Deep Generative Translation of Standard Images into Virtual High-Energy Images for Facilitating Dual-Energy Chest...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Dual-energy subtraction (DES) radiography improves visualization of soft tissues and bones.
• Specialized dual-energy imaging systems are typically required for DES.
• Existing deep learning models can generate virtual DES from high-energy (HE) to low-energy (LE) images (HE2LE).

Purpose of the Study:

• To develop a method for generating virtual DES from standard (STD) radiography images.
• To enable virtual DES without specialized dual-energy hardware.
• To enhance the diagnostic utility of conventional radiographic imaging.

Main Methods:

• A U-Net-based translation model was developed to convert STD images into virtual HE images.
• The model utilized a discriminator within the pix2pix framework.
• A dataset of 600 triplet chest radiographs (STD, HE, LE) was used for training with sixfold cross-validation over 2000 epochs.

Main Results:

• The STD2HE model generated virtual HE images with high fidelity (PSNR: 34.5, SSIM: 0.979, DISTS: 0.0257).
• Virtual LE images derived from virtual HE inputs also showed high similarity to ground truth (PSNR: 35.1, SSIM: 0.963, DISTS: 0.0455).
• Generated virtual DES images demonstrated high structural and perceptual fidelity (FID: 67.6 for BS, 73.6 for BE).

Conclusions:

• The proposed framework successfully generates virtual DES images from standard radiography.
• This approach has the potential to enhance diagnostic accuracy in radiological analysis.
• The integration of this method with existing HE2LE models offers a pathway to improved imaging without specialized hardware.