GAN objective comparison for synthetic MRI generation

What's actually being asked

Map each observed symptom to its likely mechanism, then compare the standard, least-squares, and Wasserstein GAN objectives on how well each addresses stability and image quality for this medical-imaging setting.

Answer

Diagnosing the symptoms

• Mode collapse — G covers only a narrow slice of real anatomical variability; the standard sigmoid-based minimax loss gives no explicit diversity incentive (same mechanism discussed for digit GANs).
• Unstable loss curves — alternating gradient updates on a non-convex minimax game with a saturating sigmoid D; under the original objective, loss values don't correlate well with actual sample quality, so the curves swing without telling you much useful.
• Artifacts — typical when a strided-conv/transposed-conv G is pushed by a D that locally overpowers it, repeatedly finding and reacting to small checkerboard/edge "tells," and when gradients vanish or explode under the saturating loss, leaving some regions under-corrected.
• Insufficient fine detail — a classic symptom of D saturating on big, obvious fakes and never being forced to push G on local, high-frequency texture — gradient dries up before G is pressured to refine detail.

Standard GAN objective

min_G max_D V(D,G) = 𝔼_{x∼p_data}[log D(x)] + 𝔼_{z∼p_z}[log(1 − D(G(z)))]

D has a sigmoid output and behaves as a binary classifier. Mechanism problem: once D confidently separates real MRI scans from fakes (easy early on, since initial fakes are poor), log(1 − D(G(z))) sits in a flat region of its curve and G's gradient vanishes — directly producing the vanishing-gradient and poor-fine-detail symptoms. Because the loss doesn't move smoothly with actual quality, the observed curves look unstable.

Least-Squares GAN (LSGAN)

D loss: min_D (1/2) 𝔼_{x∼p_data}[(D(x) − 1)²] + (1/2) 𝔼_{z∼p_z}[(D(G(z)))²] G loss: min_G (1/2) 𝔼_{z∼p_z}[(D(G(z)) − 1)²]

D has no sigmoid at its output — it's scored as a regression target (1 for real, 0 for fake) with squared error. Mechanism: unlike cross-entropy, squared error keeps penalizing a sample even once it's on the "correct" side of D's decision boundary but still far from it — it keeps pulling fakes continuously toward the real-data manifold instead of stopping once "good enough to fool D." This directly helps fine detail (G keeps receiving gradient to close the remaining gap) and gives smoother loss curves, since the penalty is a continuous quadratic rather than a saturating log. It reduces mode collapse somewhat but less decisively than Wasserstein-based methods.

Wasserstein GAN (WGAN / WGAN-GP)

Critic objective (real-valued score, no sigmoid), with the critic constrained to be 1-Lipschitz (weight clipping, or the softer gradient-penalty variant): max_D 𝔼_{x∼p_data}[D(x)] − 𝔼_{z∼p_z}[D(G(z))] G objective: min_G − 𝔼_{z∼p_z}[D(G(z))] WGAN-GP adds a penalty λ 𝔼[(‖∇_x̂ D(x̂)‖₂ − 1)²] on points x̂ interpolated between real and fake samples, to softly enforce the Lipschitz constraint instead of hard clipping.

Mechanism: the Wasserstein (earth-mover) distance stays meaningful and gives non-zero, informative gradients even when the real and fake distributions barely overlap — exactly the early-training regime for a hard domain like MRI scans. This is the strongest answer to mode collapse (a smoother distance landscape gives less incentive to hide in one mode) and to unstable loss curves (the critic's loss now actually tracks sample quality, something the standard GAN loss never reliably gave). The Lipschitz constraint also caps how sharply the critic can react to small pixel-level changes, reducing artifacts driven by an overly reactive D.

Comparison summary

Objective	Vanishing gradients	Mode collapse	Loss ↔ quality correlation	Fine detail
Standard GAN	high risk (sigmoid saturates)	high risk	poor	weak
LSGAN	reduced (quadratic, non-saturating penalty)	somewhat reduced	better (smooth quadratic)	improved
WGAN(-GP)	low (Lipschitz critic avoids saturation)	lowest (smooth EM distance)	best	best

Recommendation for this scenario

For synthetic MRI generation, where both texture fidelity and distributional coverage carry clinical weight, WGAN-GP is the strongest default — it addresses stability, diversity, and gives a loss signal that's actually usable for monitoring training. LSGAN is a reasonable lighter-weight fallback when full Lipschitz enforcement is too costly to train.

Exam tip

Write out all three loss formulas rather than just naming the methods — "LSGAN uses squared error" without the formula costs marks. Explicitly tie the "no sigmoid saturation" (LSGAN) and "Lipschitz-constrained critic" (WGAN) mechanisms to the specific symptom each one fixes.