Segmentation Method of Stroke Lesions Based on Bidirectional Recurrent U-Net Model

Abstract

Stroke has a high disability rate and fatality rate. It is of great clinical significance to study the automatic recognition and segmentation of stroke lesions. Convolutional neural network can not make use of time sequence correlation of medical image data, and has the problem of low feature utilization. Therefore, a bidirectional recurrent U-Net (BIRU-Net) model was proposed for segmentation of lesions. First, a recurrent neural network is introduced to replace part of the convolutional layer in U-Net with an improved attention convolutional gate recursive unit (ACGRU), so that the segmentation model is not only suitable for small-scale annotated medical image data sets, but also has the characteristics of long-term memory. Second, a dual-channel fusion training mechanism is adopted to input the forward and reverse slice data of a single view into BIRU-Net at the same time, and realize bidirectional feature fusion in the process of model forward propagation, effectively utilizing the bidirectional dependence of slice sequence. Finally, the segmentation results of each single view are refused to effectively utilize the spatial context information of data. The experimental results of ATLAS data set show that the DSC value of the proposed method reaches 62.58%. Compared with other methods at the present stage, the proposed method can segment the lesion region more accurately.

Keywords

• deep learning
• segmentation of stroke lesions
• cgru
• u-net
• bidirectional feature fusion
• multiplane fusion