Abstract

Info

Title:: Pancreas segmentation algorithm based on depth-wise convolution and tri-orientated spatial attention

Author(s):: TAN Lulu1; FENG Qianjin1; 2; 1. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China 2. Guangdong Provincial Key Laboratory of Medical Image Processing, Guangzhou 510515, China

Keywords:: Keywords: pancreas depth-wise convolution tri-orientated spatial attention cascaded network

Abstract:: Abstract: A cascaded 3D pancreas segmentation network (CPS-Net) is proposed to address the challenges in pancreas segmentation caused by its small size and complex anatomical structure. CPS-Net is composed of two components: the first part utilizes ResUNet to quickly localize the pancreas region, while the second part uses a network that fuses depth-wise convolution block and tri-orientated spatial attention module to refine the segmentation results. Specifically, depth-wise convolution block significantly enhances the differentiation between the pancreas and surrounding tissues by extracting multi-scale features layer by layer, while tri-orientated spatial attention module combines axial attention, planar attention and window attention mechanisms to comprehensively capture the detailed structure of the pancreas in a complex background. CPS-Net achieved Dice similarity coefficient, positive predictive value, sensitivity, and Hausdorff distance of 87.42%±1.58%, 87.42%±3.52%, 87.74%±4.58%, and (0.22±0.08) mm, respectively, on the NIH public dataset, demonstrating its higher pancreas segmentation accuracy and superior performance compared with the current state-of-the-art segmentation networks.