Deep Learning-Based Prediction of VMAT Dose Distribution in Nasopharyngeal Cancer Using a Modified ResNet Architecture

Purpose: Accurate prediction of dose distributions in radiotherapy treatment planning can improve efficiency and support automated plan quality assessment. This study develops a deep learning framework based on a modified ResNet architecture to predict voxel-wise dose distributions for nasopharyngeal cancer patients treated with volumetric modulated arc therapy (VMAT). Methods: A retrospective dataset of 230 patients was analyzed, with nine organs at risk (OARs) evaluated. Model performance was assessed using dose–volume histogram (DVH) metrics, including mean absolute error (MAE) and mean squared error (MSE). Results: The model demonstrated good agreement with clinical plans based on DVH metrics. Lower prediction errors were observed in simpler structures such as the eyes (MAE up to 0.02 Gy) and brainstem (MAE = 0.023 Gy), while higher errors were found in more complex and highly modulated regions such as the parotid glands (MAE up to 0.047 Gy). Prediction accuracy, based on MAE, ranked from best to worst as Right Eye, Left Eye, Brainstem, Spinal Cord, Left Optic Nerve, Right Parotid, Right Optic Nerve, Optic Chiasm, and Left Parotid. A similar trend was observed for MSE. Conclusion: These results indicate that the model effectively captures general dose distribution patterns, but remains challenged in regions with steep dose gradients. This approach shows potential to support automated treatment planning workflows and plan quality evaluation in head-and-neck radiotherapy.