Quality control of VMAT planning using artificial neural network models for nasopharyngeal carcinoma

Abstract

Objective To train individualized three-dimensional (3D) dose prediction models for radiotherapy planning, and use the models to establish a planning quality control method . Methods A total of 99 cases diagnosed as early nasopharyngeal carcinoma (NPC) were analyzed retrospectively, who received simultaneous integrated boost (SIB) with volumetric modulated arc therapy (VMAT). Seven geometric features were extracted, including the minimum distance features from each organs at risk (OARs) to planning target volume (PTV), boost targets and outline, as well as four coordinate position characteristics.89 cases were trained and 10 cases were tested based on 3D dose distribution prediction models using artificial neural network (ANN). A planning quality control method were established based on the prediction models. The dosimetric parameters including D2%, D25%, D50%, D75% and mean dose (MD) of each OAR were used as quality control indicators, and the passing criteria was defined as that the dosimetric difference between manual planning and the predicted dose should be less than 10%. The quality control method was tested with 10 plans designed by a junior physicist. Results There was no significant discrepancy between the model predicted dose and the result of expert plan in the main dosimetric indexes of 18 OARs. The dose differences of D2%, D25%, D50%, D75% and MD were all controlled within 1.2 Gy.All the 10 plans designed by a junior physicist reached the general clinical dose requirements, while by using our proposes quality control method, one of these plans was observed not optimal enough and some dosimetric parameters of spinal cord, spinal cord PRV, brainstem and brainstem PRV could be improved. After re-optimizing this plan according to the predicted values of the model, the D2% of spinal cord and brainstem decreased by 8.4 Gy and 5.8 Gy, respectively. Conclusions This study proposes a simple and convenient quality control method for radiotherapy planning. This method could overcome the disadvantage of unified dose constrains without considering patient-specific conditions, and improve the quality and stability of individualized radiotherapy planning. Key words: Radiotherapy planning; Dose prediction; Artificial neural network; Quality control; Nasopharyngeal carcinoma