Dose prediction for cervical cancer VMAT patients with a full-scale 3D-cGAN-based model and the comparison of different input data on the prediction results

Abstract

Purpose: To develop a 3D dose distribution prediction deep learning model for volumetric modulated arc radiotherapy (VMAT) of cervical cancer, and to explore the impact of different multichannel input data on the prediction accuracy, especially to prove the feasibility of dose prediction only based on computed tomography (CT) images and planning target volume (PTV) delineated contours.Methods: A total of 118 VMAT cases were collected, which were made into three datasets with different multichannel combinations. In addition to the clinical dose distribution data occupying one channel, the three datasets were as follows: Dataset-A, 7 channels, included CT images, the PTV and the organs at risk (OARs); Dataset-B, 2 channels included CT images and the PTV; Dataset-C, a single channel, included only CT images. A full-scale feature fusion 3D conditional generative adversarial network (cGAN) based dose distribution prediction architecture was proposed, with multiple losses function used as the optimization target. Under this framework, three models were obtained by training with the three datasets: Model-A, Model-B and Model-C. The following indicators were used to evaluate and compare the performance of the models: (1) the 3D dose difference map and the mean absolute error (MAE); (2) the dose-volume histogram (DVH) curve;(3) the dose index (DI) of the PTV and OARs; (4) the Dice similarity coe cient (DSC).Results: The proposed model accurately predicts the 3D dose distribution. For the twenty test patients, the MAE of Model-A is 1.1±0.2%, the MAE of Model-B and Model-C are 1.4±0.2%, and 1.9±0.3% respectively. There are no signi cant differences between the DIs of PTV (D99%, D98%, D95%, HI and CI) and OARs of Model-A and Model-B from the clinical result. The average DSC of Model-A with different isodose volumes is greater than 0.94. Model-B and Model-C follow with average DSCs greater than 0.91 and 0.86 respectively. Conclusion:We propose a new dose prediction model based on a full-scale fusion and generative adversarial architecture, con rming the feasibility of dose prediction based only on CT images and the PTV. The method proposed provides a simpler and more effective method for clinical dose assessment, radiotherapy planning assistance and automatic planning.