Room: AAPM ePoster Library
To develop a method for predicting clinical cases (rectum and left breast cancer) of 3D dose distribution, given the radiation therapy (RT) planning and segmented anatomy image, by applying deep learning techniques to a database of previously optimized and approved treatment plans.
We have modified a deep convolutional neural network (CNN) model, fusion-net (originally designed for cell structure segmentation purposes), for predicting dose from patient-specific image contours of the planning target volume and organs at risk (OAR). A database of volumetric modulated arc therapy (VMAT) plans for the fifty left breast cancer patients and 3D-conformal radiation therapy (CRT) plans for fifty rectum cancer patients used for training and validation data sets. The volume overlap error (VOE), relative volume difference (RVD), dice similarity coefficient (DSC), and dose volume histogram (DVH) were used for the quantitative evaluation of 3D dose prediction.
The deep neural network model predicted 3D dose distribution accurately except bladder and left lung structure. In case of rectum cancer, the average value of the absolute differences in [max, mean] prescription dose, specifically for each structure is [0.88%, 0.82%] (GTV), [1.24%, 1.08%] (CTV), [-0.70%, -4.68%] (bladder), and [0.72%, 1.24%] (Body) of the prescription dose. In case of left breast cancer, the average value of the absolute differences in [max, mean] prescription dose, specifically for each structure is [-0.02%, -2.10%] (PTV), [-0.15%, 1.30%] (left breast), [0.80%, -1.60%] (heart), [-1.43%, 0.30%] (left lung), -5.68%, 1.80%] (right lung), and [-0.02%, 1.30%] (body) of the prescription dose.
In this study developed a deep learning method for 3D voxel dose prediction, and shown to be able to produce accurately patient’s specific dose prediction for breast and rectum cancer. The use of deep learning techniques in clinical plans will enable accurate dose prediction, optimization to save time, and maintain high-quality plans.