Exploit Spatially Resolved Transcriptomic Data to Infer Cellular Features from Pathology Imaging Data.

Area of Science:

• Computational pathology
• Bioinformatics
• Artificial intelligence in medicine

Background:

• Deep learning in digital pathology requires extensive annotated image data, which is often scarce.
• Limited annotations for small image patches hinder model training and efficacy.

Purpose of the Study:

• To develop an innovative approach for annotating digital pathology images using paired spatial transcriptomic data.
• To introduce STpath, a novel transfer-learning neural network for predicting cell type proportions and classifying tumor microenvironments.

Main Methods:

• Leveraging paired spatially resolved transcriptomic data to generate annotations for pathology images.
• Developing and evaluating STpath, a transfer-learning neural network model.
• Utilizing three distinct breast cancer datasets for performance evaluation.

Main Results:

• Demonstrated the feasibility of using spatial transcriptomics for pathology image annotation.
• Found associations between pre-trained deep learning model features and cell type identities in pathology patches.
• STpath showed promising performance on limited training data, excelling with variable cell type proportions and high-resolution images.

Conclusions:

• STpath is a valuable AI tool for assisting pathologists in diagnostic tasks.
• The approach is scalable with the increasing availability of spatial transcriptomic data.
• Future updates to STpath are anticipated to enhance its utility in pathology.