Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers
Medical image classifiers often struggle with fairness because minority demographic groups—defined by age and sex—are underrepresented in training data. This leads to two problems: the classifier performs poorly on these groups, and because test sets are also small for these groups, it is difficult to tell if a performance gap is a real bias or just random noise. This paper introduces a method using a demographically-conditioned synthetic image generator to both improve training fairness and provide a more reliable way to audit model performance.
Using Synthetic Data for Training
The researchers found that the way synthetic data is introduced to a model is critical. They compared two methods: "joint augmentation," where synthetic and real data are mixed together, and "sequential pretraining," where the model is first trained on synthetic data and then fine-tuned on a small amount of real data.
The sequential approach proved significantly more effective. By using a balanced synthetic dataset as a pretraining prior, the model achieved high performance using only 1% of the available real training data. This "100x real-data efficiency" allowed the model to surpass the performance of a model trained on the full, biased real-world dataset, particularly in terms of fairness for the "worst-case" demographic cells.
Detecting Bias with Synthetic Test Sets
Because real-world test sets often lack enough samples for minority groups, the researchers used their generator to create large, synthetic "minority cohorts" for evaluation. They tested whether these synthetic sets could accurately rank the performance of different models compared to a "real oracle"—a large, well-powered test set.
The results showed that the synthetic estimator was highly reliable, perfectly reproducing the subgroup rankings of the real oracle. In cases where real test data was too sparse to provide a stable estimate, the synthetic cohort provided a much more consistent measurement of bias, effectively acting as a proxy for the missing real-world data.
Why Strategy Matters
A key takeaway from the study is that the composition of the data is not the only factor; the training schedule is equally important. When the researchers compared models trained on the same amount of data, the sequential "pretrain-then-fine-tune" strategy consistently outperformed joint augmentation.
The researchers conclude that synthetic data is most useful when it is used to build a foundational understanding of the demographic distribution before the model is refined on real-world data. By treating synthetic data as a representation prior rather than just extra training examples, practitioners can create models that are both more accurate and significantly fairer across diverse demographic groups.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!