HiAdapter: Histopathology-induced Adapter for Pathology Foundation Models.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning in oncology

Background:

• Pathology foundation models require efficient fine-tuning for specific tasks.
• Current methods lack generalization to varied histopathological images, including new cancers and stains, due to stain variability and tissue complexity.

Purpose of the Study:

• To develop an efficient and generalizable fine-tuning strategy for pathology foundation models.
• To address the limitations of task-agnostic fine-tuning in histopathology.

Main Methods:

• Introduced Histopathology-induced Adapter (HiAdapter) with Stain-invariant (S-Adapter) and Morphology-aware (M-Adapter) components.
• Developed Pathology Prototypical Contrastive Loss (PPCLoss) to improve feature discriminability.
• Evaluated HiAdapter on three foundation models across six benchmarks, including unseen cancers and stains.

Main Results:

• HiAdapter demonstrated significant improvements in efficiency and accuracy across multiple benchmarks.
• Achieved an average improvement of 2.15 in F1 score and 1.55 in accuracy over the second-best method.
• Showcased superior generalizability on independent datasets and WSI-level survival analysis.

Conclusions:

• HiAdapter effectively bridges low-level optical properties and high-level tissue semantics in histopathology.
• The method offers strong biological and diagnostic interpretability, with potential for patient-level diagnosis and prognosis.
• HiAdapter represents a significant advancement in fine-tuning pathology foundation models for real-world clinical applications.