Pancreas segmentation in CT based on RC-3DUNet with SOM

Deep learning-based automatic and accurate 3D pancreas segmentation plays a significant role in medical diagnosis and disease treatment, which has received a lot of attention from the medical image processing community. 3D pancreas segmentation embraces two challenges: one is that the pancreas occupies a relatively small proportion in CT images, and there is a serious class imbalance problem between foreground and background, which makes it difficult to achieve accurate segmentation. Another issue is that existing models are overly reliant on computer memory. In response to the above issues, we propose a two-stage training framework to alleviate the influence of background on segmentation results while reducing memory consumption. Among which, different from previous methods that used the entire pancreas region as input, we design a Selective Overlap Method (SOM) that can learn contextual information while reducing computing costs by selecting the most appropriate number of overlapping slices. In addition, a novel 3D segmentation model named ResConv-3DUnet (RC-3DUNet) is integrated into this framework, which not only maintains the receptive field while drastically reducing parameter number but also uses the residual information between layers to strengthen edge attention and regularizes the output by designing the supervision weights of each decoder layer. Extensive experiments were conducted on the Medical Segmentation Decathlon (MSD) pancreas segmentation and the National Institutes of Health (NIH) pancreas segmentation datasets, demonstrating that our method has a superior trade-off between accuracy and lightweight than existing approaches.