PAFU-Net: A Pyramidal Attention-Fortified U-Net for Precise Liver Segmentation in Abdominal CT scans

• Andrews Jose,
• Sujitha Juliet,
• J. Anitha,

Vol. 20, No. 2, February 28, 2026
10.3837/tiis.2026.02.005, Download Paper (Free):

• Liver Segmentation
• Attention U-Net
• Computed tomography
• Atrous Spatial Pyramid Pooling
• Convolutional neural networks

Abstract

The computer-aided diagnosis, surgical planning, treatment monitoring, and volumetric assessment of livers require proper liver segmentation by abdominal CT scans. Our architecture is PAFU-Net which is a pyramidal attention-strengthened U-Net architecture that incorporates Atrous Spatial Pyramid Pooling (ASPP) with a hybrid architecture of encoder channel attention, a bottleneck CBAM architecture, and decoder spatial attention. Moreover, ASPP-gated skip connections are used to add more multi-scale contextual representation and at the same time reducing background leakage and enhancing feature consistency within the network. Soft-tissue contrast and noise reduction prior to network inference are further improved by a lightweight preprocessing pipeline which includes slice elimination, HU-based windowing, and CLAHE. On the LiTS and 3D-IRCADb, PAFU-Net scores 0.9824 and 0.9277 on the Dice scores, respectively, and outperforms the standard U-Net and Attention U-Net models, as well as CBAM and ASPP-based ones, with much fewer parameters. The aggregate number of results shows high boundary preservation, excellent cross-dataset generalization and evident potential to be used in practical clinical settings in liver-segmentation processes.

Statistics

Cite this article