Federated Cross-Incremental Self-Supervised Learning for Medical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated cross learning excels in medical image segmentation but suffers from catastrophic forgetting due to data heterogeneity.
• The pixelwise label deficiency problem exacerbates forgetting in federated learning settings.

Purpose of the Study:

• To introduce FedCSL, a federated cross-incremental self-supervised learning method for medical image segmentation.
• To enable incremental learning across clients without knowledge forgetting, while preserving data privacy and minimizing labeled data requirements.

Main Methods:

• A cross-incremental collaborative distillation (CCD) mechanism using secure multiparty computation (MPC) to transfer knowledge between clients.
• A retrospect mechanism to optimize client training sequences and enhance interclient knowledge propagation.
• A two-stage training framework: federated self-supervised pretraining via masked image modeling (MIM) followed by supervised fine-tuning for segmentation tasks.

Main Results:

• FedCSL effectively enables incremental learning across clients without catastrophic forgetting.
• The method significantly reduces the need for large-scale, densely annotated medical datasets.
• Experimental results on public datasets show superior quantitative and qualitative performance compared to state-of-the-art methods.

Conclusions:

• FedCSL offers an effective solution to catastrophic forgetting and label deficiency in federated medical image segmentation.
• The proposed method enhances privacy, reduces data annotation burdens, and achieves high performance through collaborative distillation and self-supervised learning.