ECTU-Net: Efficient Channel Attention and Transformer Integration for Liver Tumor Segmentation in CT Images.
Authors
Affiliations (1)
Affiliations (1)
- School of Physics and Information Technology, Shaanxi Normal University, Shaanxi Key Laboratory of Ultrasonics, Xi'an, Shaanxi, China.
Abstract
U-Net and its variants have demonstrated strong local feature extraction in medical image segmentation. Efficient Channel Attention-Transformer U-Net (ECTU-Net) is proposed to recover fine boundaries and refine multi-scale features. A lightweight four-layer U-Net serves as the backbone. Efficient Channel Attention (ECA) modules enhance channel-wise feature selection, and compact Transformer blocks learn global dependencies. A Flatten-Transformer-Reshape process at the bottleneck replaces high-dimensional convolutions to capture pixel-level long-range dependencies. On LiTS17, ECTU-Net improves the Dice Similarity Coefficient (DSC) by 9.20% over the original U-Net and reduces FLOPs by 42.3%. The Transformer block adds only 0.09M parameters while improving global context modeling. ECTU-Net outperforms the comparison models in liver tumor segmentation and accurately delineates tumor regions. It combines high segmentation accuracy with parameter efficiency. However, the limited sample size restricts the generalizability of the findings. ECTU-Net enables high-precision medical image segmentation in resource-constrained settings and shows strong potential for clinical application.