A deep learning tool without muscle-by-muscle grading to differentiate myositis from facio-scapulo-humeral dystrophy using MRI

Vincent Fabry; Franck Mamalet; Anne Laforet; Mikael Capelle; Blandine Acket; Coralie Sengenes; Pascal Cintas; Marie Faruch-Bilfeld

doi:10.1016/j.diii.2022.01.012

A deep learning tool without muscle-by-muscle grading to differentiate myositis from facio-scapulo-humeral dystrophy using MRI

Diagn Interv Imaging. 2022 Jul-Aug;103(7-8):353-359. doi: 10.1016/j.diii.2022.01.012. Epub 2022 Mar 12.

Authors

Vincent Fabry¹, Franck Mamalet², Anne Laforet³, Mikael Capelle², Blandine Acket³, Coralie Sengenes⁴, Pascal Cintas³, Marie Faruch-Bilfeld⁵

Affiliations

¹ Department of Neurology, Toulouse University Hospital, 31059 Toulouse, France. Electronic address: fabry.v@chu-toulouse.fr.
² IRT-Saint Exupéry, 31400 Toulouse, France.
³ Department of Neurology, Toulouse University Hospital, 31059 Toulouse, France.
⁴ StromaLab, Bâtiment INCERE, 31100 Toulouse, France.
⁵ Department of Radiology, Toulouse University Hospital, 31059 Toulouse, Cedex 9, France.

PMID: 35292217
DOI: 10.1016/j.diii.2022.01.012

Abstract

Purpose: The purpose of this study was to assess the capabilities of a deep learning (DL) tool to discriminate between type 1 facioscapulo-humeral dystrophy (FSHD1) and myositis using whole-body muscle magnetic resonance imaging (MRI) examination without the need for visual grading of muscle signal changes.

Materials and methods: A total of 40 patients who underwent whole-body MRI examination that included T1-weighted and STIR sequences were included. There were 19 patients with proven FSHD1 (9 men, 10 women; mean age, 47.7 ± 18.0 [SD] years; age range: 20-72 years) and 21 patients with myositis fulfilling European Neuromuscular Centre criteria and European League Against Rheumatism and American College of Rheumatology criteria (11 men, 10 women; mean age, 59.3 ± 17.0 [SD]; age range: 19-78 years). Based on thigh, calf, and shoulder sections a supervised training of a neural network was performed and its diagnostic performance was studied using a 5-fold cross validation method and compared to the results obtained by two radiologists specialized in musculoskeletal imaging.

Results: The DL tool was able to differentiate FSHD1 from myositis with a correct classification percentage respectively of 69 % (95% CI: 39-99), 75% (95% CI: 48-100) and 77% (95% CI: 60-94) when thigh only, thigh and calf or the thigh, calf, and shoulder MR images were analyzed. The percentages of correct classification of the two radiologists for these later MR images were 38/40 (95%) and 35/40 (87.5%), respectively; with no differences with DL tool correct classification (P = 0.41 and P > 0.99, respectively). Among the seven patients who were misclassified by the radiologists, the DL tool correctly classified six of them.

Conclusion: A DL tool was developed to discriminate between FSHD1 and myositis using whole-body MRI with performances equivalent to those achieved by two radiologists. This study provides a proof of concept of the effectiveness of a DL approach to distinguish between two myopathies using MRI with a small amount of data, and no prior muscle signal changes grading.

Keywords: Artificial intelligence; Deep learning; Facioscapulohumeral muscular dystrophy; Magnetic resonance imaging; Myositis.

MeSH terms

Adult
Aged
Deep Learning*
Female
Humans
Magnetic Resonance Imaging / methods
Male
Middle Aged
Muscle, Skeletal / diagnostic imaging
Muscular Dystrophy, Facioscapulohumeral* / diagnostic imaging
Muscular Dystrophy, Facioscapulohumeral* / pathology
Myositis* / diagnostic imaging
Myositis* / pathology
Young Adult