OA236] Patient-specific quality assurance based on electronic portal imaging devices via artificial neural networks

Abstract

Purpose In this work, Artificial Neural Networks (ANN) are used in External Beam Radiotherapy, especially for dosimetry purposes using Electronic Portal Imaging Device (EPID), which needs frequent calibration and complex setting of dedicated software. This work is preliminarily oriented towards quality assurance of intensity-modulated radiotherapy (IMRT). This first step will allow adjusting ANN before their implementation in more complex treatments and in vivodosimetry. Methods Like humans, artificial neural networks work via two phases – learning and recognition. The supervised learning phase entails the creation of “neurons” by linking input and output data. The recognition phase consists of using input data to predict output data. During the learning phase, EPID signals and desired absorbed dose distributions (taken from treatment planning system – TPS) are used as the neural network’s input and output, respectively. Once the learning is complete, new EPID signal can be used to predict the delivered absorbed dose distribution, allowing comparison with the TPS. EPID images were taken during 6MV treatment delivery, with a dose rate of 600 MU/min, from aSi1000 EPID (Varian). Mounted with Exact-arm on a Clinac 23iX equipped with a multi-leaf collimator (120 leaves). The EPID were acquired using the half-resolution mode. 2D plane images were calculated in EclipseTM at the maximum depth dose in a water phantom. Results Learning was performed using 11 input/output datasets from IMRT treatments. All of the used datasets (both EPID inputs and absorbed dose distribution outputs) consisted of 384 × 512 pixels. Learning can be time consuming but once the ANN has been fixed, its use during the recognition phase will be instantaneous. The gamma index, γ, was used to evaluate the difference between the ANN calculated and planned distributions. γ gives the number of pixels (as a percentage) that respect a given objective. γ(2% , 2 mm) for Head and Neck cancers was found to be 99.7%, highlighting the ANN capability to predict the absorbed dose distribution based on EPIDs. Conclusions It was shown that patient-specific quality assurance of IMRT based on EPID can be performed with neural networks algorithms. Next work would be extending algorithms for in vivo dosimetry purpose.