3684 - AI-Driven Graph-Based Radiomics: Enhancing Imaging Biomarkers Reproducibility in Multi-Institutional Head and Neck Cancer Management
Presenter(s)
H. Moradmand1, J. Molitoris1, L. Schumaker1, E. Allor1, R. Krc1, P. T. Tran1, A. Sawant2, R. Mehra3, D. Gaykalova4, and L. Ren1; 1University of Maryland School of Medicine, Baltimore, MD, 2University of Maryland, School of Medicine, Baltimore, MD, 3University of Maryland Cancer Center, Baltimore, MD, 4University of Maryland School of Medicine, BALTIMORE, MD
Purpose/Objective(s): Radiomics leverages high-throughput extraction of quantitative imaging features from medical images to develop biomarkers that enhance diagnosis, prognosis, and treatment decisions. However, in head and neck squamous cell carcinoma (HNSCC), the clinical utility of radiomic biomarkers is limited by variability in imaging parameters and across institutions. Our study aimed to develop an AI-driven Graph-Based Feature Selection (Graph-FS) approach that enhances the stability and reproducibility of radiomic features, ultimately improving the reliability of imaging biomarkers for better clinical decision making.
Materials/Methods: We conducted a retrospective analysis of 752 HNSCC cases across three institutions. From each patient's gross tumor volume, 1,648 radiomic features were extracted under 36 different parameter settings (including variations in normalization scales, bin widths, and outlier thresholds). The Graph-FS approach utilizes unsupervised learning to generate feature graphs that identify clusters of correlated features and select the most representative ones. Its performance was benchmarked against Lasso, Boruta, Recursive Feature Elimination (RFE), and Minimum Redundancy Maximum Relevance (mRMR) using stability metrics (Jaccard Index [JI], Dice Sorensen Index [DSC], and Overlap Percentage [OP]) and reproducibility measures (JI). Additionally, we assessed the prognostic performance of each feature selection method for 2-year survival predictions to evaluate their clinical impact.
Results: Graph-FS improved stability within individual centers by 98.26%, achieving the highest DSC (0.62) compared to Boruta (0.01), Lasso (0.03), mRMR (0.04), and RFE (0.03). It enhanced reproducibility across multi-institutional by 98.10%, with a JI of 0.127 versus Boruta (0.005), Lasso (0.010), RFE (0.006), and mRMR (0.014). On the external validation test set, Graph-FS achieved the highest AUC (0.71) for 2-year survival prediction, outperforming Boruta (0.64), Lasso (0.61), mRMR (0.58), and RFE (0.59).
Conclusion: The AI-driven Graph-FS method significantly improves the stability and reproducibility of radiomic feature selection in a multi-institution HNSCC study. This approach supports the development of robust imaging biomarkers and enhances clinical decision-making in precision oncology.