This dissertation develops deep learning frameworks to advance brain characterization and disorder diagnosis, while addressing major challenges in neuroimaging, including high data dimensionality, limited labeled samples, interpretability, and missing modalities across datasets. I introduce representation learning models, including Siamese architectures and supervised contrastive learning, to capture subject-specific spatial and functional brain patterns. Building on these representations, I investigate the classification of Alzheimer’s disease (AD) and schizophrenia (SZ) using functional network connectivity (FNC), identifying disorder-specific signatures that contribute to differential diagnosis. For further clinical applications, I design 3D CNNs for whole-brain classification in epilepsy and demonstrate their effectiveness in distinguishing MRI-positive and MRI-negative patients. Moving to generative modeling, I propose Cycle-GAN–based approaches to impute missing modalities in Alzheimer’s disease and show that synthetic data can enhance downstream multimodal classification. Finally, I extend this work by introducing GAN-guided diffusion models that leverage cycle-consistency to synthesize cross-modality data more stably and with greater diversity than GANs alone. Collectively, these contributions advance both predictive modeling and generative synthesis for multimodal neuroimaging, providing tools for biomarker discovery, disease diagnosis, and bridging incomplete datasets.