MDSU-Net: A Multi-attention and Depthwise Separable Convolution Network for Stroke Lesion Segmentation

In recent years, deep learning is popular in medical image segmentation tasks. Due to network model construction issues, the contextual dependencies of large-scale medical images cannot sometimes be captured. We propose a novel network MDSU-Net by incorporating a multi-attention mechanism and a depthwise separable convolution within a U-Net framework. The multi-attention consists of a dual attention and four attention gates, which extracts the contextual information and the long-range feature information from large-scale images. MDSU-Net achieves a Dice coefficient of 0.5587, a precision of 0.6424, and a recall of 0.5276 on the Anatomical Tracings of Lesions After Stroke (ATLAS) dataset. Our experiment results suggest that the proposed MDSU-Net outperforms the state-of-the-art methods including U-Net and D-Unet in Dice coefficient, precision, and recall.