Integrating Machine Learning and Deep Learning Architectures for Automated, Explainable Classification of Leigh Syndrome from Magnetic Resonance Imaging

Leigh Syndrome (LS) is an incurable mitochondrial disorder, affecting approximately 1 in 40,000 individuals, and is characterized by bilaterally symmetric T2-weighted hyperintensities in the brainstem, cerebellum, and basal ganglia. LS has a poor prognosis, with a life expectancy of 2.4 years. Although Magnetic Resonance Imaging (MRI) is the primary diagnostic tool, earlier detection and monitoring could significantly improve outcomes, particularly in early-onset cases. This study develops and compares four machine learning architectures—nnU-Net, a hybrid CNN–transformer, DenseNet-121, and InceptionV3—for automated LS classification using 171 MRI scans from Nanavati Hospital (Mumbai, India). To generate a more comprehensive dataset, the dataset was augmented to 4,000 images (2,000 LS, 2,000 non-LS) and split 75/25 for training and testing. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. DenseNet-121 demonstrated the strongest performance, achieving the highest overall classification metrics and maintaining balanced predictions across classes. To improve interpretability, Local Interpretable Model-Agnostic Explanations (LIME) were applied, showing that model predictions corresponded to clinically relevant brain regions associated with LS pathology. Limitations include reliance on a single-center dataset and use of conventional structural MRI without advanced modalities such as fMRI, perfusion imaging, or spectroscopy, which may enhance diagnostic precision. Overall, this work highlights the potential of explainable, multi-model machine learning frameworks as assistive tools for LS diagnosis. Future validation across multi-center datasets and integration of advanced imaging techniques could further improve clinical applicability and accessibility.