Room: Exhibit Hall | Forum 2
Purpose: The focus of this work was to develop a machine learning strategy for performing prediction of IMRT dose distribution using only contours in planning CT.
Methods: Our approach employed deep convolutional neural network (DCNN) that learn the dose distributions generated from eighty prostate cancer patients treated with IMRT. A supervised learning approach was used for the deep learning process, where the dose for a set of voxels in the training set was taken as the label for the learning process. The training data come from existing patients with clinical dose distribution and contours of planning target volume (PTV) and organ at risks (OAR). A batch-wise Adaptive Moment Estimation method was employed for leaning and optimizing a 21-layer deep neural network model consisted only of convolutional layers. The performance of our model was evaluated by comparing the predicted dose prediction against the clinical dose distribution, where we performed four-fold cross-validation and normalized dose distribution to the maximum dose.
Results: Average gamma passing rates with standard deviation (criteria: 5%/5mm, threshold = 50%) were 78.9Â±6.66%. The average differences between clinical and predicted dose distributions were -6.76% and -4.88% in maximum dose of rectum and bladder, respectively, whereas they were -3.38% and -4.74% in mean dose of rectum and bladder, respectively. Our result showed the strong result using DCNN to predict the dose.
Conclusion: A novel DCNN method was developed to accurately and efficiently predict dose distribution from patientâ€™s contours. Our results showed the potential of DCNN to predict dose distribution, suggesting that DCNN method is useful for maintaining plan quality and saving planning time.