Accurate Segmentation Method of Sacroiliac Joint CT Images Based on Improved U-Net

To address the challenges of automatic segmentation in sacroiliac joint CT images caused by complex bone structures and narrow joint spaces, this paper proposes an improved 3D U-Net architecture that significantly enhances segmentation accuracy through embedding a Squeeze-and-Excitation (SE) channel attention module in the bottleneck layer. As sacroiliac joint segmentation is critical for early diagnosis of ankylosing spondylitis (AS), yet existing automated methods struggle with low contrast joint spaces and heterogeneous bone densities in CT images, our method utilizes the SE module's dynamic channel recalibration mechanism to enhance key feature channel weights in the deepest semantic bottleneck layer. This approach effectively resolves traditional methods' issues of over-segmentation in high-density bone cortex areas and under-segmentation in joint spaces. Evaluation on a clinical dataset of 40 sacroiliac joint CT scans demonstrates superior performance, achieving a 91.4% Dice coefficient and 84.3% IoU, representing improvements of 1.1% and 1.7% over the baseline 3D U-Net, respectively, with particular advantages in segmenting joint surface erosion areas in AS patients.