Deep Learning-Based Denoising for High-Resolution Carotid Vessel Wall MRI Using Standard Neurovascular Coils

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• High-resolution carotid vessel wall MRI is crucial for assessing atherosclerosis.
• Standard clinical coils often yield lower signal-to-noise ratio (SNR) images, limiting diagnostic quality.
• Specialized coils can improve SNR but increase costs and complexity.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) denoising method for enhancing carotid vessel wall MRI.
• To improve image quality acquired with standard head-and-neck coils.
• To enable high-resolution, multi-contrast imaging without specialized hardware.

Main Methods:

• A supervised DL model using a residual UNet architecture was trained on 55 carotid MRI scans.
• The model mapped low-SNR input images (standard coil) to high-SNR reference images (NSS coil).
• Performance was benchmarked against conventional methods using quantitative (PSNR, SSIM, SNR, CNR) and qualitative metrics.

Main Results:

• The DL denoiser substantially reduced noise while preserving vessel wall structures.
• Achieved high PSNR (>31 dB) and SSIM (>0.93) compared to reference scans.
• Significantly improved SNR and CNR in vessel wall and lumen regions, approaching reference quality.
• Maintained inner-wall edge sharpness, supporting confident plaque assessment.
• Radiologist ratings confirmed image quality improvements.

Conclusions:

• A DL-based denoising method effectively enhances high-resolution carotid vessel wall MRI quality.
• The DL method achieves image quality comparable to specialized neck surface coils.
• This approach may broaden access to advanced carotid imaging by eliminating the need for additional hardware.