Abstract
This paper introduces a hybrid supervision learning framework for classifying large-scale pathology Whole Slide Images (WSIs) by integrating a limited set of fine-grained pixel-level annotations with a vast quantity of coarse image-level labels. Utilizing an Expectation-Maximization (EM)-inspired approach to generate and refine patch-level pseudo-labels, the method avoids the memory constraints of end-to-end training. The framework significantly boosts specificity while maintaining the 100% sensitivity required for clinical safety, outperforming standard weak supervision baselines.
Introduction
In computational pathology, Whole Slide Images (WSIs) present a significant challenge due to their extreme resolution, often exceeding 100,000 x 100,000 pixels. This scale makes standard end-to-end deep learning impossible under current GPU memory constraints. Researchers typically resort to "weak supervision" using Multiple Instance Learning (MIL) based on image-level labels, which often results in high false-positive rates. Conversely, obtaining comprehensive pixel-level segmentation masks is prohibitively expensive. This study proposes a hybrid supervision approach to effectively combine a small set of expensive pixel-level annotations with a massive dataset of cheap image-level labels.
Methodology
The authors propose a two-stage, iterative framework inspired by Expectation-Maximization (EM) to train a patch-level segmentation model and a WSI-level classifier.
Stage 1: Patch Segmentation and Pseudo-label Refinement
The process begins by pre-training a patch segmentation model on available pixel-level labels. In the E-step, the model generates pixel-level pseudo-labels for all patches. These are refined using image-level labels: patches from negative images are designated as "hard negatives," while high-confidence patches from positive images are boosted. In the M-step, the model is retrained on a sampled batch of ground-truth labels, boosted pseudo-positives, and a heavy majority of hard negatives to suppress false-positive patterns.
Stage 2: Whole Slide Image Classification
The refined patch model evaluates all patches in a WSI. The top $K$ most suspicious patches are selected and fed into a secondary WSI classifier, which is trained using ground-truth image-level labels. The system iterates between these stages until the classification loss converges.
Experimental Results
The framework was evaluated on a Gastric Cancer Dataset containing 10,894 WSIs. Notably, only 200 patches were provided with fine-grained pixel-level labels.
The hybrid supervision framework achieved 89.32% specificity while strictly maintaining 100% sensitivity (AUC: 0.9906). This performance significantly outperformed an image-level-only baseline (80.40% specificity) and a pixel-level-only baseline (84.12% specificity). The model also demonstrated greater robustness, achieving 100% sensitivity at a much higher probability threshold (0.1000) compared to baselines.
Conclusion
By leveraging "asymmetric label trust"—treating negative slides as reliable sources of hard negatives—the proposed method effectively suppresses false positives. While the iterative training process adds computational overhead, the gains in specificity and clinical reliability represent a significant advancement for automated WSI analysis. Future work may focus on optimizing threshold sensitivity and streamlining the iterative pipeline.
Summary updated