Automated Classification of Neuromuscular Diseases Using Thigh Muscle MRI With Model Interpretations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Neuromuscular diseases (NMDs) diagnosis relies on clinical evaluation, electromyography, nerve conduction studies, blood tests, muscle biopsy, and genetic testing.
• Muscle MRI visualizes affected areas, identifying fatty replacement, atrophy, and edema, with distinct patterns aiding diagnosis.
• Current diagnostic methods can be invasive and time-consuming, highlighting the need for advanced, non-invasive techniques.

Purpose of the Study:

• To develop an automated approach for classifying NMDs using upper leg symptomatic MRI scans.
• To provide interpretable explanations for the classification model's decisions.
• To validate the model's performance against established diagnostic criteria and literature.

Main Methods:

• Utilized 109 Dixon upper leg MRI scans from four NMDs (LGMDR12, BMD, DM1, CMT1A) and healthy controls.
• Trained a U-Net model for muscle segmentation, calculated fat fractions, and used these as input for a random forest classifier.
• Employed SHapley Additive exPlanations (SHAP) to interpret model reasoning and compared findings with medical literature.

Main Results:

• Achieved an overall accuracy of 89% in distinguishing between NMD classes and healthy controls.
• Demonstrated high Area Under the Curve (AUC) values for all classes (0.972-0.997).
• Confirmed that model performance was consistent regardless of segmentation accuracy or feature extraction method, and SHAP explanations aligned with known muscle involvement patterns.

Conclusions:

• A fully automated method effectively distinguishes between four NMDs and healthy controls using MRI data.
• The developed approach offers an accurate and interpretable tool for NMD diagnosis.
• This automated method holds promise for improving the efficiency and accuracy of NMD diagnostics.