SC-GAN

SC-GAN synthesizes multimodal 3D neuroimaging data using a conditional generative adversarial network to generate PET, fractional anisotropy (FA), and mean diffusivity (MD) maps from T1-weighted (T1w) and FLAIR MRI inputs.

Key Features:

• 3D Conditional Generative Adversarial Network: Implemented as a 3D conditional GAN for volumetric multimodal image synthesis.
• Spectral Normalization: Controls the Lipschitz constant of the discriminator to stabilize adversarial training.
• Self-Attention Module: Captures long-range relationships between widely separated image voxels to improve synthesis fidelity.
• Feature Matching: Enforces similarity between generated images and target distribution feature representations to improve accuracy.
• Adversarial Learning Framework: Uses adversarial loss to promote realistic image generation.
• Superresolution Capability: Applied to superresolution tasks using multidimensional diffusion MRI data.
• Evaluation Metrics: Quantitative assessment uses normalized root mean square error (RMS), peak signal-to-noise ratio (SNR), and structural similarity.
• Ablation Study Findings: Identified adversarial learning and self-attention as principal contributors to performance, with spectral normalization also beneficial.

Scientific Applications:

• Multimodal Neuroimaging Synthesis: Predicts PET, FA, and MD maps from T1w and FLAIR MRI inputs for multimodal MRI-to-target modality translation.
• Diffusion MRI Superresolution: Performs superresolution on multidimensional diffusion MRI data, improving spatial detail relative to baseline models.
• Benchmarking and Quantitative Evaluation: Demonstrated lower normalized RMS error and higher SNR and structural similarity on ADNI-3, with prediction error reductions of 18% for FA, 24% for PET, and 29% for MD compared to a 2D conditional GAN, and outperforming other 3D GAN models and 3D Unet in reported evaluations.

Methodology:

Architecture and training employ a 3D conditional GAN with spectral normalization, feature matching, a self-attention module, and adversarial learning; evaluation was performed on the ADNI-3 dataset and a superresolution task using multidimensional diffusion MRI.