Efficient Lung Segmentation from Chest Radiographs using Transfer Learning and Lightweight Deep Architecture.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate lung segmentation is crucial for chest radiograph (CXR) analysis.
• Challenges include anatomical variability and image quality differences.
• Existing methods often require complex models.

Purpose of the Study:

• To propose DenseCX, an automated lung segmentation methodology.
• To leverage U-Net architectures and transfer learning for enhanced feature extraction.
• To reduce overfitting and improve generalization in lung segmentation.

Main Methods:

• DenseCX utilizes a U-Net architecture with a Dense block-based encoder.
• Transfer learning and homogeneous domain adaptation are employed.
• Evaluation performed on three public CXR datasets using a leave-one-group-out strategy.

Main Results:

• DenseCX demonstrates superior performance over conventional ImageNet initialization.
• Achieves a favorable balance between performance and model complexity.
• Exhibits improved convergence and a lighter architecture.

Conclusions:

• DenseCX offers an effective and efficient solution for automated lung segmentation in CXRs.
• The methodology shows strong generalization capabilities across diverse datasets.
• DenseCX represents a significant advancement in medical image analysis for diagnosis.