Automated segmentation of chronic stroke lesion using efficient U-Net architecture

Stroke is the most common neurological condition worldwide and causes various sequelae, such as motor impairment, cognitive deficit, and language problems. Typically, a radiologist segments the lesion directly. This is time-consuming and relies on the subjective perception of a specialist, so studies that automatically segment stroke lesions are very important in this regard. Although recent medical image segmentation studies have achieved great results using deep learning techniques, there are still many trainable parameters, and long-term dependence problems still exist. To solve these problems, we propose an efficient U-Net (eUNet) for segmenting chronic stroke lesions. The proposed e-UNet incorporates a depthwise convolution-based e-block designed to efficiently reduce the trainable parameters. A global-feature attention block (GA-block) improves segmentation performance by capturing global features between the encoder and decoder. The proposed e-UNet reduces the number of trainable parameters by 3.75 times compared to U-Net.We used the Anatomical Tracings of Lesions After Stroke (ATLAS) dataset to evaluate e-UNet. The lesions segmentation performance of the proposed e-UNet achieved 59.2%, 45.5%, 77.7%, 52.3%, and 52.3% in Dice, IoU, precision, and recall, respectively, for the test dataset (8694 2D images).