An iterative confidence-driven discrepancy-aware federated weakly supervised model for spine segmentation in 3D medical images

Abstract

Spinal disorders can cause neurological dysfunction and disability, with diagnosis relying on medical image segmentation. Traditional deep learning requires costly pixel-level annotations. Weakly supervised semantic segmentation with image-level labels reduces this cost, but CNNs struggle with long-range dependencies, leading to under-activated class activation maps (CAMs). Vision Transformers (ViTs) partly address this but cause over-activation. Data privacy constraints hinder large-scale dataset construction, and Federated Learning (FL) introduces Non-IID challenges. We propose ICDE, a federated weakly supervised framework addressing over-activation and Non-IID issues. A Progressive Reliable Mask Refinement (PRMR) module expands pseudo-labels from high-confidence regions to alleviate over-activation. A Class Prototype Optimization (CPO) module enhances PRMR via prototype–patch similarity. For Non-IID FL, a Confidence Discrepancy-aware Pseudo-label Correction (CDPC) mechanism balances global–local models for reliable pseudo-labels. Our framework achieves state-of-the-art performance on CTSpine1K, MRSpineSeg2021, and SPIDER2023, demonstrating strong robustness in addressing the complex and nonlinear characteristics of pathological spinal anatomies.