Revealing interpretable signatures of whole slide images

Context

Detecting different tissue (e.g. tumor, muscle, fibrosis) and cell types (e.g. B cells, T cells, plasma cells) from digitized whole slide images (WSI) would greatly facilitate patient diagnosis, patient inclusion to new clinical trials, and support prediction of response to treatment.

‍

Methods

We use pathologist’s annotation of tissues to train AI models called Histomics.

• Pixel-level labels
• Tile-level labels

‍

Results

Histomics serve as interpretable features to:

• Feed machine learning models
• Decypher AI-based biomarkers
• Characterize patients
• Identify patterns in subgroups of patients

‍

Impact

To identify novel biomarkers in images of specific tumoral regions that are important to better understand disease evolution and differentiated outcomes.

‍