Breast cancer radiogenomics analysis via computational perturbation on AI-driven multi-omics guided image synthesis

Radiogenomics offers a promising non-invasive approach for characterizing breast cancer (BC), yet its progress is often limited by the scarcity of cohorts containing matched imaging and multi-omics data. Recent advances in generative AI have enabled the synthesis of imaging phenotypes from genomic features, but prior work has focused on the combined influence of all genomic signals rather than isolating the effects of specific biological pathways. In this study, we introduce a perturbation-based radiogenomic framework that integrates multi-omics meta-genes with a conditional generative adversarial network (cGAN) to examine how pathway-level alterations influence synthetic BC MRI phenotypes. Seventeen meta-genes derived from Bayesian Tensor Factorization were perturbed at three levels (overexpression, base case, knockout) and synthetic DCE-MRI volumes were generated for each condition. Radiomic features were extracted using MedSAM-guided segmentation and PyRadiomics, followed by statistical evaluation using one-way ANOVA and Tukey post-hoc testing. Among the 17 pathways analyzed, only two meta-genes, representing cell cycle regulation and steroid hormone biosynthesis, produced significant and biologically interpretable changes in tumor size, heterogeneity, and textural patterns. These findings show that computational perturbation can uncover pathway-specific imaging signatures and offer mechanistic insights that complement traditional radiogenomics and explainable AI approaches. This work demonstrates the potential of perturbation-driven generative models to advance precision imaging genomics in BC.