Related Experiment Video
Updated: Oct 29, 2025
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
48.7K
Lesion probability mapping in MS patients using a regression network on MR fingerprinting
Ingo Hermann1,2, Alena K Golla3,4, Eloy Martínez-Heras5
1Computer Assisted Clinical Medicine, Medical Faculty Mannheim, Heidelberg University, Mannheim, Germany. ingo.hermann@medma.uni-heidelberg.de.
BMC Medical Imaging
|July 9, 2021
Summary
This study introduces a deep learning model for Magnetic Resonance Fingerprinting to accurately map white matter lesions in multiple sclerosis patients. The AI tool significantly speeds up lesion analysis, offering potential clinical benefits.
Area of Science:
- Neuroimaging
- Artificial Intelligence in Medicine
- Medical Image Analysis
Background:
- Magnetic Resonance Fingerprinting (MRF) using echo-planar imaging (EPI) is an advanced MRI technique.
- Accurate segmentation and mapping of white matter (WM) lesions, normal appearing white matter (NAWM), and gray matter (GM) are crucial for multiple sclerosis (MS) patient management.
- Current methods for lesion probability mapping can be time-consuming and complex.
Purpose of the Study:
- To develop and validate a regression neural network for reconstructing probability maps of WM lesions, NAWM, and GM using MRF-EPI.
- To assess the network's ability to simultaneously provide denoised and distortion-corrected quantitative maps.
- To evaluate the clinical utility of deep learning (DL) for rapid WM lesion analysis in MS.
Main Methods:
- MRF-EPI data were acquired from 42 MS patients and 6 healthy volunteers across two imaging sites.
- A U-net architecture was employed to train a deep learning model.
- The U-net was trained for denoising, distortion correction, and generating probability maps for WM lesions, NAWM, and GM.
Main Results:
- The developed neural network achieved a high Dice coefficient of [Formula: see text] for WM lesion prediction with a lesion detection rate of [Formula: see text] at a 33% threshold.
- Accurate quantitative maps for [Formula: see text] and [Formula: see text] were generated with relative deviations of 5.2% and 5.1%, respectively.
- The model demonstrated strong performance for NAWM and GM segmentation, with average Dice coefficients of [Formula: see text] and [Formula: see text] after 80% thresholding.
Conclusions:
- Deep learning presents a powerful and efficient approach for predicting WM lesion probability maps.
- The developed MRF-EPI based DL model can significantly reduce the time required for lesion analysis.
- This technology holds considerable clinical promise for improving WM lesion assessment in patients with multiple sclerosis.
