Research on automatic segmentation method of stroke lesions joint with image registration
To address the challenge of relatively low segmentation accuracy in chronic stroke lesions,we propose a novel approach for automatic segmentation in chronic stroke using joint deep image registration.The method employs a deep Laplacian pyramid image registration network to obtain brain tissue partitions in a coarse-to-fine manner within the space of diffeomorphic mappings,acquiring anatomical prior information for lesion locations.The original MRI images and the results from the registration phase are then combined and fed into a U-Net augmented with channel and spatial attention modules for lesion segmentation.Testing on the publicly available ATLAS dataset demonstrates the effectiveness of the proposed method in improving the accuracy of chronic stroke lesion segmentation.It achieves a 4.4%improvement over the classical 2D U-Net,highlighting that the deep image registration-based brain partition prior effectively enhances the model's segmentation performance.Moreover,superior tissue segmentation contributes to better lesion segmentation accuracy.
国家科技期刊平台
NSTL
万方数据
