Â鶹´«Ã½Ó³»­

Â鶹´«Ã½Ó³»­

Explore the latest content from across our publications
Join The Â鶹´«Ã½Ó³»­

Log In

Forgot Password?
Create New Account

Loading... please wait

Abstract Details

Automated Muscle Fat Fraction Quantification through MRI and Deep Learning to Track Axonal Loss in Peripheral Nerve Diseases
Neuromuscular and Clinical Neurophysiology (EMG)
Neuromuscular and Clinical Neurophysiology (EMG) Posters (7:00 AM-5:00 PM)
090
To automate segmentation on muscle MRI images through deep learning to quantify muscle fat fraction in patients with peripheral nerve diseases.
Intramuscular fat accumulation is a common feature in denervated muscles. Fat fraction (FF) can be calculated from MRI images. FF has been shown to be a promising monitoring biomarker in Charcot-Marie-Tooth type 1A. We propose to measure FF in individual muscles in patients with peripheral neuropathies, thereby track partially denervated muscles which may further increase the responsiveness and allow each patient to serve as his/her own control. However, manual segmentation of individual muscles is very laborious with human-made errors. In this study, we develop a deep learning-based model to automatically quantify FF for individual muscles.
MRI data from 24 patients with peripheral neuropathies and 19 healthy controls were manually segmented for all individual muscles, leading to 19,140 contoured muscle boundaries on 1,200 thigh muscle and 660 calf muscle images. Patients were randomly divided into two groups for either training (nthigh=23; ncalf=10) or testing (nthigh=17; ncalf=12). A deep learning-based 3D U-Net model was developed to segment individual muscles. The segmented individual muscles were also grouped to different muscle compartments. Dice coefficient (DC), Bland-Altman and Pearson Correlation analyses were performed to evaluate the similarity between the manual and automated segmentations.
The DC values varied from 0.83±0.17 to 0.98±0.02 (0.63±0.18 to 0.96±0.02) for the testing thigh (calf) muscle data. The 95% CI of Bland-Altman analysis and Pearson coefficient between manual-FF and automatic-FF were [0.49%, -0.56%] and r2=0.989 ([0.84%, -0.71%] and r2=0.971) for the testing thigh (calf) muscle data.
The automated method achieved excellent agreement with those from manual segmentations. There were very few outliers that can be readily corrected in future refining. On-going effect is to leverage the model to track partially denervated muscles longitudinally in patients with peripheral nerve diseases.
Authors/Disclosures
Yongsheng Chen, PhD (Wayne State University)
PRESENTER 		Dr. Chen has nothing to disclose.
Daniel Moiseev Daniel Moiseev has nothing to disclose.
Wan Yee Kong, MBBS (DMC) Dr. Kong has nothing to disclose.
No disclosure on file
Jun Li, MD, PhD, FÂ鶹´«Ã½Ó³»­ (Harris Methodist Hospital) The institution of Dr. Li has received personal compensation in the range of $500-$4,999 for serving as a Consultant for FDA. The institution of Dr. Li has received research support from NIH. Dr. Li has a non-compensated relationship as a Associate Editor of ACTN journal with ANA that is relevant to Â鶹´«Ã½Ó³»­ interests or activities.