FACA-Net: pancreatic CT image segmentation based on multi-domain feature enhancement and boundary morphology restoration

Abstract

Pancreatic computed tomography segmentation holds significant value in the early diagnosis and clinical intervention of pancreatic diseases. However, the pancreas’s small size, variable morphology, and blurred boundaries pose substantial challenges for segmentation. To address these challenges, FACA-Net, an improved network based on Attention U-Net, which incorporates multi-domain feature enhancement and boundary-morphology restoration modules, is proposed. Firstly, the FreqAsym feature selection module which consists of the Freqselect module and the asymmetric adaptive convolutional kernel module, is embedded within the skip-connection feature propagation path. The Freqselect module accurately captures pancreatic boundary details and suppresses irrelevant background noise through frequency-domain selection. Simultaneously, asymmetric adaptive convolutional kernels are introduced to dynamically learn and generate asymmetric receptive fields tailored to the pancreas’s elongated, irregular morphology. Then, the modified CARAFE content-aware upsampling is employed for feature upsampling and reconstruction. To further fuse the boundary details refined by FreqAsym, the FreqAsym output and decoder-side features are compressed to identical channels, fused, serving as input for the kernel prediction module of modified CARAFE. The dynamically generated adaptive kernels refine feature upsampling and preserve precise boundary information. Finally, Semantic Consistency Regularization is introduced as an auxiliary supervision strategy to facilitate multi-scale feature learning. Experimental results on the Medical Segmentation Decathlon (MSD) and NIH pancreatic datasets demonstrate significant improvements in both Dice similarity and 95% Hausdorff distance: On the MSD dataset, Dice reached 78.62% and HD95 decreased to 6.81 mm; on the NIH dataset, Dice reached 84.22% and HD95 decreased to 7.10 mm. These results validate the effectiveness of the improved model in pancreatic segmentation tasks, particularly in handling complex morphologies and regions with blurred boundaries.