Segmentation of medical images using an attention embedded lightweight network

Abstract

Accurate segmentation of computerized tomography (CT) images is of great significance to clinical diagnosis. However, because of the high similarity of gray values, it is a challenging task for CT image segmentation. The encoder and decoder based CNN architecture has greatly improved the segmentation effect, but it also encounters a bottleneck due to the information loss in the encoding process. In view of this, we proposed an image segmentation model based on a novel network architecture for medical image segmentation. To improve the efficiency and decrease the number of model parameters, we optimized the Inception module by substituting the depth-wise separable convolutions (DWSC) for the standard convolutions. Then, the optimized Inception module paired with the residual network was chosen as the backbone extractor to extract high-quality image features. Further, a hybrid attention mechanism, which consists of channel-wise and spatial attention, was incorporated into the network to realize the maximum reuse of inter-channel relationships and spatial point characteristics. In particular, the attention module was separately embedded into the contracting and expansive paths to enhance the feature extraction capability and detail restoration effects. The experimental indicators were significantly improved on the test dataset, and the intersection over union (IoU) of the proposed method reached no less than 0.9645, 0.6499, and 0.7945 on the Lung, Colon tumor, and DRIVE datasets, respectively, which demonstrated the effectiveness of the proposed method. Our code and data are available at https://github.com/xtu502/medical-image-segmentation/.