Room: Davidson Ballroom A
Purpose: To develop a deep-learning method for voxel-by-voxel dose prediction for tomotherapy.
Methods: Using previously treated plans as training data, a predictive model based on deep-learning algorithm was trained to predict a 3D dose distribution in view of patient geometrical features. First, we converted the contours and dose volumes from plan database to 3D matrix by using a developed visualization toolkit (VTK) based algorithm, and transferred to the model for correlating anatomical features and dose distributions. Nasopharyngeal cancer (NPC) patients treated by helical tomothrapy (TOMO) were studied. Using these clinically approved plans, the models were trained to predict new patient 3D dose distribution and the predictive accuracy was evaluated using the dose difference between the prediction and clinical plan. The mean and standard deviation were evaluated to assess prediction bias and precision.
Results: The predictive models based on deep-learning method demonstrated highly accurate dose distribution prediction. 100 NPC patients with TOMO plans were used to train the model and the preliminary results of the averaged voxel-by-voxel dose differences between the predicted and actual values varied from -2% to 9% corresponding to the prescription dose. The results present the feasibility of voxel-by-voxel 3D dose prediction in NPC cases.
Conclusion: The method demonstrates accurate 3D voxel-by-voxel dose predictions for NPC patients. The predicted dose map can be useful for ensuring the treatment plan quality and consistency, and guiding automatic planning.
Not Applicable / None Entered.