Background <p>This study sought to evaluate the effectiveness of a machine learning (ML) model utilizing multimodal MRI in distinguishing cerebral small vessel disease (CSVD) patients with depression (CSVD + D) from those without depression (CSVD − D).</p> Materials and methods <p>This retrospective study involved 198 participants from three centers, who were divided into training (<i>n</i> = 113; CSVD + D = 56, CSVD − D = 57), external validation 1 (<i>n</i> = 85; CSVD + D = 31, CSVD − D = 54), and 2 (<i>n</i> = 102; CSVD + D = 39, CSVD − D = 63) cohorts. Structural, functional, and diffusion tensor imaging was used to extract features, which were utilized to construct ML models based on nine ML classifiers. The efficacy of the models was evaluated through the receiver operating characteristic (ROC) analysis. SHapley additive explanations (SHAP) analysis provided deep insights into the model’s interpretability.</p> Results <p>Twelve features from multimodal MRI were finally identified. The eXtreme Gradient Boosting (XGBoost) classifier performed the best. The XGBoost-based multimodal model integrating features from all three modalities achieved high diagnostic performance, with areas under the ROC curves of 0.958, 0.893, and 0.917 in the training, external validation 1 and 2 cohorts, respectively, and accuracies of 0.929, 0.871, and 0.853. The SHAP results revealed that the key contributing features included elevated amplitude of low-frequency fluctuations in the superior frontal gyrus and reduced fractional anisotropy in the default mode network.</p> Conclusion <p>Integrating multimodal MRI and ML may improve the classification performance for identifying CSVD + D, suggesting the potential value of multimodal imaging markers in the assessment of CSVD-related depressive symptoms.</p> Clinical trial number <p>Not applicable.</p>