Three-dimensional Hierarchically Dense U-net Architecture for Knowledge-based Dose Prediction in Head and Neck Radiotherapy.

Purpose: Our objective is to investigate a novel knowledge-based planning model that utilizes the newly developed dense U-Net architecture to predict three-dimensional (3D) dose distributions in patients with head and neck cancer. Materials and Methods: We utilized data published by the American Association of Physicists in Medicine Institute, which includes treatment plans for 340 patients with head and neck cancer who were treated using intensity modulated radiotherapy treatment with a prescribed dose of 70 Gy. The data were divided into three subsets, and the proposed dense U-Net was applied to predict full volumetric dose distributions. Model performance was evaluated using the mean absolute error (MAE) between predicted and clinical doses across all subsets. Results: The MAE between the reference and predicted dose distributions for the training, validation, and testing datasets was 1.60 Gy, 3.09 Gy, and 3.14 Gy, respectively. The mean absolute dose error was measured at 1.53 Gy in the brainstem, 3.43 Gy in the left parotid, 3.29 Gy in the right parotid, 2.05 Gy in the spinal cord, 2.40 Gy in the esophagus, 3.81 Gy in the mandible, 3.35 Gy in the larynx, 1.56 Gy in planning target volume (PTV70), 1.82 Gy in PTV63, 1.91 Gy in PTV56, and 3.14 Gy in the body contour for the testing data. The proposed framework enabled rapid generation of complete 3D dose distributions for each patient's plan in a few seconds, demonstrating strong potential to accelerate clinical workflows. Conclusion: The proposed dense U-Net model demonstrated proficiency in accurately predicting dose distributions for the head and neck region, ensuring consistent quality.