Selective Prediction Forgetting for Adaptation of Weakly Supervised Localization of Histology Images
The method SFDA-DeP addresses prediction bias in weakly supervised object localization (WSOL) during domain adaptation for histopathology images, where cross-domain shifts occur. It iteratively identifies and corrects this bias by selectively penalizing predictions for dominant classes while retaining uncertain samples. This process refines decision boundaries and improves both classification accuracy and pixel-level localization under severe distribution shifts.