Predicting Progression-Free Survival Using Primary and Nodal Radiomic Features in Head and Neck Cancer

Abstract

Imaging quantified by radiomics has yielded reproducible predictions of both oncologic outcomes and treatment-related toxicity. Much of the correlation with oncologic outcomes has focused on image features from the primary tumor. However, both primary tumor and nodal features can impact outcomes, and it is known that differences in gene expression in primary tumors and nodal metastases provide distinct information regarding survival in head and neck squamous cell carcinomas (HNSCC). Nevertheless, current radiomic biomarkers often make no distinction between primary and nodal radiomic features. Therefore, we hypothesize that combining CT image features of primary tumors and nodal metastases in HNSCC will improve our ability to predict progression-free survival (PFS) compared to primary features alone.Pre-treatment CT images were prospectively collected from 115 HNSCC patients undergoing RT at our institution from 2008-2018. PFS was determined at time of last follow-up prior to 2018. Clinical data were systemically captured from patient medical charts. 107 radiomic features were extracted for the primary tumor and the ipsilateral/contralateral nodal regions. Features were pre-selected by literature review for those possibly significant in predicting loco-regional control of HNSCC. Tumor volume, correlation of gray level co-occurrence matrix (Corr-GLCM), run length non-uniformity of gray level of gray level run length matrix (RLN-GLRLM), and coarseness of neighboring gray tone difference matrix (Coarse-NGTDM) for the primary contours; and tumor volume and Corr-GLCM of the nodal contours were included in the model. A regression analysis of pre-selected features was performed using ridge regularization in the training set (n = 93 primary; n = 48 primary & nodal), and a generalized linear model was built with repeated ten-fold cross validation of patients treated from 2008-2017. Model performance was evaluated in a test set (patients treated 2017-2018) using the area under the receiver operating characteristic curve (AUC). Models were built separately for patients with primary tumor only and primary + nodal radiomic features.The AUC/sensitivity/specificity on the test set for the model with primary features (n = 22) was 0.91/1.00/0.75, and for the model with primary and nodal features (n = 18) was 0.75/0.60/0.92. The features with the highest coefficients in the combined model included primary Coarse-NGTDM (β = -0.21), primary Corr-GLCM (β = 0.0082), and nodal Corr-GLCM (β = -0.00041).In a small cohort of HNSCC patients, combining nodal features with primary features did not improve the prediction performance compared to primary features alone. This work suggests that primary tumor radiomic features may influence PFS more than nodal features. Future work is required in a larger cohort to validate these findings.