(Sunday, 7/29/2018) 5:05 PM - 6:00 PM
Room: Room 209
Purpose: In cancer radiation therapy, it is often time consuming to develop a treatment plan that is satisfactory to the attending. To increase the efficiency of treatment planning, we have been developing deep learning model based on 3D dose distribution prediction based on patient contours before any dose calculation and plan optimization. We believe such predicted dose distributions can be used by dosimetrists and physicians as the references for planning, which will greatly reduce the number of iterations and improve both planning efficiency and plan quality. In this work, we test this idea by building a dose prediction model using U-net architecture for prostate patients treated with VMAT.
Methods: 3D U-Net architecture is used for dose prediction. 86 prostate patients with prescription dose of 45Gy are divided into three datasets – 61 for training, 15 for validation, and 10 for testing. The model input includes 6 structures of 128*128*96 pixels, which are PTV, bladder, left humeral head, right humeral head, rectum, and body. The output includes one channel of 128*128*96 pixels that is predicted 3D dose distribution. The selected loss function is mean squared error between the predicted dose and the true dose.
Results: The predicted 3D dose distributions are very similar to the ground truth. The mean dose difference inside the body is within 0.1%(D95), 2.3%(D98), 3.8%(D99) of the clinical dose. The mean difference for all structures are within 15% . Dose volume histograms are also very close between prediction and ground truth.
Conclusion: The trained 3D U-Net model is proven to be useful for predicting 3D dose distributions for prostate cancer patients treated with VMAT.
Funding Support, Disclosures, and Conflict of Interest: Cancer Prevention an Research Institute of Texas(CPRIT)(IIRA RP150485)
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
