²³ Na MRI in ischemic stroke: Acquisition time reduction using postprocessing with convolutional neural networks.

Area of Science:

• Medical Imaging
• Neuroimaging
• Artificial Intelligence in Medicine

Background:

• Quantitative ²³ Na MRI measures tissue sodium concentration (TSC), crucial for cell viability but hindered by long acquisition times.
• K-space undersampling accelerates MRI acquisition but introduces noise and artifacts, limiting clinical use.
• Convolutional Neural Networks (CNNs) show promise in enhancing medical image quality and reducing artifacts.

Purpose of the Study:

• To reduce ²³ Na MRI acquisition time by 75% through k-space undersampling.
• To utilize CNNs for denoising and artifact reduction in undersampled ²³ Na MRI.
• To evaluate the impact of CNN-based postprocessing on image quality, TSC accuracy, and diagnostic potential.

Main Methods:

• Retrospective analysis of ²³ Na MRI data from 46 ischemic stroke patients (10 min acquisition).
• Simulated 2.5 min acquisition (RI) reconstructed using eight U-Net or ResNet-based CNNs trained to match full 10 min images (FI).
• Evaluation metrics included signal-to-noise ratio (SNR), structural similarity (SSIM), TSC error, and subjective neuroradiologist ratings.

Main Results:

• CNNs significantly improved SNR of undersampled data, with U-Nets achieving an SNR of 43.99 compared to 10.16 for RI.
• Three CNNs improved SSIM to 0.91 ± 0.03, and reduced TSC error by up to 15%.
• Subjective image quality ratings for CNN-processed images were significantly better than for undersampled images alone.

Conclusions:

• CNN postprocessing effectively reduces noise and artifacts in 75% undersampled ²³ Na MRI, enabling significantly faster acquisition.
• This approach maintains or improves image quality and quantitative accuracy (TSC), facilitating clinical translation of rapid ²³ Na MRI.
• Accelerated ²³ Na MRI holds potential for improved assessment of cell viability in neurological conditions like ischemic stroke.