landmark
Generative models improve fairness of medical classifiers under distribution shifts
Ira Ktena, Olivia Wiles, Isabela Albuquerque · Google DeepMind · 2024-04
Why it matters
The strongest downstream evidence in medical generative imaging: pixel-space cascaded diffusion models, with no VAE anywhere, generate synthetic training data that improves both in-distribution and out-of-distribution diagnostic accuracy and narrows demographic subgroup gaps, demonstrated across chest X-ray, histopathology and dermatology. Notably, the choice of pixel space is never argued — it is inherited from the Imagen lineage the authors were using.
What this paper does
read: full textBefore this
Medical ML models degrade under distribution shift and underperform on underrepresented subgroups, and prior generative-augmentation work in medicine mostly used GANs of limited diversity and quality, evaluated on single modalities without jointly tackling robustness and fairness together.
The problem
Data acquisition and expert labeling to fix underrepresentation are often prohibitively expensive or practically impossible because of the rarity of conditions or the scarcity of clinical expertise, so the question is whether a generative model can produce steerable synthetic augmentations that close accuracy and fairness gaps without new data collection.
The idea
Train a conditional diffusion model on available labeled and unlabeled data, sample from it so that demographic attributes are represented uniformly while the true disease prevalence is preserved, and use those synthetic images to enrich the training set of a downstream diagnostic classifier.
How it works
Generation happens directly in pixel space, with a cascaded DDPM using a UNet backbone and no VAE or other learned latent anywhere in the pipeline. A low-resolution diffusion model (64 by 64) is trained first, and where higher resolution is needed a second upsampling diffusion model is trained with the same label or label-and-attribute conditioning, following the standard cascaded-DDPM literature. The paper never argues for pixel space over a latent alternative. It simply adopts the cascaded DDPM architecture as previously published, with no discussion of memory limits, invertibility, or any other reason to avoid a latent. Three modalities are tested, namely histopathology (CAMELYON17, patch classification), chest radiology (CheXpert in-distribution, ChestX-ray14 out-of-distribution) and dermatology (a US teledermatology dataset in-distribution, Australian and Colombian datasets out-of-distribution).
Evidence
This is downstream clinical-task evidence, not just distribution metrics, and it includes a genuine reader study. In histopathology, synthetic data plus color augmentation gave a 48.5 percent relative improvement over baseline and 3.2 percent over color augmentation alone in the most skewed low-data setting, while closing the in-distribution fairness gap between hospitals by 20 absolute percentage points. In chest radiology, average AUC improved 5.2 percent out of distribution, with cardiomegaly AUC improving 21.1 percent, the sex fairness gap narrowing 44.6 percent, and the race fairness gap narrowing 31.7 percent in distribution. In dermatology, label-and-property conditioning gave 27.3 percent better high-risk sensitivity in distribution and 63.5 percent out of distribution in the more skewed setting, closing the fairness gap by a factor of 7.5 out of distribution. A reader study then checked whether the synthetic images were themselves clinically credible, three dermatologists rating 488 synthetic images reached 56.0 plus or minus 11.9 percent top-1 diagnostic accuracy and 67.7 plus or minus 12.5 percent top-3 accuracy, comparable to 101 board-certified dermatologists rating 789 real images at 54.0 plus or minus 21.1 percent top-1 and 67.1 plus or minus 22.7 percent top-3. This combination of downstream accuracy and fairness gains across three modalities, corroborated by a reader study, is stronger evidence than the FID- or SSIM-only comparisons common elsewhere in the medical section.
Limitations
The paper concedes it is not a substitute for genuine data collection and should only complement it. It concedes that a poor or biased generative model could exacerbate rather than fix subgroup disparities, that binary sex and sex-and-gender conflation and the Fitzpatrick skin-type scale carry their own known limitations, and that its conclusions are only as trustworthy as the evaluation datasets used to draw them. It also concedes that only about half of the generated dermatology images were rated diagnosable quality at all, and that memorization and privacy guarantees for rare attributes remain open problems for future work.
Why it matters
This is the strongest downstream clinical evidence in the whole medical section precisely because it combines accuracy gains, fairness gains, three independent modalities and clinician judgment of the synthetic images themselves, all without any learned latent space in the pipeline. It is also a pointed negative result for the pixel-space line's own reasoning, since the paper never argues for pixel space at all, so its strength as evidence for pixel-space generation is really strength as evidence for cascaded diffusion trained the way the Imagen lineage already did, inherited rather than justified.
Research line
Method note — the shared flow-matching interpolation
Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.