AGMR-Net: Attention Guided Multiscale Recovery framework for stroke segmentation
Automatic and accurate lesion segmentation is critical for clinically estimating the lesion statuses of stroke diseases and developing appropriate diagnostic systems. Although existing methods have achieved remarkable results, further adoption of the models is hindered by: (1) inter-class indistinction, the normal brain tissue resembles the lesion in appearance. (2) intra-class inconsistency, large variability exists between different areas of the lesion. To solve these challenges in stroke segmentation, we propose a novel method, namely Attention Guided Multiscale Recovery framework (AGMR-Net) in this paper. Firstly, a coarse-grained patch attention module in the encoding is adopted to get a patch-based coarse-grained attention map in a multi-stage explicitly supervised way, enabling target spatial context saliency representation with a patch-based weighting technique that eliminates the effect of intra-class inconsistency. Secondly, to obtain a more detailed boundary partitioning to solve the challenge of the inter-class indistinction, a newly designed cross-dimensional feature fusion module is used to capture global contextual information to further guide the selective aggregation of 2D and 3D features, which can compensate for the lack of boundary learning capability of 2D convolution. Lastly, in the decoding stage, an innovative designed multi-scale deconvolution upsampling instead of linear interpolation enhances the recovery of target space and boundary information. The AGMR-Net is evaluated on the open dataset Anatomical Tracings of Lesions-After-Stroke (ATLAS), achieving the highest dice similarity coefficient (DSC) score of 0.594, Hausdorff distance of 27.005 mm, and average symmetry surface distance of 7.137 mm, which demonstrate that our proposed method outperforms other state-of-the-art methods and has great potential in the diagnosis of stroke.