KEY POINTS
- This multicenter retrospective study analyzed 3,421 patients after eligibility filtering across four cohorts spanning multiple institutions. RADCURE contributed 2,897 patients, HN1 137, HN-PET-CT 293, and TCGA-HNSC 94 patients with paired CT and RNA sequencing for biological interpretation.
- Investigators extracted 833 radiomic features from the primary-tumor GTV and selected 100 for modeling. Instead of treating them as an ordinary vector, OmicsMap arranged correlated features close together on a two-dimensional grid so a convolutional neural network could learn inter-feature relationships as spatial patterns.
- OmicsMap consistently outperformed conventional radiomics for distant metastasis-free survival. C-indices were 0.742 vs 0.704 in RADCURE, 0.768 vs 0.722 in HN1, and 0.671 vs 0.634 in HN-PET-CT. External 1–3-year AUCs reached 0.905/0.778/0.727 in HN1 and 0.783/0.710/0.697 in HN-PET-CT.
- Adding clinical variables—sex, T stage, N stage and concurrent chemoradiotherapy—improved performance further in the external cohorts. The fusion model reached a C-index of 0.864 (95% CI 0.798–0.930) in HN1 and 0.730 (95% CI 0.656–0.803) in HN-PET-CT.
- The OmicsMap score retained independent prognostic information after adjustment in RADCURE (HR 1.71, p=0.003) and HN-PET-CT (HR 1.45, p=0.02). High- and low-risk groups also showed significantly different distant-metastasis-free survival in RADCURE (p<0.0001), HN1 (p=0.00013) and HN-PET-CT (p=0.0084).
- Cross-platform robustness was specifically examined. In the multi-institutional HN-PET-CT cohort, ComBat harmonization improved C-index from 0.617 to 0.671, while performance within RADCURE remained relatively stable across scanner vendors (0.714–0.752) and contrast protocols (0.726–0.748).
- The imaging-defined risk phenotype also had a biologically coherent transcriptomic signal. High-risk tumors were enriched for cell-cycle proliferation, hypoxia, angiogenesis, epithelial–mesenchymal transition and extracellular-matrix remodeling, whereas low-risk tumors showed stronger immune-related signatures. A derived three-gene signature—MYO16, NTSR1 and LINC01429—also stratified progression in an independent TCGA RNA cohort of 519 patients (p=0.0029).
CLINICAL TAKEAWAY
The interesting idea here is that radiomic features may contain useful information not only individually but also in how they relate to each other, and the gains persisted across two external datasets. This is stronger evidence than a typical single-center radiomics paper, but it remains retrospective prediction research: there is no evidence yet that using OmicsMap to intensify systemic or radiation treatment improves outcomes.
SOURCE
International Journal of Radiation Oncology, Biology, Physics