GAN-based disentanglement learning for chest X-ray rib suppression

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Rib-suppressed chest X-rays (CXR) improve pulmonary disease diagnosis reliability.
• Existing rib suppression methods struggle with detail preservation and residual rib removal.

Purpose of the Study:

• To develop an advanced Generative Adversarial Network (GAN)-based framework, Rib Suppression GAN (RSGAN), for effective rib suppression in CXR.
• To utilize anatomical knowledge from unpaired computed tomography (CT) images for enhanced rib suppression.

Main Methods:

• Proposed a GAN-based disentanglement learning framework (RSGAN).
• Employed a residual map to quantify intensity differences between original and suppressed CXR.
• Disentangled CXR into structure and contrast features, transferring rib structural priors from CT-derived digitally reconstructed radiographs (DRRs).
• Incorporated adaptive loss for improved rib residue suppression and detail preservation.

Main Results:

• RSGAN demonstrated superior image quality compared to existing state-of-the-art rib suppression techniques.
• Experiments were conducted on 1673 CT volumes and over 120,000 images from four CXR datasets.
• The integration of RSGAN-generated rib-suppressed CXR with original CXR improved lung disease classification and tuberculosis area detection performance.

Conclusions:

• RSGAN effectively suppresses ribs in CXR while preserving crucial details.
• The proposed method offers a significant advancement in medical image processing for improved diagnostic accuracy in pulmonary conditions.