U-Net-Based Prediction of Cerebrospinal Fluid Distribution and Ventricular Reflux Grading

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Cerebrospinal fluid (CSF) is vital for brain waste clearance.
• Altered CSF flow patterns are linked to central nervous system diseases.
• Intrathecal contrast agent distribution aids in studying CSF dynamics.

Purpose of the Study:

• To explore deep learning for predicting intrathecal gadolinium-based contrast agent distribution in the human brain.
• To assess the efficacy of a U-net-based model for forecasting CSF tracer flow patterns.
• To determine if early post-injection MRI data is sufficient for accurate predictions.

Main Methods:

• Utilized T1-weighted MRI scans at multiple time points before and after intrathecal contrast agent injection.
• Developed a U-net-based supervised learning model to predict pixel-wise signal increase at 24 hours post-injection.
• Evaluated model performance using various tracer distribution stages and baseline scans.

Main Results:

• Deep learning models trained on early (2-hour post-injection) MRI data achieved prediction accuracy comparable to models using later-stage scans.
• Model predictions aligned well with expert neuroradiologist gradings of ventricular reflux.
• The U-net model demonstrated robust performance in predicting CSF tracer distribution.

Conclusions:

• Deep learning offers a promising, efficient method for predicting CSF flow dynamics.
• Minimizing MRI scan duration using deep learning can enhance clinical workflow efficiency.
• This approach has the potential to improve patient well-being and reduce healthcare expenditures.