Radiogenomic and Deep Learning Network Approaches to Predict KRAS Mutation and Microsatellite Instability Status From Radiotherapy Plan

Abstract

<h3>Purpose/Objective(s)</h3> We aimed to investigate the role of radiogenomic and deep learning approaches in predicting the KRAS mutation and microsatellite instability (MSI) status of a tumor using the radiotherapy planning computed tomography (CT) images of 110 patients who received neoadjuvant concurrent chemoradiation for locally advanced rectal cancer. <h3>Materials/Methods</h3> After surgical resection, 30 (27.3%) tumors were found to have KRAS mutation and 2 (1.8%) had MSI-high status. We adopted both the radiogenomic and deep learning approaches to predict KRAS mutation and MSI. A total of 378 texture features were extracted from the boost clinical target volume (CTV) in the radiotherapy planning CT images. The least absolute shrinkage and selection operator method was performed to select the features associated with KRAS and MSI status. Tuning parameter (λ) was determined by 10-fold cross validation. The radiogenomic score was calculated by combining the coefficients of the selected features. Meanwhile, we developed a deep learning network based on the three-dimensional input of the CTV. Performance of models was internally validated. <h3>Results</h3> The predictive ability of the radiogenomic score model revealed AUCs of 0.73 for KRAS mutation and 0.99 for MSI status, whereas that for the deep learning model demonstrated worse performance, with AUCs of 0.63 for KRAS and 0.58 for MSI. <h3>Conclusion</h3> Compared with the deep learning network model, the radiogenomic score model was a more feasible approach to predict KRAS and MSI status. These results should be externally validated on a larger dataset