Subarachnoid Hemorrhage Image Segmentation Based On Deformable Attention Mechanism

Subarachnoid hemorrhage(SAH)is an acute neurological disease,usually caused by blood flowing into the subarachnoid space due to the rupture of intracranial aneurysm,and it belongs to the category of intracranial hemorrhage(ICH).Most SAH lesions present the complex data characteristics of diffuseness,irregularity and multi-scale,which are quite different from other ICH lesions.At present,the segmentation algorithms applied to ICH images are usually full convolution networks(FCN)or hybrid models composed of convolution module and attention module.The fully convolutional network,limited by the fixed receptive field of the convolution kernel,can’t accurately model the diffuseness of the lesions in the subarachnoid hemorrhage image.While in the hybrid model,the various attention modules used don’t calculate the correlation among pixels reasonably,and they are easy to introduce some irrelevant information in the attention calculation process,which reduces the quality of the extracted data features,resulting in poor segmentation performance.In view of the above problems,to realize the effective expression of complex data characteristics existed in SAH lesions,this paper focuses on the methods of SAH image segmentation based on deformable attention mechanism.The main contributions of this paper are as follows:(1)Propose deformable attention u-shaped network with progressively supervised learning for subarachnoid hemorrhage image segmentation.In this paper,we specifically design a deformable attention u-shaped network(DAUN).First,a deformable attention module(DAM)is embedded at the end of each encoding layer of the Res-UNet to adaptively adjust the attention domain,thus reducing the introduction of irrelevant information.Next,in order to improve the segmentation accuracy of the model for irregular edges and tiny lesions,a region-boundary-aware loss function(RBL)is used to optimize internal parameters of DAUN during training.Finally,a progressive supervised learning strategy(PSL)is designed to train the proposed DAUN model,which enables DAUN to pay balanced attention to the semantic information and location information represented by each pixel.Later,comparative experiments were carried out on our self-made subarachnoid hemorrhage image dataset(SAH-CT)and a public cell nucleus segmentation image dataset(Mo Nu Seg).Experimental results show that our DAUN outperforms other advanced methods in the segmentation of subarachnoid hemorrhage images,proving the effectiveness of the proposed method.(2)Propose u-shaped deformable transformer for subarachnoid hemorrhage image segmentation.In order to further accurately model the data characteristics of SAH lesions,a u-shaped deformable transformer(UDT)is proposed subsequently.First,we designed a multi-scale deformable attention module(MSDA)to jointly express the SAH lesions’ data characteristics of diffuseness,irregularity and multi-scale.The MSDA module is able to fuse features of different scales and dynamically adjust the attention domain of each pixel according to the input image to generate discriminative multi-scale features.Second,we propose a cross deformable attention-based skip connection structure(CDASC)to accurately segment the large amount of irregular edges presented in SAH lesions.The CDASC structure can exploit the spatial details from the encoded features to refine the spatial information represented by the decoded features.According to the comparative experimental results on SAH-CT dataset,Gla S dataset and Mo Nu Seg dataset,the rationality of the proposed method is verified.