Lightweight self supervised learning framework for domain generalization in histopathology.

Area of Science:

• Histopathology
• Computational Pathology
• Machine Learning

Background:

• Large foundation models (FMs) trained on extensive histopathology datasets show promise for advancing computational pathology.
• FMs can bridge the domain gap between training and testing datasets, improving translation opportunities.
• However, the computational and data demands of FMs limit their accessibility and widespread adoption.

Purpose of the Study:

• To introduce HistoLite, a lightweight self-supervised learning framework for domain-invariant representation learning in histopathology.
• To enable efficient learning of generalized and transferable features, overcoming the limitations of large-scale FM requirements.
• To evaluate HistoLite's performance in domain generalization using breast Whole Slide Images (WSIs) and compare it with state-of-the-art FMs.

Main Methods:

• Developed HistoLite, a self-supervised learning framework using customizable auto-encoders.
• Curated a novel dataset of WSIs from the same tissue slides scanned by two different platforms to analyze scanner bias.
• Evaluated representation shift using novel metrics (robustness index) and assessed downstream task accuracy.

Main Results:

• Most FMs exhibited susceptibility to scanner bias, indicated by embedding differences and performance drops on out-of-domain data.
• Top-performing FMs (UNI, Virchow2, Prov-GigaPath) were likely favored by large model size and training datasets.
• HistoLite demonstrated low representation shift and minimal performance degradation on out-of-domain data, with modest classification accuracy.

Conclusions:

• Scanner bias poses a significant challenge for real-world deployment of FMs in histopathology.
• HistoLite offers a promising lightweight alternative, balancing generalization and accuracy for computational pathology applications.
• The findings highlight a potential trade-off between model size, generalization capability, and accuracy in histopathology FMs.