Feasibility of reconstructing in-vivo patient 3D dose distributions from 2D EPID image data using convolutional neural networks

<i>Objective</i>. The primary purpose of this work is to demonstrate the feasibility of a deep convolutional neural network (dCNN) based algorithm that uses two-dimensional (2D) electronic portal imaging device (EPID) images and CT images as input to reconstruct 3D dose distributions inside the patient.<i>Approach</i>. To generalize dCNN training and testing data, geometric and materials models of a VitalBeam accelerator treatment head and a corresponding EPID imager were constructed in detail in the GPU-accelerated Monte Carlo dose computing software, ARCHER. The EPID imager pixel spatial resolution ranging from 1.0 mm to 8.5 mm was studied to select optimal pixel size for simulation. For purposes of training the U-Net-based dCNN, a total of 101 clinical intensive modulated radiation treatment cases-81 for training, 10 for validation, and 10 for testing-were simulated to produce comparative data of 3D dose distribution versus 2D EPID image data. The model's accuracy was evaluated by comparing its predictions with Monte Carlo dose.<i>Main Results</i>. Using the optimal EPID pixel size of 1.5 mm, it took about 18 min to simulate the particle transport in patient-specific CT and EPID imager per a single field. In contrast, the trained dCNN can predict 3D dose distributions in about 0.35 s. The average 3D gamma passing rates between ARCHER and predicted doses are 99.02 ± 0.57% (3%/3 mm) and 96.85 ± 1.22% (2%/2 mm) for accumulated fields, respectively. Dose volume histogram data suggest that the proposed dCNN 3D dose prediction algorithm is accurate in evaluating treatment goals.<i>Significance</i>. This study has proposed a novel deep-learning model that is accurate and rapid in predicting 3D patient dose from 2D EPID images. The computational speed is expected to facilitate clinical practice for EPID-based<i>in-vivo</i>patient-specific quality assurance towards adaptive radiation therapy.