Automatic Dose Prescription for Radiotherapy of Brain Metastases Using Two-Path Three-Dimensional CNN

Abstract

<h3>Purpose/Objective(s)</h3> AI modeling physicians' clinical decision making (CDM) can improve the efficiency and accuracy of the clinical practice and provide a valuable tool for initial consultations to patients seeking secondary opinions. We propose to develop an AI model that mimics physician decision process based on physician-specific or institution-specific clinical practice. We will use dose prescription as an example to demonstrate the feasibility of such models in this study. <h3>Materials/Methods</h3> This study included 148 patients with brain metastases treated by hypo-fractionated radiotherapy from 2017 to 2021. CT images and contours of the target volume and organ at risk (OAR) were extracted from the treatment planning system. A 3D convolutional neural network (CNN) architecture was built using two encoding paths with the same kernel and filters to capture the different image and contour features. Specifically, one path was built to capture the tumor feature, including the size and location of the tumor, and another path was built to capture the relative spatial relationship between the tumor and OARs. The model combines information from the two paths to make a final prediction of dose prescription. The actual prescription in the patient record was used as the ground truth for modeling training. The model performance was assessed by 8-fold cross-validation, in which each fold consisted of randomly selected 100 training, 20 validation, and 20 testing subjects. <h3>Results</h3> Of the 148 patients, 90 were treated by a single fraction (1 × 21, 1 × 22, 1 × 24 Gy) and 58 were treated by multiple fractions (3 × 9 Gy, 5 × 6 Gy) by 10 physicians based on the record. 138 (93%) cases were predicted correctly, and 10 (7%) cases were predicted incorrectly by the model. For the 10 failed cases, 8 were caused by the practice variations among different physicians, which was not accounted for by the model trained using data from a group of physicians. <h3>Conclusion</h3> This study demonstrated the feasibility of an AI model to predict dose prescription in CDM modeling. In the future, clinical parameters, such as patient performance score, tumor type, re-treatment, adjuvant therapy, will be included in the model to improve its performance further. The physician-specific model will also be explored to model individual physician's practice. Such physician-specific or institution-specific CDM models can serve as vital tools to address healthcare disparities by providing initial consultations to patients in underdeveloped areas or developing countries. It can also become a valuable QA tool for physicians to cross-check intra- and inter-institution practices.