Improving CXR Bone Suppression by Exploiting Domain-Level and Instance-Level Information.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Effective bone suppression in chest X-rays (CXRs) is crucial for detecting lung abnormalities and improving clinical diagnoses.
• Existing deep generative models for bone suppression in CXRs often lack significant improvements in downstream diagnostic tasks due to a focus on instance-specific details over domain knowledge.

Purpose of the Study:

• To develop a novel framework that integrates both instance-level and domain-level information for enhanced bone suppression in CXRs.
• To improve the effectiveness of bone suppression techniques for better lung abnormality detection and clinical diagnosis.

Main Methods:

• A hybrid approach combining cross-covariance attention blocks (CABs) and Vision Transformers (ViTs) for instance-level information capture.
• Introduction of multi-head codebook attention (MCA) to capture global, domain-level information specific to bone-suppressed CXRs.
• A two-stage training scheme involving MCA learning and a synthesis stage using a ViT-based Generative Adversarial Network (GAN).

Main Results:

• Significant improvements in bone suppression quality, with PSNR increased by 8.36% and SSIM by 2.7%.
• Enhanced performance in downstream diagnostic tasks, including lung disease classification (2.8% and 4.2% improvement) and segmentation (1.5% improvement).

Conclusions:

• The proposed framework effectively integrates instance and domain knowledge for superior bone suppression in CXRs.
• The approach demonstrates substantial benefits for clinical diagnosis by improving both image quality and diagnostic task performance.