Double Structure Constrained Cerebrovascular Segmentation Network with Sparse Labels

Deep learning-based cerebrovascular segmentation methods are difficult to segment cerebral vessels with good connectivity under sparse labels. A double structure constraint cerebrovascular segmentation network including encoder, decoder and structural attention module is proposed. The sagittal and coronal features are extracted to construct plane attention. The structure attention mechanism is constructed by combining with channel attention to constrain the cerebrovascular structure at the network level. Central line Dice loss function improved by equalization coefficient is added to Dice loss function to preserve the connectivity of vascular structure and constrain vessels at the topological structure level. Experimental results on TubeTK show that compared with the four attention networks, the DSC of the proposed method is improved by 4.58%–6.86%, the IOU is improved by 5.07%–7.47% and the center line Dice is improved by 3.26%–5.40%.