Comprehensive Summary
This study investigates the use of radiomics and machine learning (ML) to differentiate Duchenne muscular dystrophy (DMD) from Becker muscular dystrophy (BMD) in children aged 36–60 months using MRI Dixon sequences. A total of 62 male patients (41 with DMD and 21 with BMD) underwent T2-weighted Dixon MRI imaging, from which 465 radiomic features were extracted. The study focused on the gluteus maximus muscle and used a deep learning-based segmentation tool for consistent region identification. After applying statistical methods such as the Mann–Whitney U test, Pearson correlation filtering, and LASSO regression, the most predictive features were selected for five ML classification models: Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), Multilayer Perceptron (MLP), and K- Nearest Neighbors (KNN). The ML models were evaluated via five-fold cross-validation. While radiologists achieved a sensitivity of 95.1%, their specificity was only 19.0%, misclassifying most BMD patients. In contrast, ML models showed superior performance with accuracy ranging from 81.2% to 90.6%, specificity from 71.0% to 86.0%, and F1 scores from 85.2% to 92.6%. The MLP model achieved the highest performance across metrics.
Outcomes and Implications
Early and accurate differentiation between Duchenne muscular dystrophy (DMD) and Becker muscular dystrophy (BMD) in pediatric patients is vital due to the diseases' vastly different progression rates, treatment responses, and life expectancies. Misdiagnosis can lead to inappropriate use of corticosteroids or delayed access to specialized care. In this study, traditional radiologist interpretation, though sensitive (95.1%), suffered from poor specificity (19.0%), misclassifying a majority of BMD cases and limiting its clinical reliability. In contrast, the machine learning (ML) models exhibited high and balanced diagnostic performance. The best-performing model, Multilayer Perceptron (MLP), achieved an accuracy of 90.6%, a specificity of 86.0%, and an F1 score of 92.6%, reflecting a strong ability to distinguish between the two muscular dystrophy subtypes. Even the lowest-performing ML model, Logistic Regression, still outperformed radiologists in specificity (71.0%) while maintaining robust accuracy (81.2%). The clinical significance of these results lies in the potential for AI to dramatically reduce misclassification rates, particularly for BMD, leading to more precise treatment planning and resource allocation. Radiomics extracted 465 quantitative features from the gluteus maximus muscle, allowing ML to leverage nuanced, non-visible imaging cues to drive diagnostic decisions. This method replaces subjective interpretation with objective, repeatable, and data- driven outcomes. Moreover, the use of deep learning-based segmentation tools enhances consistency in image analysis, and the entire diagnostic workflow remains non-invasive, making it suitable for young children who often require minimal imaging discomfort. These advancements suggest that integrating AI into early neuromuscular diagnostic pathways could standardize detection, shorten time to diagnosis, and lay the groundwork for more effective, personalized care interventions. Scaling these findings through multicenter studies with larger cohorts could validate the generalizability of this AI-based approach, paving the way for regulatory approval and clinical adoption in pediatric neurology and radiology. The machine learning models in this study, particularly the MLP (accuracy: 90.6%, specificity: 86.0%, F1 score: 92.6%), demonstrated much higher diagnostic precision. These results suggest that integrating radiomics and ML into diagnostic workflows can enhance early-stage identification of muscular dystrophy subtypes, supporting more informed therapeutic decisions. Radiomics allows the quantification of subtle MRI features not easily detectable by human eyes, providing reliable evidence for early classification. This non-invasive, reproducible method holds promise for broader application and could be expanded through multicenter studies to improve generalizability and long-term patient outcomes.