Generative Adversarial Network (GAN)

The core mechanism is a two-player minimax game. The generator creates synthetic samples from random noise, while the discriminator attempts to distinguish them from real data. The generator's loss is directly tied to the discriminator's success, creating a powerful feedback loop that forces the generator to produce statistically indistinguishable outputs. This dynamic is mathematically formalized as:

• Generator (G): Learns the true data distribution p_data(x).
• Discriminator (D): Estimates the probability that a sample came from the training data rather than G.
• Objective: min_G max_D V(D, G) = E_xp_data[log D(x)] + E_zp_z[log(1 - D(G(z)))]. This competitive training is what allows GANs to capture complex, high-dimensional distributions like those of medical scans.

A critical failure condition where the generator finds a single, highly convincing output that consistently fools the discriminator, and produces only that limited variety of samples. In medical imaging, a collapsed GAN might generate the same generic-looking chest X-ray regardless of the input noise, failing to capture the full diversity of patient anatomy and pathology. This is a primary challenge in GAN training, as it indicates the generator is not learning the full data distribution. Mitigation strategies include:

• Minibatch discrimination: Allowing the discriminator to look at multiple samples at once to detect a lack of diversity.
• Unrolled GANs: Defining the generator's loss with respect to an unrolled optimization of the discriminator to prevent exploitation of a locally weak discriminator.
• Wasserstein loss: Using the Earth Mover's distance to provide more meaningful gradients, especially when the real and generated distributions have minimal overlap.

A fundamental extension where both the generator and discriminator are conditioned on auxiliary information, such as a class label or a semantic layout map. This transforms the GAN from an uncontrolled generator into a steerable tool. In medical contexts, a cGAN can be conditioned on a semantic label map of organs to synthesize a corresponding, photorealistic CT slice, or on a patient's age and sex to generate a context-appropriate scan. This is the foundational architecture for tasks like:

• Image-to-Image Translation: Converting an MRI to a synthetic CT scan.
• Lesion Insertion: Placing a realistic tumor into a healthy scan by conditioning on a binary mask.
• Super-Resolution: Enhancing a low-resolution image by conditioning on the low-res input.

A highly influential training methodology that replaces the standard binary cross-entropy loss with the Wasserstein distance (Earth Mover's distance). This metric provides a smoother, more meaningful gradient signal, even when the real and generated distributions are far apart. The key innovations are:

• Critic instead of Discriminator: The network is called a critic, and it outputs a score rather than a probability, aiming to maximize the difference between scores for real and fake samples.
• Lipschitz Constraint: The critic must be a 1-Lipschitz function, enforced by a gradient penalty on random interpolations between real and generated data. This stabilizes training dramatically.
• Benefits: WGAN-GP is far less prone to mode collapse and provides a loss curve that correlates with sample quality, making it a robust default choice for generating high-fidelity medical images.

A technique designed to prevent the discriminator from overfitting when training data is limited, a common scenario in medical imaging. Instead of augmenting the dataset, ADA dynamically applies a set of augmentations to both real and generated images before they enter the discriminator. The key mechanism is an adaptive hyperparameter that controls augmentation probability based on the degree of overfitting:

• Overfitting Heuristic: Monitors the gap between discriminator outputs on real vs. generated sets.
• Augmentation Pipeline: Includes pixel blitting, geometric transformations, and color transforms.
• Stabilization: By forcing the discriminator to become invariant to these augmentations, ADA prevents it from memorizing the training set and provides a stronger learning signal to the generator, enabling high-quality synthesis from datasets of only a few thousand images.

A training strategy where both the generator and discriminator start by learning low-resolution versions of the image and progressively add new layers that model increasingly fine details. This dramatically stabilizes training because the networks first learn the large-scale structure of the data distribution before tackling high-frequency details. The process works as follows:

• Phase 1: Train on 4x4 pixel images to capture gross anatomy and pose.
• Phase 2-N: Smoothly fade in new layers to double the resolution to 8x8, 16x16, and so on, up to the target resolution (e.g., 1024x1024).
• Medical Benefit: This coarse-to-fine approach is highly effective for generating globally coherent anatomical structures with realistic fine textures, avoiding the common pitfall of plausible textures on an impossible anatomy.

A critical failure condition in GAN training where the generator network loses diversity and produces a highly limited set of outputs. Instead of capturing the full statistical distribution of real medical images, the generator finds a few samples that reliably fool the discriminator and exploits them repeatedly. In medical imaging, this is catastrophic—a collapsed model might generate only one type of tumor morphology, rendering the synthetic dataset useless for training a robust diagnostic classifier. Mitigation strategies include minibatch discrimination, unrolled GANs, and Wasserstein loss functions.

The gold-standard quantitative metric for evaluating synthetic image fidelity. FID measures the Wasserstein-2 distance between feature distributions extracted from real and generated images using a pre-trained Inception network. A lower FID score indicates higher visual quality and diversity. For medical imaging, achieving clinically acceptable FID is non-negotiable—it correlates with whether a downstream diagnostic model trained on synthetic data will generalize to real patient scans. Typical targets depend on modality and anatomy.

A perceptual metric that quantifies the structural degradation between a generated image and a reference ground truth. Unlike pixel-wise error metrics like MSE, SSIM evaluates image quality based on three components: luminance, contrast, and structure. In synthetic CT generation, SSIM is used alongside FID to ensure that generated Hounsfield Unit values preserve anatomical edges and tissue boundaries. A score of 1.0 indicates perfect structural identity; clinically acceptable thresholds are typically above 0.85.