SplitFed-CL: New AI Framework Fixes Medical Image Segmentation with Noisy Labels
A co-learning method uses a global teacher to correct unreliable labels in medical imaging.
Medical image segmentation models often suffer from heterogeneous label quality across different clinical sites, especially when using federated learning for privacy. SplitFed-CL addresses this by combining split federated learning (which reduces client-side computation) with a co-learning paradigm. The framework features a global teacher model that guides local student models to identify and correct unreliable annotations. Reliable labels are used for direct training, while unreliable ones are refined via a weighted student-teacher refinement process. Additionally, consistency regularization improves robustness to input perturbations, and a trainable weighting module dynamically balances multiple loss terms.
To better simulate real-world annotation errors, the authors introduce a novel difficulty-guided strategy that mimics human-like boundary-centric mistakes, where perturbation levels depend on shape complexity. SplitFed-CL was evaluated on two multiclass segmentation datasets with controlled synthetic noise and one binary dataset containing genuine annotation errors. The framework consistently beat seven state-of-the-art baselines, achieving higher segmentation quality and stronger robustness. This work offers a practical solution for training accurate medical image segmentation models across institutions without sacrificing data privacy.
- Combines split federated learning with co-learning to preserve data privacy while reducing client-side computation.
- Uses a global teacher model to detect and refine unreliable labels, with a trainable weighting module for adaptive loss balancing.
- Outperforms seven state-of-the-art baselines on multiclass segmentation with synthetic noise and binary segmentation with real-world label errors.
Why It Matters
Enables more accurate medical image analysis across institutions while preserving data privacy and handling real-world label noise.