Using AI Networks to Improve Low-Field Brain Images

Overview

Low-field Magnetic Resonance Imaging (MRI) promises to make MRI more accessible and affordable. However, this accessibility comes with a tradeoff: lower image quality that can limit clinical usefulness. Artificial intelligence (AI) methods have the potential to improve low-field MRI image quality and facilitate clinical point-of-care decisions.

Researchers in the Information Technology Laboratory's AI Research, Measurement, and Standards Division (AI-RMS) are exploring the use of AI to create high-quality Apparent Diffusion Coefficient (ADC) maps from lower-quality low-field MRI scans. ADC is an MRI parameter used in multiple neurological applications, including stroke diagnosis and oncology. If AI can generate ADC maps that reliably bridge the quality gap between low-field and high-field MRI, it could make high-quality diagnostic neuroimaging far more accessible.

Approach

The research team is exploring the use of AI to create high-quality Apparent Diffusion Coefficient (ADC) maps from 64 mT MRI scans. ADC is a field-strength-invariant quantitative MRI parameter used in multiple neurological applications, including stroke diagnosis and oncology.

The team trains AI models using a database of co-registered MRI brain scans of healthy participants acquired at both 64 mT and 3 T. The AI models are trained to output brain ADC maps from diffusion-weighted images acquired at 64 mT, using the 3 T scans as ground truth.

To assess run-to-run variability and improve robustness, the team trains 25 individual AI models. Rather than relying on a single model's output, the researchers generate an ensemble average across all 25 models. This ensemble approach yields several benefits:

• Improves the range of ADC values compared to the raw low-field maps
• Decreases noise levels in the resulting images
• Returns maps with greater clinical value than the original low-field ADC maps
• Diminishes the appearance of "hallucinations" — artifacts visible in some individual AI model outputs that do not correspond to real anatomical features

The quality of the AI-inferenced ADC maps is evaluated by a working radiologist for both quantitative and qualitative accuracy in comparison to the 3 T ground truth.

Impact

If AI can reliably bridge the quality gap between low-field and high-field MRI, it could make diagnostic neuroimaging far more accessible. Low-field scanners are portable, less expensive, and do not require the shielded rooms that high-field systems demand. AI-enhanced low-field imaging could bring point-of-care brain scanning to emergency departments, rural clinics, and resource-limited settings where high-field MRI is unavailable.

This work also advances NIST's measurement science mission by developing methods to evaluate the fidelity of AI-reconstructed medical images — a challenge that extends well beyond MRI to any domain where AI is used to enhance or reconstruct scientific measurements from limited input data.