Adaptation of Weakly Supervised Localization in Histopathology by Debiasing Predictions

A collaborative research team has published a new paper, "Adaptation of Weakly Supervised Localization in Histopathology by Debiasing Predictions," introducing a method called SFDA-DeP. This technique solves a major real-world problem for AI in medicine: when a model trained to find cancer in tissue slides at one hospital fails at another due to differences in staining protocols or scanners. These Weakly Supervised Object Localization (WSOL) models, which only need image-level labels, often develop a strong bias toward predicting the most common tissue types in new data, degrading their diagnostic accuracy.

Standard adaptation methods, which use self-training, accidentally reinforce this initial bias. SFDA-DeP formulates the problem differently, using a concept inspired by machine unlearning. The algorithm works in cycles, first identifying target images where the model is over-predicting a dominant class with low confidence. It then selectively reduces the model's predictive certainty for these uncertain cases, preventing the bias from being amplified. A concurrently optimized pixel-level classifier helps restore the model's ability to pinpoint the exact location of abnormalities within the slide.

The team validated SFDA-DeP on major histopathology benchmarks including Glas and the CAMELYON datasets, which involve cross-organ and cross-center analysis. The results showed consistent improvements over state-of-the-art domain adaptation baselines for both the high-level classification task (identifying if cancer is present) and the precise localization task (outlining the cancerous region). This represents a significant step toward deploying reliable, unbiased AI diagnostic assistants that can generalize across diverse clinical settings without needing costly re-labeling of data.