Spatially-aware H&E Deep Learning Features of Single Cells Show Cancer & Cell Type Classification Abilities

Document Type

Article

Publication Date

2026

JAX Location

JGM

Abstract

H&E-stained whole-slide images (WSIs) are continuously used for top-tier histological diagnoses. Artificial intelligence and machine learning (AI/ML) models have been shown to improve the speed and accuracy of these histological analyses. This study aims to build and evaluate the performance of a cell mosaic workflow within the spatial transcriptomics quantification (STQ) pipeline to extract cell-centered, spatially dependent and independent deep learning features for downstream analysis in three categories: cancer/non-cancer, cell type (cancer-associated fibroblasts), and cancer cell subtype classifications for breast cancer and cervical cancer samples. The cell mosaic workflow generally displayed better performance compared to Nuclei.io, a traditional histological image pre-processing tool, in terms of accuracy, f1, and ROC-AUC scores. The greatest number of inaccuracies occurred near tumor boundaries and in regions of heterogeneity, though this effect is less intense in the cell mosaic workflow.

Please contact the Joan Staats Library for information regarding this document.

COinS