WavPAC-Net: a wavelet-based hybrid encoder network with permuted axial-channel feature enhancement for liver segmentation

The liver, core to human metabolism and detoxification, is threatened by chronic diseases like cirrhosis that increase hepatocellular carcinoma (HCC) risk, with early diagnosis dependent on accurate liver segmentation in medical imaging. Current clinical segmentation methods generally suffer from low efficiency, strong subjectivity, and scarcity of AI-driven commercial products. Although deep learning technologies have been applied to liver segmentation, mainstream architectures still exhibit inherent limitations: CNN struggles with long-range dependencies while Transformer lacks fine detail representation. This study proposes WavPAC-Net, an innovative wavelet-based fusion network incorporating: (1) a hybrid CNN-Transformer encoder combining ConvNeXt-Small branch for precise hepatic parenchymal texture capture and Swin Transformer branch for long-range dependency modeling; (2) a Wavelet Domain Fusion Module (WDFM) that decomposes the features extracted by CNN and Transformer into low-frequency and high-frequency sub-bands respectively, and performs adaptive weighted fusion; and (3) a Permuted Axial-Channel Feature Enhancer (PACFE) that implements cross-dimensional refinement of high-dimensional fused features through collaborative three-dimensional spatial modeling and channel dimension permutation. In systematic comparisons with 20 state-of-the-art segmentation models, WavPAC-Net demonstrates statistically significant advantages in both segmentation accuracy and boundary fidelity. Evaluated on the CirrMRI600+ dataset (T2-weighted MRI), WavPAC-Net achieved Dice score of 93.16%, IoU of 88.45%, and MCC of 93.03%, with a computational complexity of 19.88 GFLOPs. Multi-modal validation on the public LiTS CT dataset further confirmed the model’s excellent performance, achieving Dice score of 96.91%, IoU of 94.32%, and MCC of 96.79%. Overall, WavPAC-Net provides an effective balance between segmentation accuracy and computational complexity, supporting its potential for practical liver segmentation applications.