Program Information
Predicting Three Dimensional Dose Distribution with Deep Convolutional Neural Networks
Y Yuan*, T Tseng , M Chao , R Stock , Y Lo , Mount Sinai Medical Center, New York, NY
Presentations
TU-RPM-GePD-JT-1 (Tuesday, August 1, 2017) 3:45 PM - 4:15 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: To investigate the feasibility of using deep convolutional neural networks (CNN) to predict 3-D dose distribution
Methods: Using clinically approved plans as training data, a fully CNN model with 19 layers and 290,129 trainable parameters was trained to predict 3D dose distribution using patient-specific geometric parameters. Twenty-eight prostate patients were included in this study, which were planned with RapidArc to prescription dose of 45 Gy. From each OAR, a distance transform was firstly applied to generate a map where each pixel value represents the distance from this pixel to the border of the given OAR. These OAR distance maps served as the input to the CNN model and the output was the predicted 3D dose distribution. Instead of developing sophisticated human-craft features, we focus on designing appropriate network architecture and training strategies to ensure effective and efficient learning with limited training data. Our CNN model consists of two pathways, in which contextual information is aggregated via convolution and max-pooling in the convolutional path and full image resolution is restored via deconvolution and up-sampling in the deconvolutional path. Batch normalization was employed to reduce internal covariate shift while dropout and image augmentation were used to reduce overfitting. The predictive accuracy was evaluated by comparing the clinical plan with the predicted dose distribution
Results: With four-fold cross validation, the average predicted dose on PTV was found to be 46.5 ± 1.15 Gy, while the planned dose was 46.8±0.24 Gy. The average pixel-wise dose difference on PTV was 4.73%. All the predicted dose distributions met the OAR dose constraints in our institute
Conclusion: The preliminary results demonstrate that deep fully CNN has the potential to predict 3D dose distribution given patient’s geometric characteristics. We are expanding our database to improve the prediction accuracy and extending the framework to other sites
Contact Email: