LCRE-Net: A Lightweight Cross-Scale Residual Enhancement Network for Lung Segmentation in CT Images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of lung and lesion regions in CT images is vital for diagnosing and assessing lung diseases.
• Current segmentation methods face challenges like limited receptive fields, poor cross-scale feature interaction, unstable boundaries, and high complexity.

Purpose of the Study:

• To propose a lightweight cross-scale residual enhancement network (LCRE-Net) for improved lung and lesion segmentation in CT images.
• To address limitations of existing methods, enhancing practical applicability in clinical settings.

Main Methods:

• Utilized a pre-trained Pyramid Vision Transformer v2 as the encoder backbone.
• Introduced zero-initialized residual paths to stabilize feature representation, especially for small lesions.
• Developed a cross-scale attention pyramid module for adaptive fusion of semantic and spatial features.
• Incorporated a lightweight feature enhancement path with receptive field blocks and edge enhancers to manage boundary artifacts and improve multi-scale context.

Main Results:

• LCRE-Net achieved high segmentation performance on three public lung CT datasets.
• Reported average Dice similarity coefficients of 0.9845, 0.8544, and 0.8611.
• Demonstrated low model complexity and stable performance across different datasets.

Conclusions:

• LCRE-Net offers a promising solution for accurate and efficient lung and lesion segmentation in CT images.
• The network's design effectively tackles challenges related to feature representation, scale interaction, and boundary delineation.
• LCRE-Net shows potential for practical application in lung disease diagnosis and quantitative assessment.