An Efficient and Robust Method for Chest X-ray Rib Suppression That Improves Pulmonary Abnormality Diagnosis

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Healthcare

Background:

• Suppression of thoracic bone shadows on chest X-rays (CXRs) is crucial for accurate pulmonary disease diagnosis.
• Existing methods include unsupervised physical models (accurate but slow) and supervised deep learning models (fast but limited by ground truth data).
• There is a need for efficient and robust methods for rib suppression in CXRs.

Purpose of the Study:

• To develop a generalizable and efficient workflow for chest X-ray rib suppression by integrating physical and machine learning techniques.
• To improve the accuracy and speed of bone shadow removal in CXRs.

Main Methods:

• A two-stage pipeline was developed: (1) generation of ground truth (GT) bone-eliminated pairs using a physical model, and (2) training a supervised deep learning network (SADXNet) on these pairs.
• SADXNet, a densely connected U-shaped network, was designed for image denoising and rib suppression.
• A loss function combining peak signal-to-noise ratio (PSNR) and multi-scale structural similarity index measure (MS-SSIM) was employed.

Main Results:

• SADXNet effectively suppressed rib edges near the lung wall and vertebrae without obscuring important features like vessels or abnormalities.
• The model achieved near-zero RMSE compared to physical model GTs, with prediction times under 1 second.
• Downstream tasks showed significant improvements: 3.23% and 6.62% AUC increase for lung nodule detection and disease localization, respectively, with substantial reductions in false positives.

Conclusions:

• The proposed SADXNet, trained on physically generated image pairs, provides robust sub-second rib suppression with high fidelity.
• Quantitative results from downstream tasks validate the superiority of SADXNet and highlight the effectiveness of training ML models with physically generated datasets.
• The study provides the training images and SADXNet model for further research and application.