DO-UNet:Improved UNet Model for CT Image Segmentation using DBB and Octave Convolution

Abstract Artificial intelligence techniques such as machine learning and deep learning are prevalent in the field of medical imaging, especially image segmentation. In the field of medicine, image segmentation helps to diagnose diseases accurately and hence has sparked interest. Neural-driven image segmentation methods, especially U-Net, are prominent.U-Net, although effective, suffers from severe information loss during downsampling and insufficiently detailed texture of the segmentation results, which limits its application. To solve this problem, we introduce DO-UNet, which strengthens the structure of U-Net. DO-UNet employs jump links at the downsampling layer, preserves high-resolution features, and suppresses data loss. Dense Block with Bottleneck (DBB) replaces the 3*3 convolutional block to efficiently extract spinal features and improve performance. Octave convolution extends the model during upsampling, improving accuracy by expanding the feature space and minimizing parameters. Experiments on the Verse2019 and Verse2020 datasets show that DO-UNet outperforms other models in terms of dice coefficients and accuracy. Visualization results confirm DO-UNet's superiority in accurately segmenting spinal structures. The model excels in capturing details and shows robustness and interpretability. These results highlight the practical significance and promising future of DO-UNet in spine segmentation.