Abstract LB175: Weakly supervised deep learning enables spatially resolved cell type inference from H&E histopathology

Key Points

• This research aims to develop a framework for inferring cell types from traditional histopathology images without requiring complex annotations.
• Developed the SLIDE-EX deep learning framework using routine H&E whole-slide images and slide-level labels from RNA sequencing.
• Tested the framework on breast cancer samples with spatial transcriptomics data and expert annotations.
• Compared the performance of SLIDE-EX against HoVer-Net, a model designed for nucleus-level detection.
• SLIDE-EX achieved area under the curve values of 0.82 for cancer cells, 0.85 for stromal cells, and 0.79 for lymphocytes.
• SLIDE-EX outperformed HoVer-Net across all cell types, indicating its superior predictive capability.
• The model predicted cell type-specific gene expression for thousands of genes, enriching for cellular functions and improving chemotherapy response prediction.

Abstract