A hybrid 3D-CNN and CGAN approach for bipolar disorder diagnosis using VBM and SBC features from multimodal MRI

Abstract

Neuroimaging technologies like structural and functional MRI (sMRI/fMRI) have identified biomarkers for bipolar disorder (BD). Deep learning approaches have shown promising performance in analyzing such data; however, many multimodal models rely on complex architectures with limited biological grounding of their inputs. We developed a diagnostic model using sMRI and fMRI data from 49 BD patients and 121 healthy controls (UCLA CNP study). Rather than raw images, we extracted biologically meaningful features: voxel-based morphometry (VBM) from sMRI and seed-based functional connectivity (SBC) from fMRI. These features were processed using a dual-branch 3D convolutional neural network (3D-CNN). Conditional Generative Adversarial Networks (CGANs) performed feature-space augmentation exclusively on training data after dataset splitting. Model performance was evaluated using nested 5-fold cross-validation. The proposed multimodal framework achieved a mean accuracy of 93.49% and an AUC of 0.97 during repeated nested cross-validation. On an independent held-out test set, the model achieved an accuracy of 93.02%, sensitivity of 83.33%, and specificity of 96.77%. This outperformed the single-modality sMRI model (81.39%) and the three-seed fMRI model (88.37%). CGAN-based augmentation consistently improved classification accuracy across repeated cross-validation runs, with a mean gain of 3.02% (95% bootstrap CI: 1.86–4.19%). These findings suggest that multimodal integration of structural and functional neuroimaging features may improve bipolar disorder classification performance. The proposed framework leverages biologically meaningful feature representations and feature-space augmentation, providing a proof-of-concept approach for multimodal neuroimaging-based classification. Further validation on independent and multi-site datasets is required to assess the robustness, generalizability, and potential translational relevance of the proposed framework.