Purpose: Brachytherapy (BT) combined with external beam radiotherapy (RT) is the current standard of care for cervical cancer patients. Accurate segmentation of organs at risk (OAR) is a prerequisite step for RT planning. In this study, we propose a two-step automatic multi-organ segmentation method using convolutional neural network (CNN) in female pelvic magnetic resonance (MR) images.
Methods: Our segmentation model consists of two cascaded steps; 1) organ localization and 2) fine segmentation of individual organs. The first step employs a 3D U-Net to localize each organ from which organ-specific regions of interest (ROIs) are produced. The original MR volume is then cropped, based on the ROIs, yielding organ-specific volumes for the next step. For fine segmentation, modified 3D Dense Block (BD) is embedded into a 3D U-Net-like architecture to simultaneously capture intra- and inter-slice information while facilitating information flow in the network. A total of 221 3D T2-weighted MR volumes from 135 patients, were used for this study. Training images were augmented using shift and rotation operations and the networks were trained for 200 epochs using a total of 804 MR volumes.
Results: The proposed method was applied to segment the bladder, rectum, and sigmoid. Compared to the expert radiation oncologists’ manual segmentations, average dice similarity coefficients (DSC, mean ± standard deviation) for 20 test cases are 0.93±0.04 (bladder), 0.87±0.03 (rectum), and 0.78±0.09 (sigmoid). Our methods was also compared to the standard 3D U-Net and 3D Dense U-Net and showed superior performance (p<0.05). Using this newly created model, the segmentation takes 7-8 seconds per each organ.
Conclusion: We proposed a novel fully automated multi-organ segmentation algorithm for female pelvic MR images. Our experimental results demonstrate that the developed method is accurate, fast, and reproducible, potentially reducing the burden of tedious manual image contouring and thus improving efficiency in RT planning.