Real-time, inline quantitative MRI enabled by scanner-integrated machine learning: a proof of principle with NODDI

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroimaging

Background:

• Advanced quantitative MRI (qMRI) techniques are often limited to research settings due to complex, offline parameter estimation.
• This hinders the clinical translation and adoption of powerful qMRI methods.

Purpose of the Study:

• To develop and integrate a real-time, inline parameter estimation system for qMRI using neural networks (NNs).
• The goal is to enable "clinical mode" qMRI, facilitating broader clinical use.

Main Methods:

• Customized Siemens Image Calculation Environment (ICE) pipeline to deploy trained NNs via ONNX Runtime.
• Trained two fully-connected NNs offline using the neurite orientation dispersion and density imaging (NODDI) model.
• Demonstrated inline estimation in vivo and evaluated performance with synthetic data.

Main Results:

• Successfully integrated NNs into ICE for inline, whole-brain NODDI parameter estimation in under 10 seconds.
• NNMLE estimates showed consistency with conventional methods and comparable accuracy with better noise robustness.
• NNGT offered higher noise robustness but with compromised accuracy.

Conclusions:

• Real-time inline parameter estimation using this framework overcomes a major barrier to clinical qMRI adoption.
• The generalizable approach allows efficient integration of advanced qMRI techniques into clinical workflows.