Effect of Kernel Functions on the Performance of Support Vector Regression Algorithm in Predicting Patient-Specific Organ Doses from CT Scans

<i>Background:</i> CT examinations are commonly utilized for the diagnosis of internal diseases. The X-rays emitted during CT scans can elevate the risks of developing solid cancers by causing DNA damage. The risk of CT scan-induced solid cancers is intricately linked to the organ doses specific to each patient. The Support Vector Regression (SVR) algorithm exhibits the capability to swiftly and accurately predict organ doses. Kernel functions, including linear, polynomial, and radial basis (RBF) functions, play a crucial role in the overall performance of SVR when predicting patient-specific organ doses from CT scans. Therefore, it is imperative to investigate the influence of kernel selection on the comprehensive predictive effectiveness of SVR. <i>Purpose:</i> This study investigates the impact of kernel functions on the predictive performance of SVR models trained by radiomics features, and to pinpoint the optimal kernel function for predicting patient-specific organ doses from CT scans. <i>Methods:</i> CT images from head and abdominal CT scans were processed using DeepViewer<sup></sup>, an auto-segmentation tool for defining regions of interest (ROIs) within their organs. Radiomics features were extracted from the CT data and ROIs. Benchmark organ doses were calculated through Monte Carlo simulations. SVR models, utilizing the radiomics features, were trained with linear-, polynomial-, and RBF kernels to predict patient-specific organ doses from CT scans. The robustness of the SVR prediction was examined by applying 25 random sample splits with each kernel. The mean absolute percentage error (MAPE) and coefficient of determination (R<sup>2</sup>) were compared among the kernels to identify the optimal kernel. <i>Results:</i> The linear kernel obtains better overall predictive performance than the polynomial and RBF kernels. The SVR trained with the linear kernel function achieves lower MAPE values, below 5% for head organs and under 6.8% for abdominal organs. Furthermore, it shows higher R<sup>2</sup> values exceeding 0.85 for head organs and going beyond 0.8 for abdominal organs. <i>Conclusions:</i> Kernel selection severely impact the overall performance of SVR models. The optimal kernel varies with CT scanned parts and organ types indicating the necessity to conduct organ-specific kernel selection.