AB1145 FULLY CONVOLUTIONAL NEURAL NETWORK-BASED SEGMENTATION OF INDIVIDUAL MUSCLES IN MR IMAGES USING MUSCLES AND BORDERS PARCELLATIONS. (June 2019)

• Record Type: Journal Article
• Authors: Fournel, Joris; Troter, Arnaud Le; Guis, Sandrine; Bendahan, David; Ghattas, Badih
• Is Part Of: Annals of the rheumatic diseases. Volume 78(2019)Supplement 2
• Page Start: 2034
• Page End: 2034
• Publication Date: 2019-06
• Subjects: Rheumatism -- Periodicals; 616.723005
• DOI: 10.1136/annrheumdis-2019-eular.2804
• Languages: English
• ISSNs: 0003-4967

Abstract

Abstract : Background: Segmentation of individual muscles in MR images is challenging considering the poor contrast between muscles and the large variability between and within subjects. Objectives: The segmentation performance of the Bayesian SegNet network was assessed for the four individual muscles of the quadriceps group. In addition to the classes corresponding to each muscle, we analyzed the effect of adding four additional classes corresponding to muscle borders. We also investigated the network performance taking into account each muscle individually or the whole set of muscles. The corresponding results were compared with those obtained using a conventional multi-atlas method. Methods: For the training phase, a dataset of 500 images was used while the testing phase was performed for two other datasets with 140 images each. Four different variants of the same network were assayed considering simultaneous segmentation of individual muscles (On5), separate segmentation of individual muscles (Fn2) and the use of additional classes related to muscle borders in both cases (On9 and Fn3). Results: All approaches largely outperformed the results of a multi-atlas strategy. The higher DSI values i.e. 0.96 ± 0.01 for the rectus femoris muscle, 0.93 ± 0.01 for the vastus intermedius muscle, 0.94 ± 0.03 for the vastus lateralis muscle and 0.96 ± 0.01 for the vastus medialis muscle were obtained with the On9 and Fn3 networks i.e. taking into account the muscle borders labels in addition to the muscle labels. Conclusion: Deep-learning based methods are optimal for the segmentation of thigh muscles and the corresponding efficiency can be improved when considering labels for muscles together with borders. Disclosure of Interests: None declared