Validation of a Denoising Method Using Deep Learning-Based Reconstruction to Quantify Multiple Sclerosis Lesion Load
T Yamamoto1, C Lacheret2, H Fukutomi1
1From the Institut de Bio-imagerie (T.Y., H.F., L.D., V.D., T.T.), University Bordeaux, Bordeaux, France.
AJNR. American Journal of Neuroradiology
|July 28, 2022
Summary
Deep learning denoising enhances accelerated 3D FLAIR MRI scans for multiple sclerosis (MS) lesion quantification. This strategy improves image quality and accuracy, potentially shortening MRI examination times for MS patients.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Neuroscience
Background:
- Accurate quantification of white matter (WM) lesion load is critical for managing multiple sclerosis (MS).
- Accelerated imaging techniques are desirable to reduce patient burden and increase throughput.
- Deep learning-based reconstruction offers potential for image quality enhancement.
Purpose of the Study:
- To evaluate the efficacy of combining accelerated 3D FLAIR MRI with deep learning-based denoising for MS lesion quantification.
- To determine if this combined strategy can maintain diagnostic accuracy while reducing scan time.
Main Methods:
- Prospective examination of 28 MS patients using four accelerated 3D FLAIR sequences with varying scan times.
- Reconstruction of each sequence with and without deep learning-based denoising.
- Assessment of image quality using Likert scale, SNR, and CNR.
- Quantitative evaluation of manual and automatic lesion segmentation against ground truth using multiple metrics (e.g., Dice, Hausdorff).
Main Results:
- Accelerated FLAIR sequences showed image quality deterioration, but deep learning denoising significantly improved subjective and quantitative metrics.
- Denoising improved lesion detection, recovering contours and identifying lesions missed in standard accelerated scans.
- The 2-minute 35-second FLAIR with denoising achieved comparable lesion quantification to the reference 4-minute 54-second scan.
Conclusions:
- Deep learning-based denoising is a valuable strategy for recognizing MS lesions in accelerated FLAIR acquisitions.
- This approach can effectively shorten MRI scan times in clinical practice without compromising essential diagnostic information.
- The combination offers a promising method for efficient and accurate MS lesion load assessment.
