CXR-Seg: A Novel Deep Learning Network for Lung Segmentation from Chest X-Ray Images.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning, especially neural networks, is crucial for medical image analysis, enhancing diagnostic accuracy.
• Accurate segmentation and classification of thoracic organs in chest X-rays remain challenging.
• Existing methods require improvement for reliable lung segmentation and abnormality detection.

Purpose of the Study:

• To introduce CXR-Seg, a novel deep learning architecture for semantic lung segmentation in chest X-ray images.
• To address the limitations of current methods in achieving precise segmentation and enhancing diagnostic reliability.
• To improve the accuracy and generalization of lung segmentation in thoracic imaging.

Main Methods:

• Developed CXR-Seg, a network featuring an EfficientNet encoder, spatial enhancement module, transformer attention module, and multi-scale feature fusion.
• Employed a pre-trained EfficientNet for feature extraction and integrated attention mechanisms for enhanced feature fusion.
• Evaluated performance on four public datasets: MC, Darwin, Shenzhen (chest X-rays), and TCIA (brain MRI).

Main Results:

• CXR-Seg achieved high performance metrics, including Jaccard index (up to 95.63%) and Dice coefficient (up to 97.76%) on chest X-ray datasets.
• Demonstrated superior accuracy, sensitivity, and specificity across multiple public datasets.
• Outperformed existing state-of-the-art methods in lung segmentation tasks.

Conclusions:

• The proposed CXR-Seg network significantly improves semantic segmentation of lungs from chest X-ray images.
• The architecture demonstrates enhanced performance and generalization capabilities compared to current methods.
• CXR-Seg offers a promising advancement for reliable thoracic image analysis and diagnosis.