AB0945 An Atlas Based Automatic Segmentation of The Human Thigh Muscles: A Promising Approach for Muscle Volume Quantification in Longitudinal Studies

• Published in: Annals of the rheumatic diseases Vol. 75; no. Suppl 2; p. 1224
• Main Authors: Le Troter, A., Fouré, A., Guye, M., Confort-Gouny, S., Mattei, J.-P., Guis, S., Gondin, J., Salort-Campana, E., Bendahan, D.
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Limited 01.06.2016
• ISSN: 0003-4967; 1468-2060
• DOI: 10.1136/annrheumdis-2016-eular.1385

BackgroundAtlas segmentation is a powerful method for automatic structural segmentation of several sub-structures in many organs. However, so far, very few studies have assessed this method for skeletal muscle segmentation.ObjectivesIn this study, we present an atlas-based pipeline we have developed for segmentation of quadriceps muscles from magnetic resonance images obtained in twenty five young healthy males and illustrate the potential of this approach for a longitudinal follow-up in a group of 7 subjects.MethodsT1-weighted axial MR images from the thigh were recorded once in a group of 25 subjects (G1) and twice (a few days apart) in another group of 7 subjects (G2). The multi-atlas library was built from G1. In order to do so, MR images were initially automatically segmented for bone, muscle and fat tissues and each muscle of the thigh anterior compartment was manually delineated. The same process was applied to G2 in order to obtain the ground-truth set of images. Then each MRI dataset from G1 was registered to each MRI dataset of G2 using a non-linear registration process. An optimal fusion method was then used applied in order to obtain the highest DICE index, a similarity index between the atlas-based manual segmentation and the ground-truth set of images. A single-atlas version of the pipeline was used without the fusion process and a similar validation was performed for a longitudinal follow-up study.ResultsIn control subjects, the results for each quadriceps muscle show a mean DICE similarity coefficient higher than 0.85. While the multi-atlas pipeline did not provide sensitive enough for muscle volume quantification, the single-atlas version proved to be far better.ConclusionsIn the present study we reported two segmentation pipelines based on atlases. The examples provided in a control population demonstrated a robust method which could be useful for muscle quantification in the fields of neuromuscular disorders, sports medicine and rehabilitation.Disclosure of InterestNone declared