3D stripe CD-LinkNet: liver vessel segmentation network integrating stripe attention and 3D convolutional operations

Accurate segmentation of liver vessels is crucial for liver cancer surgery planning. The hybrid network combining Transformer and Convolutional Neural Network (CNN) can extract local and global features and achieve high performance in medical image segmentation. However, current hybrid networks use square patches to embed image regions, which may limit the extraction of long stripe features, especially for liver blood vessels in Computed Tomography (CT) images. To address this problem, we propose a new hybrid network to extract long stripe features for the segmentation of liver blood vessels in CT images. This network is a U-shaped 3D network that uses CNN encoders to extract local features and cascades a Transformer block at the bottom layer to extract long-term representations. The Transformer block consists of three components: strip based self attention, square based self attention, and two cascaded convolutional layers. The extraction of stripe features based on stripe attention, combined with the long-distance features extracted by square based attention and the local features extracted by convolutions, can be better used for segmentation of liver blood vessels. To evaluate this network, we tested on the public 3D-IRCADb-01 dataset and compared with existing approaches. Experimental results demonstrate that our method achieves the highest segmentation accuracy, with a Dice score of 0.6661. The proposed approach is specifically designed for vessel segmentation within the liver region and can be integrated with liver segmentation techniques.