LCRE-Net: A Lightweight Cross-Scale Residual Enhancement Network for Lung Segmentation in CT Images.
Authors
Affiliations (6)
Affiliations (6)
- School of Medicine, Sichuan University of Science and Engineering, Zigong, 643000, People's Republic of China. [email protected].
- Intelligent Perception and Control Key Laboratory of Sichuan Province, Sichuan University of Science and Engineering, Sichuan Province, No. 519 Huixing Road, Zigong, 643000, People's Republic of China. [email protected].
- School of Medicine, Sichuan University of Science and Engineering, Zigong, 643000, People's Republic of China.
- Intelligent Perception and Control Key Laboratory of Sichuan Province, Sichuan University of Science and Engineering, Sichuan Province, No. 519 Huixing Road, Zigong, 643000, People's Republic of China.
- Department of Dermatology, Zigong Fourth People's Hospital, Sichuan University of Science and Engineering School of Medicine, Zigong, 643000, People's Republic of China.
- School of Automation and Information Engineering, Sichuan University of Science and Engineering, Zigong, 643000, People's Republic of China.
Abstract
Accurate segmentation of lung and lesion regions from CT images is crucial for the diagnosis and quantitative assessment of lung diseases. Existing methods for lung CT segmentation suffer from limitations such as insufficient effective receptive field, limited cross-scale feature interaction, unstable boundary delineation, and high complexity, restricting their practical application. To address these challenges, we propose a lightweight cross-scale residual enhancement network (LCRE-Net) designed to segment lung and lesion regions from lung CT images. LCRE-Net adopts a pre-trained Pyramid Vision Transformer v2 as the encoder backbone. To alleviate the information dilution problem caused by small lesions, we embed zero-initialized residual paths at deep pyramid stages of the encoder to enhance the stability of feature representation. Simultaneously, we propose a novel module called the cross-scale attention pyramid module, which adaptively fuses high-level semantic features with mid-level spatial details through learnable weights. Furthermore, we construct a lightweight feature enhancement path consisting of efficient receptive field blocks and edge enhancers, effectively suppressing artifact boundary interference while enhancing multi-scale context modeling capabilities. Experimental results show that LCRE-Net achieves competitive segmentation performance on three publicly available lung CT datasets, achieving average Dice similarity coefficients of 0.9845, 0.8544, and 0.8611, respectively. The proposed LCRE-Net maintains low model complexity and stable performance across datasets.