Noise-robust iterative knowledge distillation for MIL-based weakly supervised histopathology segmentation

Accurate tumor segmentation in histopathology remains challenging when only image-level labels are available. Multiple instance learning (MIL) can provide more localized and boundary-aware segmentation cues than conventional class activation mapping (CAM), but MIL-derived pseudo-masks are often noisy and may amplify errors when directly used as dense supervision. To address this limitation, we propose IFKD, an iterative fusion-knowledge distillation framework for MIL-based weakly supervised histopathology segmentation. IFKD uses MIL predictions to initialize the teacher and progressively refines the pseudo-supervision through multi-level output-guided distillation, weighted cross-entropy regularization, and scheduled teacher–student alternation. In each distillation cycle, the frozen teacher provides a stable fused segmentation prediction to supervise multi-level student outputs, while the optimized student is promoted to guide the next cycle. On Camelyon16, IFKD achieves 85.6 ± 0.1 F1, 80.0 ± 0.1 IoU, and 19.0 ± 0.4 HD, improving the VGG-MIL baseline by +3.6 F1, +5.0 IoU, and reducing HD by 8.2. On DigestPath 2019, IFKD achieves 85.4 ± 0.1 F1, 77.7 ± 0.1 IoU, and 14.0 ± 0.8 HD, improving VGG-MIL by +6.0 F1, +7.4 IoU, and reducing HD by 10.0. The framework also consistently improves ResNet-MIL, Swin-MIL, and SA-MIL, demonstrating its general applicability as a plug-in refinement framework across MIL backbones and achieving the best performance among the compared weakly supervised methods.