Article Text

Download PDFPDF

FRI0582 Change in muscle volume and muscle fat fraction as potential non-invasive biomarkers of disease progression: machine learning framework for quantitative analysis of mri data
  1. D. Fischer1,
  2. P. Hafner1,
  3. S. Schmidt1,
  4. M. Hinton2,
  5. J. Gonzalez2,
  6. O. Kubassova2
  1. 1Division of Neuropaediatrics, University of Basel Children’s Hospital, Basel, Switzerland
  2. 2IAG, London, UK


Background Change in muscle volume and muscle fat fraction are potential non-invasive biomarkers of disease progression in a number of diseases, including sporadic inclusion body myositis1 and osteoarthritis2. Their measurement from magnetic resonance images (MRI) usually involves time consuming manual segmentations of the images by trained readers, which limits the use of these biomarkers in clinical research and practise.

Objectives In this study, we present a novel unsupervised k-means-classifier based image processing framework, developed for machine learning approach to segmentation of thigh muscles from MRI and the subsequent calculation of fat fraction from Dixon images. Further, we present validation of the new approach against manual segmentation using longitudinal retrospectively acquired data.

Methods Axial MR images from the upper thighs including in-phase and out of phase sequences were recorded in a group of 8 subjects at baseline and at a follow up. The 16 imaging time points were manually segmented by an expert reader, who delineated the muscle. For these regions the mean fat fraction was calculated from the in-phase and out of phase Dixon images. The fully automated segmentation was then run on the same images and the resulting fat fractions compared with the manual results. The proposed k-means approach classifies each image pixel according to signal intensity and creates image masks for bone, muscle and fat in three dimensions. The pixel counts from bone, muscle and fat are automatically measured to produce the volume and mean fat fraction.

Results We compared the mean fat fraction for the two approaches and found linear correlation was good (R2=0.9396). Manual segmentations typically took 40 min or more to execute, compared to the automated segmentations, which required less than 5 min on a standard desktop computer

Abstract FRI0582 – Figure 1

Example Fat Fracton Map

Conclusions In the present study, we reported a segmentation framework based on unsupervised k-means to measure muscle volume and fat fraction. It offers time savings versus manual segmentations and correlates well with fat fraction measurements. This could be useful for muscle quantification in the fields of osteoarthritis, sports medicine and rehabilitation. Further studies are planned to compare sensitivity of automatically acquired measures to clinical progression.

Acknowledgements 1. Amato AA et al. Neurology. 2014 83( 24:2239–46.

2. Beattie, K., Davison, M.J., Noseworthy, M. et al. Rheumatol Int (2016) 36: 855.

Disclosure of Interest None declared

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.