Objective. Convolutional neural networks have been widely adopted for medical image segmentation with their outstanding feature representation capabilities. As the segmentation accuracy gets constantly updated, the complexity of networks increases as well. Complex networks can achieve better performance but require more parameters and are hard to train with limited resources, while lightweight models are faster but cannot fully utilize the contextual information of medical images. In this paper, we focus on better balancing the efficiency and accuracy.Approach. We propose a correlation-enhanced lightweight network (CeLNet) for medical image segmentation, which adopts a siamese structure for weight sharing and parameter saving. Through the feature reuse and feature stacking of parallel branches, a point-depth convolution parallel block (PDP Block) is proposed to reduce the model parameters and computational cost while improving the feature extraction capability of encoder. A relation module is also designed to extract feature correlations of input slices, which utilizes global and local attention to enhance feature connections, while reducing feature differences through element subtraction, and finally obtains contextual information of associated slices to improve the segmentation performance.Main results. We conduct extensive experiments on the LiTS2017, MM-WHS and ISIC2018 datasets, and the proposed model consumes merely 5.18M parameters but achieves excellent segmentation performance, specifically, a DSC of 0.9233 in LiTS2017 dataset, an average DSC of 0.7895 on MM-WHS dataset and an average DSC of 0.8401 on ISIC2018 dataset.Significance. CeLNet achieves state-of-the-art performance in multiple datasets while ensuring lightweight.
Keywords: contextual learning; lightweight; medical image; multi-slice; segmentation.
© 2023 Institute of Physics and Engineering in Medicine.