Stroke lesion segmentation model based on convolutional fusion and an improved attention mechanism

Ischemic stroke is an acute cerebral blood circulation disorder caused by cerebral artery stenosis or occlusion with high rates of fatality and disability. Magnetic resonance imaging has shown high efficiency in the rapid diagnosis of ischemic stroke. However, the segmentation of stroke lesions is challenging because of their small size and blurred boundaries with normal tissues. Manually segmenting the lesions is subjective, laborious, and time-consuming. Therefore, we propose a segmentation model for stroke lesions that is based on the fusion of convolution and an improved attention mechanism. First, to extract richer context information, we propose a convolutional fusion encoding module that realizes the fusion of two-dimensional convolution and three-dimensional convolution on the contraction path in the encoding end, and enhance features using the convolutional block attention mechanism module. Then, we propose an improved residual-attention gate hybrid decoding module for integrating features with different resolutions focusing on the target region, thus improving the segmentation detail of the edges of small lesions. Lastly, we use a stroke lesion dataset to evaluate its performance in the experiment. The results demonstrate that, compared with the existing models such as TransUNet, SAN-Net, and W-Net, the Dice similarity coefficient of the proposed model exhibits better (0.6193) performance.