Click here to

Session: Radiomics [Return to Session]

A Radiogenomics Study Using a Network-Based Unbalanced Optimal Mass Transport Method in Head and Neck Squamous Cell Carcinoma

J Oh1, H Veeraraghavan1, E Katsoulakis2, A Apte1, J Zhu3*, Y Yu1, N Lee1, V Hatzoglou1, A Tannenbaum3, N Riaz1, J Deasy1, (1) Memorial Sloan Kettering Cancer Center, New York, NY, (2) Department of Veterans Affairs, Tampa, FL, (3) Stony Brook University, Stony Brook, NY


SU-H330-IePD-F9-6 (Sunday, 7/10/2022) 3:30 PM - 4:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 9

Purpose: We investigated whether radiomics features derived from contrast enhanced computed tomography (CT) are correlated with immune cell types in head-and-neck squamous cell cancer (HNSCC) patients.

Methods: DICOM data of pre-treatment CT scans for 77 HNSCC patients were downloaded from The Cancer Imaging Atlas. Primary tumors were manually contoured utilizing CERR software by a single physician and reviewed by two experienced radiologists. Using the CERR radiomics toolbox, 303 radiomic features were extracted from contoured tumors. Due to dental artifacts, first-order statistics and texture features were computed from artifact-free slices in 2D, whereas shape-based features were extracted in 3D. For the 77 samples, 22 immune cell types quantified by CIBERSORT were also analyzed. A network-based computational method derived from a partial correlation network analysis coupled with graphical LASSO, was applied to the CIBERSORT scores and radiomic features, separately. For each resulting network component, an unbalanced extension of optimal mass transport (OMT) was employed to compute the pairwise distance of all samples, because of its ability to handle unnormalized data. Kaplan-Meier survival analysis was then conducted with the log-rank test. Further, correlation analysis between CIBERSORT scores and radiomic features represented by the form of networks was performed via the resulting distance matrices.

Results: The network analysis resulted in a single network, consisting of 6 immune cell types, and 6 network components from radiomic features. Kaplan-Meier analysis on 2 sub-groups, derived from the unbalanced OMT distance in a radiomic network, showed a significant p=0.04 for overall survival. Spearman correlation (r) tests identified a radiomic network highly correlated with the network of immune cell types, yielding r=−0.24 with p=0.036.

Conclusion: Unbalanced OMT applied to radiomic features enabled the identification of radiomic network likely associated with overall survival in HNSCC, and further analysis showed the possible correlation between immune cell types and radiomic features.


Statistical Analysis, Image Analysis, Modeling


IM- Dataset Analysis/Biomathematics: Informatics

Contact Email