Room: Karl Dean Ballroom A2
Purpose: This work presents a direct prediction of fluence-map from patient anatomy for 7-field prostate intensity modulated radiotherapy (IMRT) using deep-neural-network.Materials &
Methods: The plan optimization for IMRT involves the process of finding optimal fluence-map with a given specific patient anatomy. We aim to substantially enhance the planning efficiency by directly predicting the fluence-map from the given patient anatomic information by use of a deep neural network. We constructed a network architecture similar to the U-Net, except for replacing all the max-pooling layers by convolutional layers with a stride of 2 for a better performance. Total 261 prostate plans (7-field IMRT) for 44-Gy dose delivery were collected for training with the input data being prepared by organ-labeled cross-sectional images perpendicular to the beam direction. During training, the network was designed to minimize the mean-absolute difference between true fluence-maps generated by EclipseTM treatment planning system (TPS) and predicted ones by deep-neural-network. For evaluation, we chose 20-patient data that were not used in training the network. The predicted fluence-map was compared to the true fluence-map using mean absolute error (MAE). The dose distributions produced from the true and predicted fluence-maps were used to compare Dâ‚‰â‚… of the planning target volume (PTV), and Dâ‚…â€™s of the rectum and bladder (i.e., the minimum dose received by the hottest 5%).
Results: MAE of the fluence-map produced by the deep-neural-network was 5.14Ã—10â?»Â³. From the dosimetric aspect, for the 20 test cases, mean errors of PTV Dâ‚‰â‚…, rectum Dâ‚…, and bladder Dâ‚… were -1.86 Gy, -1.29 Gy, and -0.76 Gy, respectively.
Conclusion: In this study, we constructed a deep-neural-network predicting fluence-map from the patient anatomy for prostate IMRT with a high fidelity. To our best knowledge, this is the first attempt to utilize deep-neural-network for predicting fluence-map in IMRT.