Reproduce “Continual Learning for Abdominal Multi-Organ and Tumor Segmentation” on OpenResearch

About this paper

Yixiao Zhang, Xinyi Li, Huimiao Chen, Alan Yuille, Yaoyao Liu, Zongwei Zhou

A framework for continual multi-organ and tumor segmentation in medical imaging leverages pseudo-labeling and a novel architecture with image-aware, class-specific segmentation heads, overcoming catastrophic forgetting without requiring past data. This approach demonstrates superior performance over existing baselines, achieving a Dice score of 0.787 on old classes and 0.636 on new classes in a three-step learning process on the JHH dataset, while maintaining computational efficiency.

Key insights

• Pseudo-labeling, even without complex distillation, can effectively preserve knowledge of old classes during continual learning when direct access to past data is restricted.
• Designing independent, image-aware, class-specific output heads on a shared backbone significantly reduces interference between learned classes, thereby mitigating catastrophic forgetting.
• Leveraging semantic information from pre-trained models like CLIP can provide effective guidance for generating class-specific parameters, enhancing the model's ability to adapt to novel classes.

Summary by alphaXiv.