Room: Karl Dean Ballroom C
Purpose: Deep learning has demonstrated its tremendous power in organ segmentation. It uses a network to directly map an input image to an organ mask. However, success has been mainly achieved in organs with a relatively simple shape. When it comes to sigmoid colon segmentation, the method encounters difficulties. Due to complex 3d morphology, sigmoid can appear at different regions in a CT slice. The large inter-patient variation of shape and appearance poses further challenges. A human typically segments sigmoid in a slice-by-slice fashion. Contour in the slice is determined based on the information in the neighboring slices. Motivated by this fact, we develop an iterative deep-learning approach that mimics a humanâ€™s thought to segment sigmoid colon.
Methods: The main idea of our approach is to develop a network that takes a CT slice, sigmoid contour on this slice, and the neighboring CT slice as input and predict sigmoid contour on the neighboring CT slice. We employed a U-net structure trained with data from six patient cases. To prevent over-fitting in train session, we made 72500 training data sets via data augmentation through image translation, rotation, and scaling. After the network was trained, we tested it in two patient cases that were not included in training. We started the segmentation with an initial sigmoid contour drawn at a CT slice separating rectum and sigmoid. We then sequentially loop through CT slices inferior-to-superior direction and then superior-to-inferior direction with multiple rounds. This looping captures the entire sigmoid despite its complex morphology.
Results: In the train stage, dice similarity coefficient (DSC) was >0.94 after 30 epochs. In the test data, the sigmoid was successfully segmented with DSC of 0.86.
Conclusion: We developed a novel iterative deep learning approach to enable accurate segmentation of sigmoid colon with a complex morphology.
Brachytherapy, Treatment Planning, Segmentation