An attention mechanism and multi-feature fusion network for medical image segmentation

Abstract

Recently, deep learning has been applied to medical image segmentation. However, existing methods based on deep learning still suffer from several disadvantages, such as blurred edge segmentation of image lesion regions and weak context information extraction. To tackle these problems, this paper proposes an attention mechanism and multi-feature fusion network with the encoder-decoder structure for medical image segmentation. In the proposed network, the convolutional group encoder module and the self-attention module are applied to divide images. The convolutional group encoder uses multiple convolution and dilated convolution to enhance the multi-scale information capturing capability of the model. The extracted image features will be useful for precise segmentation. Moreover, the self-attention module is introduced into the network for mining and complementing the edge details of segmented images. In the proposed model, convolutional group encoders and self-attention are applied repeatedly to capture changes in contextual relationships and continuously refine boundary information. Several experiments have been conducted on the BUSI and ISIC datasets to verify the effectiveness of the proposed method. Compared with other methods, the proposed method can achieve better segmentation results.