Backpropagation Neural Network Implementation in Volumetric Modulated Arch Therapy of Brain Cancer Dose Prediction

Introduction: The quality of volumetric modulated arc therapy (VMAT) planning is highly subjective and varies due to differences in planner’s experience. This process is time-consuming and involves multiple iterations to achieve clinical goals. Recent advancements in artificial intelligence (AI) offers an objective approach to improve the efficiency of VMAT planning. Material and Methods: In this study, the backpropagation neural network with 5-fold cross-validation model was employed to train the extracted Radiomics and dosiomics features from organ contours DICOM RT structure and dose distribution DICOM RT dose using 178 VMAT technique brain cancer patients. The Radiomics and dosiomics features represent the organ shapes and dose distribution quantitatively to increase the prediction accuracy. The Mean Squared Error and paired t-test was used in model evaluation. The treatment planning quality parameters, homogeneity index (HI) and conformity index (CI), was evaluated from both predicted and clinical dose. Results: The paired t-test indicated no significant differences (p-value > 0.05) in organs at risk (OAR) and planning target volume (PTV). The p-value for the left optic nerve is the lowest among average dose (Dmean) and maximum dose (Dmax), respectively 0.1456 and 0.0662. The average HI was 0.084±0.036 (predicted) and 0.089±0.073 (clinical), and CI was 0.938±0.107 (predicted) and 0.957±0.136 (clinical). Conclusion: The p-value for predicted parameters suggest that neural network-based dose prediction using Radiomics and dosiomics features produces results comparable to the manual treatment planning by medical physicists (overall testing dataset MSE = 0.0355).