Deep interpretable radiogenomic workflow deciphers tumor microenvironment from breast MRI and identifies clinician-interpretable biomarkers.

The tumor microenvironment (TME) influences tumor prognosis and response to immunotherapy. However, current TME assessments primarily rely on invasive pathology slices. Moreover, tumor heterogeneity poses challenges in identifying reliable biomarkers for accurate assessments of TME. We present a general interpretable workflow that deeply correlates magnetic resonance imaging (MRI) and TME. This workflow deconvolutes bulk data to infer a reliable TME profile, enables unsupervised lesion annotation with incorporated TME information, and identifies cancer imaging biomarkers and subtypes using interpretable radiomic features that are readily understandable by clinicians. Interpretable modules of gene and image data improve biomarker discovery and clinical application. The customized deconvolution outperforms existing baselines across multiple datasets, and it initially revealed an inverse relationship between the proportion of cancer-associated fibroblasts (CAFs) and T-cell infiltration in triple-negative breast cancer (TNBC). The radiogenomics model achieved an accuracy of 0.87 in predicting the proportion of CAFs and identified novel, robust microenvironment imaging biomarkers, specifically associated with CAFs. The radiomic features we identified for subtyping exhibited consistent distributions across breast cancer patients and obtained an average accuracy of more than 0.8 in five multicenter validations.