Pathologists examine tumor tissue under the microscope to classify cancer - but the molecular information encoded in those same tissue images extends far beyond what human eyes can perceive. Tumor mutational burden (TMB), the total number of mutations in a cancer cell's DNA, is one such hidden feature: high TMB tumors look similar to low TMB tumors to the naked eye, yet high TMB predicts better response to immunotherapy.
For endometrial cancer, classifying tumors as aggressive vs non-aggressive guides decisions about adjuvant treatment - whether to add chemotherapy or radiation after surgery. Current classification requires expensive molecular testing, but if deep learning could predict molecular class from routinely-collected H&E stained slides, classification could be done faster and cheaper at any institution with a digital pathology scanner.
This study developed TR-MAMIL - Truncated ResNet-based Multilayer Attention Multiple Instance Learning - a deep learning framework specifically designed for the challenges of analyzing gigapixel whole slide images (WSIs) for EC subtyping and TMB prediction.
Whole slide images are enormous - a single scanned slide contains billions of pixels - making it computationally impossible to analyze them as a single image. Multiple Instance Learning (MIL) solves this by dividing each slide into thousands of small patches and treating the entire collection of patches as a 'bag', where only the bag-level label (e.g., aggressive vs non-aggressive) is known during training.
TR-MAMIL's key innovation is its multilayer attention mechanism, which learns to identify which patches within the slide are most informative for classification. Rather than treating all patches equally, the model assigns attention scores to each patch - effectively learning which microscopic regions are diagnostically relevant, analogous to a pathologist knowing to focus on specific tissue areas.
The architecture uses a Truncated ResNet as its backbone - a modified version of the ResNet deep learning architecture that was pre-trained on natural images and then fine-tuned on pathology images. The truncation removes later layers to create a feature extractor optimized for the scale of features relevant in pathology tiles, balancing between computing cost and feature richness.
The study used 918 whole slide images from 529 endometrial cancer patients in The Cancer Genome Atlas (TCGA) - one of the largest public repositories of cancer tissue slides with matched molecular data. This scale of training data is essential for deep learning models, which require thousands of examples to learn robust features.
Ground truth labels for aggressive vs non-aggressive EC subtype were derived from molecular testing data: p53abn and high-grade serous tumors were classified as aggressive, while POLEmut, MMRd, and NSMP low-grade endometrioid tumors were classified as non-aggressive. TMB values were obtained from mutation calling on whole-exome sequencing data matched to each patient's tissue slide.
Performance was evaluated using 5-fold cross-validation and compared against seven state-of-the-art deep learning methods for WSI analysis. Survival analysis using Kaplan-Meier curves assessed whether the model's predicted subtype correlated with actual patient survival, providing biological validation beyond classification metrics.
For EC subtype classification (aggressive vs non-aggressive), TR-MAMIL achieved an AUROC of 88% - indicating that when the model assigns a higher probability of aggressive class to a patient, that patient is truly aggressive 88% of the time compared to random assignment. This performance level is approaching clinical utility for a screening application.
For TMB prediction in aggressive EC, the model achieved an AUROC of 82%, demonstrating that the morphological features visible in H&E slides do correlate meaningfully with the underlying mutation burden. For TP53 mutation prediction, AUROC reached 78% - again showing that a specific molecular alteration leaves visual traces detectable by deep learning.
TR-MAMIL outperformed all 7 comparison methods on the EC subtype classification task. The survival analysis confirmed that patients classified as aggressive by the model had significantly shorter survival than those classified as non-aggressive, validating that the model's classifications reflect true biological differences rather than statistical artifacts.
The most immediate clinical application is as a triage tool before molecular testing. In institutions with limited molecular testing capacity, TR-MAMIL could flag cases as high-probability aggressive (requiring urgent molecular workup and treatment planning) vs low-probability aggressive, helping prioritize testing queue and expedite care for high-risk patients.
TMB prediction from slides is particularly valuable because high TMB (typically defined as more than 10 mutations per megabase) is an FDA-approved biomarker for pembrolizumab (Keytruda) eligibility across solid tumors. Currently, TMB measurement requires specialized next-generation sequencing that is not available at all oncology centers. A slide-based surrogate, even imperfect, could extend immunotherapy access to underserved settings.
The model's attention maps - visualizations showing which tissue regions the model focused on - align with known pathological features of aggressive EC, such as areas with abnormal nuclei, high mitotic activity, and architectural disruption. This interpretability builds pathologist trust and suggests the model is learning clinically meaningful features rather than spurious patterns.
This study makes two primary contributions: first, the development of the TR-MAMIL architecture with its multilayer attention mechanism, which outperforms existing methods for EC WSI analysis; second, the demonstration that TMB and TP53 status - molecular features not previously thought to be visible in histology - can be predicted from H&E images with clinically meaningful accuracy.
Limitations include the single-institution training dataset (TCGA is not fully representative of all clinical settings), the binary classification approach (aggressive vs non-aggressive), and the need for additional validation in prospective clinical settings. External validation on slides from non-TCGA institutions with different staining protocols and scanner types will be crucial before clinical deployment.
The broader implication is that routine histopathology slides contain far more molecular information than was previously appreciated. As deep learning methods mature and training datasets grow, slide-based prediction of the full molecular profile of a tumor - without any additional laboratory testing - may become a practical clinical reality for endometrial cancer and many other tumor types.