Enhancing the Generalizability of Deep Learning-Based Models for Lung Field Segmentation in Chest Radiographs Using...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Lung field segmentation (LFS) in chest X-rays (CXR) is crucial for diagnosing respiratory diseases.
• Deep learning models struggle with precise LFS due to poor image contrast and overlapping structures, leading to poor generalizability.

Purpose of the Study:

• To develop a robust and generalizable two-phase hybrid approach for accurate LFS in CXR.
• To overcome limitations of existing deep learning models in handling image variations and complex anatomical structures.

Main Methods:

• A two-phase hybrid method combining heuristic edge detection (Laplacian of Gaussian and Canny filters) with deep learning (U-Net architectures).
• Phase 1: Extracting and refining multiscale edge features using LoG and Canny filters.
• Phase 2: Fusing enriched feature maps with original contrast inputs to train U-Net models.

Main Results:

• The edge-assisted approach significantly improved LFS performance across three benchmark datasets (MC, SH, JSRT).
• Deep attention U-Net achieved a Dice coefficient of 0.9815 and Jaccard score of 0.9624 on the JSRT dataset.
• The method demonstrated improved Dice and IoU gains compared to baseline models and showed strong cross-dataset validation results.

Conclusions:

• The proposed hybrid method offers a reliable and effective solution for LFS in diverse and challenging clinical settings.
• Enhanced LFS accuracy and generalizability contribute to improved computer-assisted diagnosis and patient care for respiratory conditions.