White matter lesion segmentation by fusing multi-scale features and attention

In order to solve the problems of low segmentation accuracy of white matter leisions in the current magnetic resonance brain imaging and easy missed recognition of small lesions, an improved U-Net model with multi-scale features extraction and attention mechanism was proposed for automatic segmentation of white matter leisions. Firstly, the Multi-Scale Convolution Module (MSCM) is introduced to expand the network width, so that improving the feature extraction ability of the network. Secondly, the Hybrid Down-Sampling Module (HDSM) is introduced to reduce the loss of the information in the down-sampling process through fusing features of coarse-grained and fine-grained down-sampling extraction. Simultaneously, the Cross-Layer Fusion Module (CLFM) is introduced to reduce the semantic difference of the peer layers by fusing encoder and decoder information at the both ends of the skip connection. Finally, seperated attention mode is adopted in the encoder, spatial attention and channel attention are designed respectively according to the characteristics of deep and shallow layers in feature extraction, in order to enhance the network's attention to the lesion area. The experimental results on the MICCAI2017 dataset show that the proposed algorithm can effectively segment white matter lesions, especially small lesions can be accurately segmented. Compared with other literature algorithms of the same task on the public data set provided by MICCAI2017 WMHs segmentation challenge, the proposed algorithm has effectively improved the recall and the DSC, reaching 0.834 and 0.803 respectively, which indicates that the algorithm in this paper is an effective automatic white matter hyperintensity segmentation algorithm.