High-Fidelity CT Image Denoising with De-TransGAN: A Transformer-Augmented GAN Framework with Attention Mechanisms

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Processing

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but introduces noise and artifacts, degrading image quality.
• Existing denoising methods often struggle to balance noise reduction with the preservation of fine anatomical structures.
• Diagnostic confidence in LDCT imaging is compromised by image quality issues stemming from dose reduction.

Purpose of the Study:

• To introduce De-TransGAN, a transformer-augmented Generative Adversarial Network for high-fidelity LDCT image denoising.
• To enhance image quality in LDCT scans, thereby improving diagnostic confidence.
• To develop a clinically viable solution for radiation dose reduction in CT imaging without sacrificing diagnostic utility.

Main Methods:

• De-TransGAN integrates convolutional layers with transformer blocks to capture both local textures and long-range dependencies.
• Channel-spatial attention modules (CBAM) are embedded to focus on diagnostically critical structures.
• A hybrid discriminator (PatchGAN and ViT) and WGAN-GP training enhance stability and realism, using L1, SSIM, and VGG perceptual losses.

Main Results:

• De-TransGAN significantly outperformed state-of-the-art denoising models on multiple benchmark datasets.
• Achieved high quantitative metrics on LDCT head images (PSNR: 44.92 dB, SSIM: 0.9801).
• Qualitative assessments on a private clinical dataset demonstrated excellent generalization, noise suppression, and preservation of fine anatomical details.

Conclusions:

• De-TransGAN offers a robust and effective solution for LDCT image denoising.
• The proposed method enables significant radiation dose reduction in CT imaging.
• De-TransGAN represents a clinically viable approach to maintain diagnostic quality in low-dose CT scans.