Automatic Bladder Delineation on MR Images Using a Convolution Neural Network for Online Image-Guided Radiotherapy

Abstract

Adaptive radiotherapy for bladder cancer is challenging due to its inter- and intra-fractional anatomical variations appeared during the course of treatment. In order to adapt for these changes, anatomical information of patient body needs to be monitored on time. Magnetic resonance imaging (MRI) is able to acquire patient images continuously without any radiation exposure, whereas fast bladder delineation method for MR images is still developing technique. This study aims to develop a method that automatically extracts bladder contours from the MR images using a convolutional neural network (CNN). Total 1,060 pairs of a pelvic T2 weighted axial MR image and its segmented bladder contour, which were delineated by an experienced radiation oncologist, were collected from 100 patients. All MR images had 2.5 mm in slice thickness and were acquired with an MRI scanner. This dataset was divided into 840 image pairs (80 patients) for training, 113 (10) for validation and 107 (10) for test. A U-net-based 2D CNN model was generated. Input and label of the model were set to 2D MR image and its contour, respectively. 100-epoch end-to-end training was applied for the CNN model with 5 mini-batches and data augmentation of random image flip and 90° rotations. Delineation accuracy of the trained CNN model was evaluated with Dice similarity coefficient (DSC). The CNN model is able to calculate bladder contours with high accuracy. A median DSC value over all test data was 94.4%. Despite of this high accuracy, relatively low DSC values were observed for MR images including only bladder margins. We succeeded in developing a bladder delineation model for MR images with high accuracy. Model performance for MR images with bladder margins could be insufficient with 2D CNN, and 3D CNN is expected to improve accuracy. This technique will be utilized in image-guided radiotherapy.