Super-Resolution GAN

Unlike traditional pixel-wise losses like Mean Squared Error (MSE) that produce overly smooth textures, SRGANs use a perceptual loss based on high-level feature maps extracted from a pre-trained deep network, such as VGG-19. This loss minimizes the semantic difference between the generated high-resolution image and the ground truth, compelling the generator to reconstruct fine anatomical textures and edges that are perceptually convincing to a radiologist. The total loss is a weighted combination of content loss and adversarial loss.

The generator architecture relies on deep Residual-in-Residual Dense Blocks (RRDB) to increase network capacity without suffering from vanishing gradients. Key structural elements include:

• Dense connections: Each layer receives feature maps from all preceding layers, improving information flow.
• Residual scaling: The output of dense blocks is scaled down before being added to the identity pathway, stabilizing training.
• Skip connections: Long-range skip connections preserve low-frequency structural information, allowing the network to focus on learning high-frequency details.

The discriminator network is trained to distinguish between original high-resolution medical scans and the generator's super-resolved outputs. This adversarial push forces the generator to move beyond minimizing mathematical error and instead produce images that fall within the manifold of realistic medical imagery. For medical applications, this is critical for hallucinating plausible micro-architectural details like trabecular bone patterns or subtle lesion boundaries that are absent in the low-resolution input.

Standard quantitative metrics like PSNR and SSIM often fail to correlate with diagnostic quality. SRGANs are frequently evaluated using Mean Opinion Score (MOS) testing, where trained radiologists blindly rate the diagnostic fidelity of the super-resolved images against reference high-resolution scans. This human-centric evaluation ensures the model enhances clinically relevant features rather than introducing visually plausible but anatomically incorrect artifacts, a critical safety check for regulatory clearance.

A common failure mode in medical super-resolution is the introduction of high-frequency artifacts that mimic pathology. Advanced SRGAN variants incorporate specific countermeasures:

• Total Variation (TV) regularization: Penalizes rapid pixel intensity changes to suppress noise amplification.
• Frequency-domain losses: Compare the Fourier spectra of generated and real images to ensure accurate reconstruction of spatial frequencies.
• Multi-scale discriminators: Evaluate image realism at multiple resolutions to catch artifacts at different scales.

For CT and MRI volumes, applying 2D SRGANs slice-by-slice introduces inter-slice flickering artifacts. Dedicated 3D SRGAN architectures use 3D convolutional kernels in both the generator and discriminator to learn spatial continuity along the z-axis. This ensures that super-resolved anatomical structures, such as blood vessels or organ boundaries, remain smooth and continuous when scrolling through a volumetric reconstruction, maintaining diagnostic integrity across all planes.