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

Oregon Convention Center 

Digital pathology is a fast-growing field where leveraging machine learning methods uncovers meaningful imaging features, but it faces the hurdles of sparse annotations, particularly for small pathology image segments. We propose a novel approach that employs spatially resolved transcriptomic data for annotations, though it faces challenges like annotation uncertainty from transcriptomic data and inconsistent image resolutions. We established the viability of this approach and developed a novel transfer-learning neural network model, STpath (Spatial Transcriptomics and pathology images), designed to predict cell type proportions or classify tumor microenvironments. Our findings revealed the intrinsic link between pre-trained deep learning features and cell identities in pathology image segments. Tested on two breast cancer datasets, STpath demonstrated robust performance, effectively handling samples with diverse cell type proportions and high-resolution images despite limited training data. As the influx of spatially resolved transcriptomic data continues, we foresee ongoing updates to STpath, shaping it into an invaluable AI tool for pathologists, enhancing diagnostic accuracy.

Digital and computational pathology

Whole slide imaging

Spatially resolved transcriptomics

Artificial Intelligence

Transfer learning

Cell type proportion and tumor microenvironment 

Section on Statistics in Imaging