A Lightweight Lung Region Image Segmentation Model with Attention Mechanisms

Abstract

Image segmentation of lung regions is helpful for the diagnosis of lung diseases. The existing lung segmentation networks with high segmentation accuracy face difficulty in leveraging accuracy and speed in practical clinical application platforms due to high computational loads. To address this issue, we propose a lightweight lung segmentation network based on U-Net, which consists of a residual depth-separable module, an attention module, and a multi-receptive field feature fusion module. Depthwise separable convolutions are used to achieve lightweight. To prevent a drop in accuracy, we add a scSE attention module to the encoder to help the model effectively highlight the target area during feature extraction and pay more attention to the foreground pixels. In addition, a lightweight multi-receptive field feature fusion module is designed to alleviate the loss of spatial information caused by pooling and better adapt to the multi-size features of the lung region. The proposed network is evaluated on the Luna16 and the NSCLC-Radiomics datasets. Compared with the standard U-Net model, the proposed model maintains the original accuracy and reduces the number of parameters by 69.3%.