Encoding Inter-Radiomic Feature Relationships As Image-Based Representations for Improved Distant Metastasis Prediction and Heterogeneity Phenotyping In Head and Neck Cancer
Abstract
Purpose
Conventional radiomics modeling assumes handcrafted features to be independent variables, overlooking structured inter-feature relationships that encode tumor heterogeneity and limit robust generalization across institutions. We propose a novel interaction-aware representation learning framework that explicitly encodes inter-radiomic feature relationships to improve distant metastasis-free survival (DMFS) prediction and biologically meaningful heterogeneity phenotyping in head and neck (HN) cancer.
Methods
A multi-institutional study was conducted on 3,421 HN cancer patients across twelve centers (including RADCURE, HN1, HN-PET-CT, and TCGA-HNSC). To overcome the limitations of isolated feature analysis, we developed OmicsMap to explicitly encode statistical dependencies between radiomic features into a spatially structured, 2D data-driven topology. A convolutional neural network was trained on these maps to extract interaction-aware prognostic signatures, which were fused with clinical variables. Model performance was assessed using the concordance index (C-index) and time-dependent area under the receiver operating characteristic curve (AUC) across RADCURE, HN1, and HN-PET-CT cohorts. Furthermore, we performed radiogenomic analysis using RNA-seq data from TCGA-HNSC cohort to decode the biological underpinnings of the imaging-defined risk phenotypes.
Results
The OmicsMap model demonstrated robust generalizability across independent cohorts, achieving C-indices of 0.742 (RADCURE), 0.768 (HN1), and 0.671 (HN-PET-CT), representing a 5.8-6.4% improvement over conventional radiomics models. The OmicsMap-clinical fusion model further improved stratification, yielding C-indices up to 0.864 and time-dependent AUCs of 0.727-0.895. The fusion model effectively separated high- vs low-risk patients for DMFS and overall survival across cohorts (P<0.01). Radiogenomic analysis revealed distinct biological phenotypes, with low-risk tumors enriched in immune activation pathways, while high-risk tumors exhibited fibrosis-associated microenvironmental signatures characterized by proliferation, hypoxia, and epithelial-mesenchymal transition.
Conclusion
Encoding inter-radiomic feature relationships with OmicsMap improves CT-based distant metastasis prediction and enables biologically meaningful phenotyping of tumor heterogeneity in HN cancer.