A deep radiomics model called OmicsMap was developed to capture relationships among radiomic features, characterize tumor heterogeneity, and predict distant metastasis using CT images in patients with head and neck cancer. Findings from a large, multicenter study were presented at the American Society for Radiation Oncology (ASTRO) Annual Meeting (Abstract 164).  

“OmicsMap provides a structured way to encode integrated feature relationships and improves prognostic modeling beyond conventional radiomics in head and neck cancer,” said presenting author Andrew Heider, BS, Assistant Clinical Research Coordinator, Radiation Oncology – Radiation Physics at Stanford University. “It improves CT-based prediction of distant metastasis while also capturing biologically meaningful tumor heterogeneity.” 

Model Methods 

Radiomics provides a noninvasive way to obtain valuable information about tumor phenotype, heterogeneity, and treatment response from medical images. A typical radiomics pipeline includes image acquisition, tumor delineation, feature extraction and selection, and machine learning modeling. However, radiomics data are often high-dimensional and noisy, making feature classification and pattern recognition challenging. Traditionally, radiomics features have been flattened into one-dimensional vectors for analysis in lower-dimensional spaces, potentially obscuring relationships among features and other complex information. The researchers hypothesized that preserving these relationships could provide additional prognostic information about tumor heterogeneity and distant metastasis.  

Deep radiomics uses deep learning techniques to extract high-dimensional features from radiological data. 

OmicsMap is an interaction-informed deep radiomics framework that transforms tabular radiomic features into a spatially structured, image-like representation in which the spatial organization encodes relationships among features. The OmicsMap workflow includes image acquisition, region-of-interest delineation, and feature extraction and selection, followed by the construction of patient-specific omics maps that represent spatial arrangements and feature dependencies.  

The OmicsMap model consists of an interaction matrix aligned with a projection matrix to generate patient-specific omics maps, along with a convolutional neural network to predict survival across multiple discrete follow-up intervals.  

“The idea is to place related features close together on the grid, and this spatial layout preserves interaction structure and enables convolutional neural networks to better capture complex patterns,” Mr. Heider explained.  

Study Methods 

The study included 3,421 patients with head and neck cancer from four independent cohorts across 12 institutions in the United States, Canada, and the Netherlands. All patients had histologically confirmed head and neck cancer, available pretreatment CT images with primary gross tumor volume segmentation, and complete prognostic follow-up data.  

The largest cohort was drawn from the RADCURE dataset, which was used for model development and divided into training (n = 1,883), validation (n = 580), and testing (n = 434) sets. The remaining patients were reserved for external validation. A cohort of 94 patients from The Cancer Genome Atlas was also used for radiogenomic analysis and biological interpretation. 

The investigators compared OmicsMap with several benchmark models, including nonlinear radiomics models, CT-based three-dimensional (3D) convolutional neural network models, and a shuffled OmicsMap control model. The control model used the same convolutional neural network architecture but with a disrupted spatial arrangement of features. The researchers also analyzed SHapley Additive exPlanations (SHAP) values to assess the contributions of individual features to the model’s predictions.  

Study Findings 

The OmicsMap model demonstrated consistent prognostic performance across cohorts. In the development cohort, the model incorporating clinical variables achieved an area under the curve (AUC) of 0.727 at 1 year, 0.741 at 2 years, and 0.731 at 3 years. In an independent validation cohort, the corresponding AUCs were 0.865, 0.895, and 0.869, respectively.  

Calibration showed good agreement between the predicted and observed 2-year distant metastasis–free survival rates.  

Across cohorts, the model consistently stratified patients into high- and low-risk groups according to their predicted distant metastasis–free survival. The risk stratification was also associated with progression-free survival in The Cancer Genome Atlas cohort (P = .0088).  

Compared with benchmark models, OmicsMap improved the concordance index by 3.2% to 14.3% over nonlinear radiomics models, by 2.1% to 12.5% over a CT-based 3D convolutional neural network model, and by 2.1% to 5.5% over the shuffled OmicsMap control model. The main advantage of OmicsMap over the other models came from “the structured organization of inter-feature relationships rather than from model complexity alone,” Mr. Heider explained.  

SHAP analysis identified high-order wavelet-transformed texture heterogeneity features as the most influential predictors in the OmicsMap model.  

OmicsMap visualizations revealed distinct spatial patterns between high- and low-risk patients. High-risk tumors showed greater enrichment of pathways associated with cell proliferation, DNA replication, extracellular matrix remodeling, hypoxia, angiogenesis, and epithelial mesenchymal transition. These findings, including increased matrix remodeling activity, were consistent with a more aggressive tumor phenotype and a more fibrosis-associated tumor microenvironment. In contrast, low-risk tumors exhibited greater immune-related activity.  

The researchers also developed a three-gene prognostic signature that was validated in The Cancer Genome Atlas cohort and associated with a shorter progression-free interval.  

They plan to incorporate additional deep imaging features into the model and prospectively validate the framework in more diverse patient cohorts.  

DISCLOSURES: The study authors reported no conflicts of interest. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.