J Yuan, R Kashani*, N Zaorsky, D Spratt, Y Zheng University Hospitals, Cleveland Medical Center, Cleveland, 44106 OH

• 
  R Kashani

TU-D1030-IePD-F6-2 (Tuesday, 7/12/2022) 10:30 AM - 11:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: To predict dose distribution for prostate radiation volumetric modulated arc therapy (VMAT) using convolutional neural network of deep learning.

Methods: Twenty five VMAT plans for prostate cancer were retrospectively selected that treated the prostate and seminal vesicles. The dose calculation volume was resampled to 256 x 256 x number of slices. The contours (PTV, bladder, rectum, right and left femoral head and body) were converted into binary mask volumes. With the density volume obtained from the CT image, there were 7 channels in the input to the neural network. The dose distributions were normalized by setting the PTV mean dose to 1. A convolutional deep network model, U-net, was used to train on these contour-to-dose paired dataset. A fivefold cross-validation procedure was performed in which the 20 cases were used as the training dataset and the remaining 5 cases were used as the test set. The metrics for evaluating the accuracy of predictions were voxel-wise mean absolute error (MAE), mean squared error (MSE) between the predicted dose and the true dose, and Dice similarity coefficient (DSC) for isodose volumes between 10% and 100% of the prescribed dose.

Results: The mean value with standard deviation of MAE and the square root of MSE from our model for overall cases is 3.6% ± 1.3% and 6.9% ± 1.7% of prescription dose, respectively, and for the DSC is 0.82 ± 0.04. The inaccuracy occurs at low-dose regions away from the PTV. The time needed to make a patient dose prediction is less than 5 seconds.

Conclusion: The model is able to reasonably predict the dose distribution for prostate radiotherapy based on a patient’s contours. Further improvement for the low-dose regions will include taking the beam arrangement into account.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email