Background: Pediatric leukodystrophies are rare inherited white matter disorders characterized by heterogeneous neuroimaging patterns that complicate accurate and early subtype classification. Despite their clinical importance for treatment planning, genetic counseling, and prognosis assessment, automated artificial intelligence-based diagnostic tools remain limited for these disorders.

Methods : In this study, we propose a comprehensive machine learning and deep learning framework for the classification of five clinically relevant leukodystrophy subtypes using multimodal axial brain MRI, including T1, T2, and FLAIR sequences acquired from 115 pediatric patients. Following image pre-processing and atlas-based segmentation, region-level representations were extracted and three complementary classification strategies were investigated: (i) radiomics-based machine learning models using handcrafted imaging features; (ii) an end-to-end 3D transfer-learning approach employing pre-trained CNN architectures, including ResNet18, ResNet50, and DenseNet121; and (iii) hybrid models in which CNN-derived deep features were used either independently or in combination with radiomic descriptors for supervised classification.

Results : The results demonstrated that fine-tuned transfer learning models outperformed radiomics-based and hybrid approaches across all evaluation metrics. Among the evaluated architectures, the best-performing end-to-end model, DenseNet121, achieved an accuracy of 95.50%, a precision of 89.62%, a recall of 89.19%, an F1-score of 88.88%, and a specificity of 97.12%.

Conclusion : These findings indicate that 3D transfer learning can effectively capture complex spatial and hierarchical volumetric MRI patterns for multiclass pediatric leukodystrophy classification and may serve as a promising decision-support tool in clinical settings with limited data availability.