A detail-oriented super-2D network for pulmonary artery segmentation

In medical image segmentation, automatic pulmonary vascular segmentation is very important for the diagnosis of pulmonary vascular lesions. However, the pulmonary vascular has problems such as a small cross-sectional area and morphological approximation to other pulmonary interstitial tissues, which results in low segmentation accuracy for fine vessels. Moreover, the 3D and 2D networks applied to medical image segmentation have high computational costs and low segmentation accuracies, respectively. To address these issues, we propose a detail -oriented super -2D network, named DS2Net, that focuses on the semantic reinforcement of fine vessels and computational cost optimization. Aiming at the high computational costs of 3D networks and the low segmentation accuracies of 2D networks, a super -2D segmentation pattern is adopted, which uses consecutive computed tomography (CT) slices as inputs to build 3D correlations in 2D network and provides volume spatial information to the network with a lower computational cost. Then, to achieve improved segmentation accuracy for fine vessels in the pulmonary vascular segmentation task, we propose an OffsetConv algorithm, which extends the feature scope of fine vessels via convolutional diffusion combined with morphological diffusion and solves the problem regarding the ease of fine vessel feature loss during multi scale transformation. Qualitative and quantitative experimental results obtained on the public Parse22 pulmonary artery segmentation dataset show that our proposed DS2Net is superior to other 2D networks that are widely used in medical image segmentation and achieves segmentation accuracy close to that of the 3D benchmark network under the premise of a much lower computational cost. Concretely, our DS2Net achieves average DSC scores of 78.16% for the overall CT slices and 61.53% for a single CT slice on Parse22, and it significantly surpasses other networks in terms of fine vessel segmentation. Code is available at https://github.com/LuffyMonsterB/DS2Net.