Domain-incremental white blood cell classification with privacy-aware continual learning.

Area of Science:

• Medical diagnostics
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• White blood cell (WBC) classification is essential for diagnosing hematological conditions.
• Domain shifts in medical imaging (e.g., sample source, imaging conditions) challenge deep learning models.
• Traditional models suffer catastrophic forgetting, and foundation models degrade with distribution shifts.

Purpose of the Study:

• To develop a continual learning (CL) strategy to prevent forgetting in foundation models for WBC classification.
• To address performance degradation caused by domain shifts in dynamic clinical environments.
• To enable privacy-preserving data replay for robust model training.

Main Methods:

• Proposed a generative replay-based CL strategy for WBC classification.
• Utilized lightweight generators to create synthetic latent representations of past data.
• Conducted extensive experiments using four datasets and four backbone models (ResNet50, RetCCL, CTransPath, UNI).

Main Results:

• Conventional fine-tuning methods showed performance degradation on prior tasks and struggled with domain shifts.
• The proposed CL strategy effectively mitigated catastrophic forgetting.
• Model performance was preserved across varying data domains and task orders.

Conclusions:

• The generative replay-based CL strategy offers a practical solution for maintaining reliable WBC classification.
• This approach enhances model robustness in real-world clinical settings with evolving data distributions.
• The method ensures sustained diagnostic accuracy despite domain variations and continuous learning.