Artificial intelligence is increasingly capable of extracting quantitative information from routine medical images that may not be apparent to human observers. For radiation oncology, an important question is whether these imaging signatures can move beyond prediction alone to provide robust, interpretable and biologically meaningful biomarkers that ultimately inform patient care.
The study represents an interesting step in this direction. The investigators developed an interaction-informed deep radiomics framework using pretreatment CT imaging to predict distant metastasis-free survival (DMFS) in patients with head and neck cancer. Importantly, the study included 3,421 patients from four independent cohorts, allowing the investigators to evaluate model performance across institutions rather than relying on a single-institution dataset.
From a medical physics and quantitative imaging perspective, one particularly interesting aspect is how the investigators represent radiomic information. Conventional radiomics typically extracts a large number of predefined imaging features and analyzes them as individual variables. However, tumor heterogeneity is complex, and the relationships among imaging features may carry information that individual features cannot capture. Using the OmicsMap framework, the investigators transformed high-dimensional radiomic features into two-dimensional topological maps, enabling the deep-learning model to explicitly incorporate feature-feature interactions.
The model demonstrated consistent prognostic performance across independent cohorts, with C-indices ranging from 0.671 to 0.768. When key clinical factors were incorporated, performance improved further, reaching a C-index of 0.864 in the HN1 cohort and 0.730 in the HN-PET-CT cohort. These results are encouraging because external validation remains one of the major challenges for AI and radiomics models in medical imaging.
Equally important is the investigators’ effort to understand what drives the predictions. SHAP analysis identified wavelet-transformed texture heterogeneity features, including those related to gray-level nonuniformity, local homogeneity and entropy, as important contributors. This suggests that quantitative patterns of intratumoral heterogeneity on routine CT images may contain clinically relevant prognostic information.
Perhaps the most intriguing finding is the connection between these imaging phenotypes and underlying tumor biology. In patients with matched CT and RNA sequencing data, imaging-defined high-risk tumors were associated with fibrosis-prone, immune-excluded, proliferative and hypoxic programs, whereas lower-risk tumors demonstrated more immune-active characteristics. Establishing such radiogenomic relationships may help us move from asking simply whether an AI model works to understanding what biological processes the model may be detecting.
There are still important steps before this approach can influence clinical care. Prospective validation will be necessary, particularly to evaluate robustness across scanners, acquisition protocols and diverse patient populations. In addition, improved prognostic discrimination does not necessarily mean that treatment decisions based on these predictions will improve patient outcomes.
Nevertheless, this work illustrates a promising direction for AI in radiation oncology. The future of imaging biomarkers may depend not simply on building more powerful predictive models, but on developing models that are generalizable, interpretable and connected to tumor biology. If these findings can be prospectively validated, routine pretreatment CT could potentially provide information beyond anatomy and treatment planning, helping identify patients who may benefit from closer surveillance, treatment intensification or enrollment in risk-adapted clinical trials.
Abstract 164, Feature Interaction-Informed Deep Radiomics Unveils Tumor Heterogeneity and Predicts Distant Metastasis-Free Survival in Head and Neck Cancer: A Multicenter Study, was presented during SS 13: AI Applications in Outcome Prediction of the 68th ASTRO Annual Meeting.
Published on: September 29, 2026
