Prediction of 3D Dose Distribution in Head and Neck Cancer with Convolutional Neural Network

Document Type

Conference Proceeding

Publication Date

10-1-2021

Publication Title

Cancer Clin Trials

Abstract

Background: Intensity-modulated radiation therapy (IMRT) is a commonly used technique for head and neck cancer, but the treatment planning process can be time-consuming due to the inherent complexity of the iterative optimizations. Knowledge-based planning (KBP) has been proposed as a solution to expedite the process. The currently commercially available method uses regression-based machine learning approach to predict dose-volume histograms to drive the optimizer in the inverse planning. Objectives: In this study, we aim to show the efficacy of convolutional neural network in predicting the 3-dimensional (3D) radiation dose distribution directly as part of KBP in head and neck cancer. Methods: We obtained the data of 340 patients with head and neck cancer from a public dataset as part of the 2020 OpenKBP challenge hosted by the American Association of Physicists in Medicine. The data for each patient consists of downsampled computed tomography (CT) images (128 x 128 x 128 voxels, with voxel size of approximately 3.5mm x 3.5mm x 2mm), binary masks for planning target volumes (prescribed to 70, 63, and 56 Gy) and 7 organs-at-risk (OAR), and the 3D radiation dose distribution delivered by IMRT technique with 9 equispaced coplanar beams. A 3D U-Net with encoder-decoder architecture was used as the model for prediction. The inputs to the model consisted of the simulation CT image with normalized Hounsfield units, masks for the PTVs, and a combined mask for the OARs. Based on the input information, the model predicted a 3D (128 x 128 x 128 voxels) radiation dose distribution in Gy. Three hundred cases were randomly selected as the training set, and the rest of 40 patients were used for testing only. Ten-fold cross validation was used during training, with mean absolute error (MAE) between the predicted and real doses within the body as the loss function. Adam was used as the optimizer with a learning rate of 0.0001. The algorithm was implemented in Python 3.8 with Tensorflow 2.2 as the framework. The network training was performed on both Google Colab and NVIDIA Tesla V100 GPU. The model performance was assessed by comparing the MAE between the predicted and real dose distributions. Results: The model was trained for 200 epochs, and model convergence was confirmed based on the validation loss (MAE<3.0 Gy for the last 10 epochs). On the independent test set consisting of 40 patients, the mean MAE for the predicted 3D dose distribution was 2.62 Gy per voxel (range, 1.00-6.67 Gy), 3.8% of the prescribed dose of 70 Gy. Conclusions: We showed that, with the CT images and contours of the PTVs and OARs as input, convolutional neural networks are excellent at predicting the radiation dose distributions that are similar to clinically delivered doses.

Volume

44

Issue

10

First Page

S141

Last Page

S142

Find It @ Sladen

Share

COinS