (ROIs) for identifying abnormalities is a leading problem in medical imaging. Using machine learning for this problem generally requires manually annotated ground-truth segmentations, demanding extensive time and resources from radiologists. This work presents a weakly supervised approach that utilizes binary image-level labels, which are much simpler to acquire, to effectively segment anomalies in 2D magnetic resonance images without ground truth annotations. We train a generative adversarial network (GAN) that converts cancerous images to healthy variants, which are used along with localization seeds as priors to generate improved weakly supervised segmentations. The non-cancerous variants can also be used to evaluate the segmentations in a weakly supervised fashion, which allows for the most effective segmentations to be identified and then applied to downstream clinical classification tasks. On the Multimodal Brain Tumor Segmentation (BraTS) 2020 dataset, our proposed method generates and identifies segmentations that achieve test Dice coefficients of 83.91%. Using these segmentations for pathology classification results with a test AUC of 93.32% which is comparable to the test AUC of 95.80% achieved when using true segmentations.

Jay J. Yoo, Khashayar Namdar, Sean Carey et al. · utoronto

Objectives: To develop and evaluate a radiomics machine learning model for detecting liver fibrosis on CT of the liver. Methods: For this retrospective, single-centre study, radiomic features were extracted from Regions of Interest (ROIs) on CT images of patients who underwent simultaneous liver biopsy and CT examinations. Combinations of contrast, normalization, machine learning model, and feature selection method were determined based on their mean test Area Under the Receiver Operating Characteristic curve (AUC) on randomly placed ROIs. The combination and selected features with the highest AUC were used to develop a final liver fibrosis screening model. Results: The study included 101 male and 68 female patients (mean age = 51.2 years $\pm$ 14.7 [SD]). When averaging the AUC across all combinations, non-contrast enhanced (NC) CT (AUC, 0.6100; 95% CI: 0.5897, 0.6303) outperformed contrast-enhanced CT (AUC, 0.5680; 95% CI: 0.5471, 0.5890). The combination of hyperparameters and features that yielded the highest AUC was a logistic regression model with inputs features of maximum, energy, kurtosis, skewness, and small area high gray level emphasis extracted from non-contrast enhanced NC CT normalized using Gamma correction with $γ$ = 1.5 (AUC, 0.7833; 95% CI: 0.7821, 0.7845), (sensitivity, 0.9091; 95% CI: 0.9091, 0.9091). Conclusions: Radiomics-based machine learning models allow for the detection of liver fibrosis with reasonable accuracy and high sensitivity on NC CT. Thus, these models can be used to non-invasively screen for liver fibrosis, contributing to earlier detection of the disease at a potentially curable stage.

Jay J. Yoo, Khashayar Namdar, Farzad Khalvati · utoronto

Training machine learning models to segment tumors and other anomalies in medical images is an important step for developing diagnostic tools but generally requires manually annotated ground truth segmentations, which necessitates significant time and resources. This work proposes the use of a superpixel generation model and a superpixel clustering model to enable weakly supervised brain tumor segmentations. The proposed method utilizes binary image-level classification labels, which are readily accessible, to significantly improve the initial region of interest segmentations generated by standard weakly supervised methods without requiring ground truth annotations. We used 2D slices of magnetic resonance brain scans from the Multimodal Brain Tumor Segmentation Challenge 2020 dataset and labels indicating the presence of tumors to train the pipeline. On the test cohort, our method achieved a mean Dice coefficient of 0.691 and a mean 95% Hausdorff distance of 18.1, outperforming existing superpixel-based weakly supervised segmentation methods.

Felicia Liu, Jay J. Yoo, Farzad Khalvati

Large Language Models (LLMs) have shown strong performance in text-based healthcare tasks. However, their utility in image-based applications remains unexplored. We investigate the effectiveness of LLMs for medical imaging tasks, specifically glioma classification and segmentation, and compare their performance to that of traditional convolutional neural networks (CNNs). Using the BraTS 2020 dataset of multi-modal brain MRIs, we evaluated a general-purpose vision-language LLM (LLaMA 3.2 Instruct) both before and after fine-tuning, and benchmarked its performance against custom 3D CNNs. For glioma classification (Low-Grade vs. High-Grade), the CNN achieved 80% accuracy and balanced precision and recall. The general LLM reached 76% accuracy but suffered from a specificity of only 18%, often misclassifying Low-Grade tumors. Fine-tuning improved specificity to 55%, but overall performance declined (e.g., accuracy dropped to 72%). For segmentation, three methods - center point, bounding box, and polygon extraction, were implemented. CNNs accurately localized gliomas, though small tumors were sometimes missed. In contrast, LLMs consistently clustered predictions near the image center, with no distinction of glioma size, location, or placement. Fine-tuning improved output formatting but failed to meaningfully enhance spatial accuracy. The bounding polygon method yielded random, unstructured outputs. Overall, CNNs outperformed LLMs in both tasks. LLMs showed limited spatial understanding and minimal improvement from fine-tuning, indicating that, in their current form, they are not well-suited for image-based tasks. More rigorous fine-tuning or alternative training strategies may be needed for LLMs to achieve better performance, robustness, and utility in the medical space.