Validation of deep learning techniques for quality augmentation in diffusion MRI for clinical studies

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Diffusion MRI (dMRI) is crucial for white matter analysis.
• Reduced gradient directions in dMRI can impact clinical findings, as seen in migraine studies.
• Deep learning (DL) and artificial intelligence (AI) offer potential for enhancing dMRI data quality.

Purpose of the Study:

• To evaluate DL techniques for improving dMRI quality in clinical applications.
• To assess if AI in medical imaging leads to loss or introduction of clinical information.
• To investigate the impact of reduced angular resolution in dMRI on migraine patient classification.

Main Methods:

• Fourteen teams used DL to enhance dMRI metrics (FA, AD, MD) from data with 21 gradient directions.
• The goal was to achieve results comparable to data with 61 gradient directions.
• Evaluations used image quality metrics and Tract-Based Spatial Statistics (TBSS) for migraine patient comparison.

Main Results:

• Most DL techniques improved detection of statistical differences between episodic and chronic migraine groups.
• DL methods increased false positives, with a linear relationship between new true and false positives.
• Performance varied, with some methods showing bias and failing to replicate original data distribution.

Conclusions:

• DL can enhance dMRI data but carries a risk of introducing false positives.
• Generalizing AI tasks across diverse clinical cohorts requires caution due to potential data alteration.
• Extreme care is needed when using AI for harmonization or synthesis in clinical studies with heterogeneous data.