Multi-Feature 3D Deep Learning for Predicting Antidepressant Treatment Response from Structural MRI

Objective Accurate prediction of treatment response in Major Depressive Disorder (MDD) remains a significant challenge due to individual neuroanatomical heterogeneity and the inherent scarcity of labeled clinical data. To address these limitations, we propose MAN-MDDNet, a 3D deep learning framework designed for early and precise prognosis by leveraging multi-feature structural magnetic resonance imaging (sMRI). Methods MAN-MDDNet leverages 3D anatomical representations learned from large-scale medical datasets to overcome the “small-sample” bottleneck typical in neuroimaging, ensuring robust feature convergence and stability. The model is inherently brain-structure-driven and uniquely fuses four complementary features—gray matter volume/density and white matter volume/density—to capture a holistic view of structural alterations. The integration of a spatial attention mechanism (SEBlock) enables the architecture to selectively prioritize biologically relevant regions. The model was rigorously validated across two independent clinical centers (adults from Hangzhou and adolescents from Jinan). Results MAN-MDDNet significantly outperformed traditional machine learning and baseline deep learning models, achieving accuracies of 85.90%–87.18% across the two clinical centers. Notably, t-SNE revealed high structural heterogeneity in the non-remission group, indicating complex resistance mechanisms. Conclusion This study presents an interpretable deep learning framework that integrates multi-feature sMRI to predict antidepressant treatment response, which may assist in personalized management of MDD.