A Deep Learning Approach to Infer Cellular Features from Pathology Imaging Data

Abstract

Recent developments in single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) provide unprecedented opportunities for studying individual cells and their organization. While these techniques are underutilized in clinical practice due to cost and logistical challenges, we propose an innovative alternative. We suggest leveraging scRNA-seq and ST data to train a deep learning model for histologically stained images, particularly hematoxylin and eosin (H\&E) stained whole slide images, commonly used in disease assessment. In the burgeoning field of digital pathology, where deep learning excels in extracting meaningful imaging features, limited annotations for small image segments pose a challenge. To address this, we introduce STApath, a transfer-learning neural network model that automates patch-level annotation by exploiting paired ST data for model training and is therefore capable of predicting cell type proportions and classifying tumor microenvironments. Despite challenges such as uncertain annotations from spatial transcriptomics and image resolution disparities, our work establishes the viability of this pipeline. STApath's evaluation on breast cancer datasets demonstrates promising results even with limited training data and varying proportions and resolutions. Anticipating an influx of ST data, ongoing STApath updates hold the potential to become an invaluable AI tool for pathologists, streamlining diagnostic tasks.