neurogan-3d

NeuroGAN-3D methodology for high-fidelity 3D generative super-resolution of resting-state fMRI (rs-fMRI) spatial maps. Uses a GAN architecture to enhance spatial resolution of volumetric functional br

SKILL.md

GitHub

NeuroGAN-3D: High-Fidelity 3D Generative Super-Resolution for fMRI

Enhances the spatial resolution of resting-state fMRI (rs-fMRI) spatial maps using a generative adversarial network tailored to volumetric neuroimaging computational demands.

Problem Statement

Spatial resolution of rs-fMRI-derived spatial maps determines the ability to:

  • Localize functional units with precision
  • Perform reliable brain parcellation
  • Detect subtle, spatially-specific neurobiological alterations (development, aging, disease)

Existing super-resolution methods are not optimized for volumetric neuroimaging or fail to capture fine-grained spatial patterns in functional connectivity maps.

Architecture

Generator Network

Low-Res 3D fMRI volume → Encoder → Latent representation → Decoder → High-Res 3D volume
  • 3D convolutional encoder capturing spatial structure
  • Residual learning for high-frequency details
  • Volumetric upsampling layers (3D transposed convolutions)

Discriminator Network

High-Res 3D volume (real/generated) → 3D CNN → Real/Fake classification
  • 3D convolutional discriminator evaluating spatial realism
  • Patch-based discrimination for localized quality assessment

Key Contributions

  1. First 3D GAN specifically designed for rs-fMRI spatial maps
  2. Significantly outperforms conventional interpolation baselines in preserving fine-grained spatial patterns of intrinsic functional networks
  3. Preserves biologically meaningful connectivity patterns while enhancing spatial resolution
  4. Enables downstream analysis at higher effective resolution

Implementation Pattern

import torch
import torch.nn as nn

class Generator3D(nn.Module):
    def __init__(self, in_channels, out_channels, base_filters=32):
        super().__init__()
        # Encoder
        self.enc = nn.Sequential(
            nn.Conv3d(in_channels, base_filters, 3, padding=1),
            nn.LeakyReLU(0.2),
            nn.Conv3d(base_filters, base_filters*2, 3, stride=2, padding=1),
            nn.LeakyReLU(0.2),
            nn.Conv3d(base_filters*2, base_filters*4, 3, stride=2, padding=1),
            nn.LeakyReLU(0.2),
        )
        # Decoder with upsampling
        self.dec = nn.Sequential(
            nn.ConvTranspose3d(base_filters*4, base_filters*2, 4, stride=2, padding=1),
            nn.ReLU(),
            nn.ConvTranspose3d(base_filters*2, base_filters, 4, stride=2, padding=1),
            nn.ReLU(),
            nn.Conv3d(base_filters, out_channels, 3, padding=1),
        )

    def forward(self, x):
        return self.dec(self.enc(x))

Training Considerations

  • Use perceptual loss + adversarial loss for preserving biological structure
  • Consider cycle consistency for validation
  • Evaluate with both image quality metrics and downstream FC analysis
  • Validate preserved connectivity patterns against ground truth (when available)

Datasets and Validation

  • Tested on rs-fMRI spatial maps
  • Evaluated against conventional interpolation baselines (trilinear, bicubic)
  • Measured by spatial fidelity metrics and preservation of functional network patterns

Activation Keywords

  • neurogan-3d, fMRI super-resolution, 3D GAN neuroimaging, volumetric brain maps, rs-fMRI enhancement, spatial resolution fMRI, generative super-resolution brain, functional map upsampling, GAN neuroimaging