Channel Attention HalfU-Net for Skin Lesion Segmentation

U-shaped architecture have gain popularity in the image segmentation task, especially for biomedical image. Regardless of its good performance, U-shaped architecture generally has huge number of parameters. Considering the resource constraint in actual medical environment, the needs lightweight model is essential. In order to overcome this problem, we propose Channel Attention HalfU-Net. The proposed method implement the existing medical image segmentation model named HalfU-Net. HalfU-Net can reduce the computational cost by simplify the decoder part of the UNet. Channel attention is assigned between ghost model for every stage in encoder part to capture more relevant feature. This leads to performance improvement in skin lesion segmentation problem. Despite that our model achieve state-ofthe-arts performance by surpassing U-Net in aspects of DSC and IoU metrics by 0.86% and 3.48% respectively, the number of parameter of our proposed method is more than 200 times less then U-Net.