Abstract:Dose volume histogram (DVH) metrics are widely accepted evaluation criteria in the clinic. However, incorporating these metrics into deep learning dose prediction models is challenging due to their non-convexity and non-differentiability. We propose a novel moment-based loss function for predicting 3D dose distribution for the challenging conventional lung intensity modulated radiation therapy (IMRT) plans. The moment-based loss function is convex and differentiable and can easily incorporate DVH metrics in any deep learning framework without computational overhead. The moments can also be customized to reflect the clinical priorities in 3D dose prediction. For instance, using high-order moments allows better prediction in high-dose areas for serial structures. We used a large dataset of 360 (240 for training, 50 for validation and 70 for testing) conventional lung patients with 2Gy $\times$ 30 fractions to train the deep learning (DL) model using clinically treated plans at our institution. We trained a UNet like CNN architecture using computed tomography (CT), planning target volume (PTV) and organ-at-risk contours (OAR) as input to infer corresponding voxel-wise 3D dose distribution. We evaluated three different loss functions: (1) The popular Mean Absolute Error (MAE) Loss, (2) the recently developed MAE + DVH Loss, and (3) the proposed MAE + Moments Loss. The quality of the predictions was compared using different DVH metrics as well as dose-score and DVH-score, recently introduced by the AAPM knowledge-based planning grand challenge. Model with (MAE + Moment) loss function outperformed the model with MAE loss by significantly improving the DVH-score (11%, p$<$0.01) while having similar computational cost. It also outperformed the model trained with (MAE+DVH) by significantly improving the computational cost (48%) and the DVH-score (8%, p$<$0.01).
Abstract:Deep learning (DL) 3D dose prediction has recently gained a lot of attention. However, the variability of plan quality in the training dataset, generated manually by planners with wide range of expertise, can dramatically effect the quality of the final predictions. Moreover, any changes in the clinical criteria requires a new set of manually generated plans by planners to build a new prediction model. In this work, we instead use consistent plans generated by our in-house automated planning system (named ``ECHO'') to train the DL model. ECHO (expedited constrained hierarchical optimization) generates consistent/unbiased plans by solving large-scale constrained optimization problems sequentially. If the clinical criteria changes, a new training data set can be easily generated offline using ECHO, with no or limited human intervention, making the DL-based prediction model easily adaptable to the changes in the clinical practice. We used 120 conventional lung patients (100 for training, 20 for testing) with different beam configurations and trained our DL-model using manually-generated as well as automated ECHO plans. We evaluated different inputs: (1) CT+(PTV/OAR)contours, and (2) CT+contours+beam configurations, and different loss functions: (1) MAE (mean absolute error), and (2) MAE+DVH (dose volume histograms). The quality of the predictions was compared using different DVH metrics as well as dose-score and DVH-score, recently introduced by the AAPM knowledge-based planning grand challenge. The best results were obtained using automated ECHO plans and CT+contours+beam as training inputs and MAE+DVH as loss function.