Virtual Imaging-Guided Harmonization of Chest X-rays Using a Generative Adversarial Neural Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Processing

Background:

• Chest radiography (CXR) exhibits significant variability due to acquisition and post-processing.
• This variability impacts image quality and downstream diagnostic performance.

Purpose of the Study:

• To reduce variability in CXR acquisition and post-processing.
• To assess if image harmonization improves CXR quality and diagnostic accuracy.

Main Methods:

• A generative adversarial network (GAN) was trained using virtual images from a Virtual Imaging Trial (VIT).
• The GAN maps non-harmonized CXRs to a noise-free reference using a dual U-Net architecture.
• Evaluation involved virtual, phantom, and clinical datasets using various image quality metrics (GIQMs, CIQMs).

Main Results:

• Harmonized CXRs showed significantly improved metrics (lower NRMSE, higher PSNR and SSIM) compared to non-harmonized images.
• Image quality metric variability was substantially reduced on phantom data, outperforming other harmonization methods.
• Clinical data analysis revealed suppressed acquisition variability and improved diagnostic accuracy on harmonized images.

Conclusions:

• VIT-guided GAN training effectively harmonizes CXRs, reducing quantification variability and standardizing image representation.
• The virtual-to-clinical strategy is scalable and generalizable, enabling standardized CXR appearance and reliable detection across institutions.