Abstract: The automatic segmentation of brain white matter lesions has an important auxiliary role in the clinical diagnosis and research of brain diseases. At present, researchers mainly use deep learning method to solve the problem of automatic segmentation of white matter lesions. Although some achievements have been achieved, there are still problems of low segmentation accuracy and small lesions can’t be segmented precisely. In this paper, a fully convoluted 3D deep neural network model is proposed, which integrates residual, pyramid pooling and attention mechanism. In this model, the residual net is used to avoid the gradient disappearance; pyramid pooling is used to aggregate more context information; attention mechanism is used to locate the reign of interest. All modules are connected in order to build a convolutional module chain with strong learning ability, and the up and down sampling are attached at both ends of the chain to form a complete end-to-end deep neural network model. The experiment is carried out on the MICCAI 2017 data set. Experimental results show that compared with other methods, the DSC score of this paper is 0.762, the recall rate is 0.727, the accuracy rate is 0.801, the specificity is 0.991, and the segmentation results are better than those mentioned in other literatures.

Key words: segmentation, white matter hyperintensities, pyramidal pooling, attention mechanism