Decoding pulsatile patterns of cerebrospinal fluid dynamics through enhancing interpretability in machine learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Biomedical Engineering

Background:

• Cerebrospinal Fluid (CSF) analysis is crucial for disease diagnosis.
• Phase-contrast magnetic resonance imaging (PC-MRI) captures CSF flow dynamics.
• Automatic CSF segmentation in PC-MRI is challenging, and explained ML methods are unexplored.

Purpose of the Study:

• To develop lightweight machine learning (ML) algorithms for CSF lumen segmentation in spinal PC-MRI.
• To utilize velocity-encoded images as features for ML models.
• To enhance the interpretability of ML models using SHAP for CSF dynamics.

Main Methods:

• Utilized a dataset of 57 PC-MRI slabs from 3T MRI scanner (control and idiopathic scoliosis participants).
• Trained ML models on 2176 time series images, interpolating frames for consistent feature size.
• Employed fivefold cross-validation and SHAP for model interpretability.

Main Results:

• Achieved high accuracy with an eXtreme Gradient Boosting (XGB) model: 0.99% precision, 0.95% recall, 0.97% F1 score.
• SHAP analysis provided insights into the contribution of pulsatile features for CSF lumen pixel classification.
• Demonstrated the effectiveness of ML models in feature extraction and selection from PC-MRI pulsatile data.

Conclusions:

• Developed novel ML models for CSF segmentation from PC-MRI pulsatile data.
• The explained ML approach offers valuable insights for domain experts in understanding CSF dynamics.
• This work advances automated analysis and interpretation of CSF flow in medical imaging.