A 3D Liver Semantic Segmentation Method Based on U-shaped Feature Fusion Enhancement

Semantic segmentation of 3D liver images is crucial for clinical diagnosis and case analysis. The majority of existing methodologies depend on the U-Net architecture, which has demonstrated limitations in effectively modeling long-range dependences. This shortfall impacts its ability to accurately capture subtle variations at the liver margin, often leading to segmentation inaccuracies. To overcome these limitations and enhance the local features within the global feature context, as well as to integrate multi-level semantic information, this study introduces an advanced 3D liver semantic segmentation method based on u-shaped feature fusion enhancement. The methodology begins with employing the Swin Transformer and CNNs as encoder and decoder respectively, subsequently proposing a novel U-shaped Feature Fusion (UFF) module that captures and merges multi-level features using a weighted mapping strategy. Ultimately, the features merged by the UFF module are accurately restored in terms of position and contour information through neural network learning, culminating in more accurate 3D liver semantic segmentation. Experimental results show that our SwinTU-Unet is superior in accuracy for liver segmentation. Compared to the current CNNs, transformers or their hybrids, our model shows a clear performance advantage.