Endometrial cancer is preceded by a spectrum of lesions ranging from benign changes to precancerous hyperplasia to frank malignancy. Accurate classification of these lesions on tissue biopsies is critical for guiding treatment, but visual assessment by pathologists is time-consuming and subject to variability between observers.
Digital pathology converts glass biopsy slides into gigapixel digital images called whole-slide images (WSIs). These images contain vastly more information than a pathologist can assess manually, making them prime candidates for automated analysis by deep learning algorithms.
A key obstacle is that training deep learning models typically requires detailed pixel-level annotations from expert pathologists -- an extremely expensive and labor-intensive process. Weakly supervised learning approaches aim to achieve strong performance using only slide-level labels (the final diagnosis), avoiding the need for exhaustive manual annotation.
The study used a framework called Dual-Stream Multiple Instance Learning (DSMIL), a form of weakly supervised deep learning. In this approach, each slide is treated as a collection (or "bag") of small image patches. The model learns which patches are most diagnostically relevant without being told explicitly where the disease is located.
Patches were encoded using ResNet50, a well-validated deep neural network pretrained on large natural image datasets and fine-tuned on histopathology images. An attention mechanism then assigned importance scores to each patch, allowing the model to focus on the most informative regions of the slide.
The dataset comprised 885 whole-slide images from 442 patients, classified into four categories: endometrial carcinoma (EC), atypical endometrial hyperplasia (AEH), endometrial hyperplasia (EH), and normal endometrium (NE). This four-class structure reflects the clinical spectrum from healthy tissue to cancer.
To improve feature extraction, the researchers incorporated automated cell segmentation using two advanced tools: nnU-Net v2 for nuclei segmentation and HoverNet for simultaneous nuclear detection and classification by cell type.
By identifying individual cells and their types -- including epithelial cells, stromal cells, and inflammatory cells -- the model could extract richer, biologically meaningful features from each patch rather than relying solely on raw pixel intensities.
This combination of tissue-level patch analysis and cell-level segmentation represents a multi-scale approach, mimicking how an expert pathologist simultaneously considers overall tissue architecture and individual cellular details when making a diagnosis.
The DSMIL model achieved an impressive area under the ROC curve (AUROC) of 0.9776 -- very close to perfect -- for distinguishing among the four lesion categories. Overall classification accuracy was 89.14%, indicating strong agreement with expert pathological diagnoses.
The model performed best at identifying frank endometrial carcinoma and normal endometrium, which represent the two ends of the diagnostic spectrum. Distinguishing the intermediate categories -- atypical hyperplasia from non-atypical hyperplasia -- was harder, consistent with known challenges in human pathology.
Attention heatmaps generated by the model highlighted diagnostically relevant regions such as glandular crowding, cellular atypia, and stromal invasion. These maps correspond closely to the features pathologists use for diagnosis, suggesting the model has learned clinically meaningful patterns.
A system achieving near-expert accuracy with only slide-level labels has significant practical value. It could serve as a screening tool to flag high-priority cases, reduce pathologist workload, and provide a second opinion for borderline diagnoses.
The attention-based visualization is particularly valuable for clinical adoption because it provides explainability -- pathologists can see which tissue regions drove the model's decision, allowing them to validate or override the AI assessment rather than treating it as a black box.
Weakly supervised models are also more scalable than fully supervised alternatives, since generating slide-level labels from existing pathology reports requires no additional annotation work. This makes deploying such models in resource-limited settings more feasible.
This study demonstrates that weakly supervised deep learning can classify endometrial lesions across the full spectrum from normal to cancer with high accuracy, without requiring exhaustive manual annotation of training data.
The integration of multi-scale analysis -- combining tissue patch features with cell-level segmentation -- proved effective at capturing the diagnostic complexity of endometrial histopathology. Future work could explore larger multi-center datasets and integration with molecular data to further improve performance.
As digital pathology infrastructure expands globally, AI tools like DSMIL have the potential to standardize diagnostic quality, reduce inter-observer variability, and ultimately improve patient outcomes through faster and more accurate diagnosis.