MRI-based radiomic signature and its association with genomic complexity in breast tumor heterogeneity.

Abstract

563 Background: Breast cancer is inherently heterogeneous, posing challenges for effective treatment. Uncovering the relationship between imaging features and genomic profiles could improve patient stratification. In this study, we evaluated whether radiomic features can capture the underlying genomic complexity of breast tumors, potentially offering a non-invasive means to better characterize tumor heterogeneity. Methods: We analyzed 284 breast cancer patients using an integrated radiogenomic approach. MRI-derived radiomics features were extracted and clustered using unsupervised learning methods, resulting in 12 distinct clusters. We then analyzed these clusters against matched whole-genome sequencing and transcriptome data, focusing on heterogeneity-related radiomics features. Clustering was performed using dynamic tree cutting after hierarchical clustering of 10 principal components derived from 214 radiomic features. Results: We identified distinct patterns of tumor heterogeneity among the 12 identified clusters, named according to descending cluster size (range: 10-46). Clusters 9, 4, and 3 exhibited the highest homogeneity (in that order), with cluster 9 being the most homogeneous overall. Cluster 12, 11, 8, 7, and 5 showed varying degrees of heterogeneity, while clusters 1 and 2 were moderately heterogeneous. Cluster 1-3 were HER2-enriched (PAM50). Clusters 1 and 2 together had ERBB2 amplifications (33%; Fisher’s exact test, P = 0.056), whereas cluster 3 was HER2-positive (IHC) without amplifications. Cluster 1 leaned toward the basal-like subtype, while 3 leaned toward luminal A. Cluster 2 was enriched in luminal B (50%; P = 0.012). Cluster 1-3 were distinguishable by their degree of radiomics-quantified heterogeneity. Cluster 4 was enriched in high Myc expression (17%; P = 0.059). Cluster 5 was enriched in whole-genome-based HRD (40%; P = 0.01) and basal-like (52%; P = 0.001). Cluster 6 was deprived of TP53 mutations (37%; P = 0.04), had low tumor mutational burden, and was characterized by small volume but higher surface-volume ratio, suggesting irregular shape. Cluster 8 was enriched in PIK3CA mutations (60%, P = 0.046), cluster 10 was enriched in CHEK2 mutations (9%; P = 0.039), cluster 11 showed high TERT (40%, P = 0.005) and CDKN2A (40%; P = 0.048) expression, cluster 12 was predominantly post-menopausal (80%; P = 0.47), and both clusters 10 and 12 exhibited low ESR1 expression (20%; P = 0.035). Conclusions: This comprehensive radiogenomic analysis demonstrates that MRI-based radiomics features can effectively capture tumor heterogeneity patterns that correlate with specific genomic alterations in breast cancer. The identification of 12 distinct clusters, each with characteristic genomic features, provides new insights into the biological basis of tumor heterogeneity, potentially opening new avenues for breast cancer subtyping.