Compositional and interpretable representation of histology using AI foundation models and sparse autoencoders.

Area of Science:

• Histopathology
• Computational Pathology
• Biomedical Imaging

Background:

• Hematoxylin and eosin (H&E) staining is a cornerstone of histopathology for over 150 years.
• High-plex spatial profiling offers single-cell resolution for protein and RNA expression but does not replace H&E imaging.
• Interpreting deep learning-based computational pathology (CPath) models in biological terms remains a significant challenge, limiting their use in spatial profiling.

Purpose of the Study:

• To develop a human-in-the-loop computational framework to identify human-interpretable histopathology features from H&E images.
• To leverage CPath foundation models (FMs) and sparse autoencoders (SAEs) for feature decomposition and identification.
• To integrate morphological features with molecular spatial profiling data.

Main Methods:

• A human-in-the-loop computational framework combining CPath foundation models (FMs) and sparse autoencoders (SAEs).
• Decomposition of FM embeddings to automatically identify diverse histopathology features.
• Application of the FM-SAE model to H&E images of pulmonary diseases (tuberculosis, lung cancer).

Main Results:

• The FM-SAE model successfully identified diverse, human-interpretable histopathology features in H&E images.
• Human-machine interaction augmented and accelerated expert interpretation of pulmonary disease cases.
• Generated annotations enabled morphology-aware integration of tissue architecture with molecular spatial profiling.

Conclusions:

• The developed framework enhances the interpretability of CPath models for H&E imaging.
• This approach accelerates expert analysis and improves the integration of morphological and molecular data in spatial profiling studies.
• The methodology holds promise for advancing the diagnosis and research of diseases like tuberculosis and lung cancer.