Attention-based ROI Discovery in 3D Tissue Images

TL;DR

This paper presents a novel self-supervised graph attention network method for automatically discovering biologically relevant regions in complex 3D tissue imaging data, enhancing spatial interaction analysis without manual labeling.

Contribution

Introduces SSGAT, a self-supervised multi-layer graph attention network, for automatic ROI discovery in 3D tissue images, coupled with an interactive visualization interface.

Findings

Effectively identifies meaningful immune microenvironments

Reveals complex spatial bioreactions visually

Operates without manual region labels

Abstract

High-dimensional tissue imaging generates highly complex 3D data containing multiple biomarkers, making it challenging to identify biologically relevant regions without an expert user specifying manual labels for regions of interest. We introduce an approach to automatically identifying regions of interest (ROIs) in the 3D microscopy data. Our approach is based on a novel self-supervised multi-layer graph attention network (SSGAT), coupled with a React interactive interface wrapped around Vitessce. SSGAT employs an adversarial self-supervised learning objective to identify meaningful immune microenvironments through marker interactions. Our method reveals complex spatial bioreactions that can be visually assessed to assess their distribution across tissue. Index Terms: Biomedical visualization, graph attention networks,self-supervised learning, spatial interaction analysis.