Mitigating catastrophic forgetting in Multiple sclerosis lesion segmentation using elastic weight consolidation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate multiple sclerosis (MS) lesion segmentation is vital for tracking disease progression.
• Deep learning models face domain shift challenges, impacting performance across different datasets.
• Standard transfer learning (TL) can cause catastrophic forgetting, degrading performance on original data.

Purpose of the Study:

• To apply Elastic Weight Consolidation (EWC) for domain-incremental learning in MS lesion segmentation.
• To evaluate EWC's effectiveness in mitigating catastrophic forgetting during domain adaptation.
• To assess EWC's performance in few-shot learning scenarios for MS lesion segmentation.

Main Methods:

• Utilized a 3D U-Net architecture for MS lesion segmentation.
• Trained models on public datasets (WMH2017, Shifts) and fine-tuned on an in-house dataset.
• Compared standard transfer learning (TL) with Elastic Weight Consolidation (EWC) in full and few-shot training.

Main Results:

• EWC achieved a 10% F-score improvement with only 3 target domain images.
• Five target domain images with EWC yielded results comparable to full dataset training.
• EWC reduced catastrophic forgetting by 8%-19% compared to TL, with performance drops of 20-37%.

Conclusions:

• EWC effectively enables deep learning models to adapt to new domains while retaining prior knowledge.
• This approach significantly reduces catastrophic forgetting in MS lesion segmentation.
• EWC demonstrates potential for developing more generalizable deep learning models for clinical MS applications with minimal data needs.