Image Segmentation in 3D Brachytherapy Using Convolutional LSTM

Purpose: The accuracy of the segmentation of the target lesion and at-risk surrounding organs is important for cervical cancer patients treated with three-dimensional (3D) brachytherapy. However, the nature of brachytherapy, organ deformities, metal induced artifacts caused by applicators, and limited operating time make this process a challenge. Deep learning segmentation has recently emerged as an approach to this problem. The basic concept proposed in this paper is accurate segmentation of a pelvic computed tomography image dataset such that patients can obtain more accurate radiotherapy. Methods: In this work, we propose a solution based on 3D U-Net and Long Short-Term Memory (LSTM). The model was trained and tested on a computed tomography pelvic image dataset comprising 51 patients who underwent 3D brachytherapy. The organs that required segmentation included the bladder, bowels, sigmoid colon, rectum, and uterus. We also used self-ensemble to improve segmentation accuracy. The Dice coefficient was used as the evaluation metric to determine the segmentation results for all of the organs under consideration. Results: The proposed model was evaluated using organ segmentation obtained from 10 patients with newly diagnosed stage I-IVA cervical cancer. The test results showed that segmentation conducted using the proposed model resulted in the following Dice coefficient values (mean, ± standard deviation): 87 (± 6.3) %, 72 (± 9.1) %, 72 (± 7.8) %, 95 (± 3.5) %, 86 (± 4.9) %, 77 (± 8.4) %, 73 (± 10.2) %, and 93 (± 3.5) % for the HRCTV, GTV, bowels, foley, bladder, rectum, sigmoid colon, and uterus, respectively. Conclusion: This method shows that the combination of 3D U-Net and LSTM and self-ensemble post-processing has high potential for segmentation of a pelvic computed tomography dataset.