A Generic Approach to Lung Field Segmentation From Chest Radiographs Using Deep Space and Shape Learning

Area of Science:

• Medical imaging analysis
• Computer-aided diagnosis
• Pediatric radiology

Background:

• Existing computer-aided diagnosis (CAD) for lung field segmentation primarily targets adults, with limited application to pediatric chest radiographs (CXR).
• Conventional statistical shape models (SSMs) struggle to adapt to the significant lung shape variations during pediatric development.
• Accurate lung field segmentation is crucial for quantitative analysis and disease detection in medical imaging.

Purpose of the Study:

• To develop a generic lung field segmentation framework for chest radiographs (CXR) applicable to both adult and pediatric cohorts.
• To address the limitations of traditional methods in accommodating diverse lung shapes, particularly during pediatric development.
• To enhance lung capacity estimation by including the retro-cardiac region in segmentation.

Main Methods:

• A novel deep representation learning framework utilizing ensemble space learning for robust object localization.
• Implementation of marginal shape deep learning for efficient estimation of shape deformation parameters.
• A generic segmentation approach designed to handle significant shape variations in lung fields across a wide age range.

Main Results:

• Achieved a high mean Dice similarity coefficient of 0.96 ±0.03 on 668 CXRs, encompassing the retro-cardiac region.
• Demonstrated superior performance in segmenting lung fields for both pediatric and adult subjects compared to conventional methods.
• The proposed method is computationally faster than traditional SSM-based iterative segmentation techniques for comparable accuracy.

Conclusions:

• The developed framework provides accurate and efficient lung field segmentation for chest radiographs across a broad age spectrum, including pediatric patients.
• The novel deep learning approach effectively handles substantial lung shape variations and integrates the retro-cardiac region for improved lung capacity estimation.
• The computational simplicity and generic nature of the framework suggest potential applications in segmenting other deformable anatomical structures.