Continual Nuclei Segmentation via Prototype-Wise Relation Distillation and Contrastive Learning.

Area of Science:

• Computational pathology
• Computer vision
• Artificial intelligence

Background:

• Deep learning excels at multi-type nuclei segmentation but struggles with catastrophic forgetting when learning new classes incrementally.
• Class-incremental continual learning aims to update models with new data without access to previous datasets, a critical challenge in dynamic environments.

Purpose of the Study:

• To propose a novel continual nuclei segmentation method that mitigates catastrophic forgetting and improves learning of new classes.
• To enable models to incrementally learn new nuclei types without compromising performance on previously learned classes.

Main Methods:

• Feature-level knowledge distillation incorporating prototype-wise relation distillation to maintain inter-class relation similarity for old classes.
• Prototype-wise contrastive learning with hard sampling to enhance intra-class compactness and inter-class separability of features.
• Incremental updating of models for new nuclei classes without accessing prior training data.

Main Results:

• The proposed method effectively avoids forgetting knowledge of old nuclei classes while facilitating the learning of new ones.
• Experiments on MoNuSAC and CoNSeP benchmarks show superior performance compared to existing competitive continual learning methods.
• The approach demonstrated improved performance on both previously learned and newly introduced nuclei classes.

Conclusions:

• The developed continual learning strategy offers a robust solution for nuclei segmentation in dynamic settings where new cell types emerge.
• The combination of prototype-wise relation distillation and contrastive learning is effective in preserving and enhancing feature representations for continual nuclei segmentation.