An investigation of deep learning techniques for automatic pelvic CT scan segmentation

Radiotherapy treatment planning requires accurate delineation of organs at risk (OAR), which is typically a manual and time-consuming process. This research aims to explore the feasibility of using deep learning algorithms as an automatic tool for segmenting CT scan images. Accordingly, the performance of several convolutional neural networks (CNNs), including U-Net, Residual U-Net, and SegResNet, was compared as tools for automatic segmentation of OARs in pelvic CT scans (bladder, prostate, rectum, left femoral head, and right femoral head) against manual segmentation by specialists. This study involved 238 patients for prostate segmentation and 218 patients for the other four organs. The models' performance was assessed using metrics such as the Dice similarity coefficient, Jaccard index, and Hausdorff distance. The SegResNet model, providing the best performance, achieved Dice coefficients of 0.956, 0.832, 0.864, 0.980, and 0.985 for the bladder, prostate, rectum, left femoral head, and right femoral head, respectively. In summary, the results indicate that convolutional neural networks can accurately segment organs at risk in radiotherapy planning, with accuracies above 95% for bones and bladder, and over 83% for the rectum and prostate, while also speeding up the segmentation process.