Continual Learning for Histopathology Image Classification in Class Incremental Learning

Yuanyuan Wu, Yu Zhao, Jun Bai, Anca L. Ralescu

Keywords: Continual Learning, Catastrophic Forgetting, Prompt-based CL Methods, Histopathology Images.

TL;DR: Our paper mainly focuses on verifying the application of continual leanring methods on histopathology images under class incremental learning scenario.

Abstract: Continual Learning (CL) is increasingly important for developing adaptive clinical AI models; however, its application to histopathology remains challenging due to strict privacy constraints, expanding diagnostic categories, and substantial staining variability. In this work, we investigate CL for histopathology image classification under a realistic Class-Incremental Learning (CIL) scenario using the NCT-CRC-HE-100K dataset. We benchmark representative regularization-based, replay-based, architecture-based, and prompt-based CL methods to provide a comprehensive evaluation of existing approaches for digital pathology. Among these, prompt-based CL methods have recently emerged as a promising direction by leveraging a frozen pretrained backbone and lightweight learnable prompts to mitigate Catastrophic Forgetting (CF) without storing exemplars or requiring task identifiers during inference. To understand how these methods perform under practical constraints, we analyze the impact of exemplar-free requirements, limited buffer sizes, and training-time budgets across CL paradigms. We further compare four commonly used normalization strategies and find that dataset-level normalization consistently yields the strongest performance. Our results show that replay-based methods achieve the highest accuracy when sufficient memory and training time are available, while prompt-based methods provide competitive exemplar-free performance, making them a practical option for continual adaptation in privacy-sensitive digital pathology workflows.

Primary Subject Area: Application: Histopathology

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Reproducibility: https://github.com/AILabLLL/CLMedical

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 41