Investigating the Effect of Synthetic Medical Images on Fairness in Medical Deep Learning Research

PROJECT ABSTRACT This grant proposal describes a research project that aims to improve the fairness of deep learning models for pathology detection in chest radiographs. Health equity, the state in which everyone has a fair and just opportunity to attain their highest level of health, is the cornerstone of a fair and just society. However, racial

and ethnic minority groups in the United States experience higher rates of illness and death across a wide range of health conditions. With the adoption of Artificial Intelligence (AI) and Deep Learning (DL) in healthcare, there is growing concern about increased disparities through the use of algorithms.

To address this issue, the research team proposes to leverage generative modeling to better represent minority groups in training data. Specifically, they will train DL models that detect 14 pathologies from publicly available chest radiographs with patient age, sex and race information. They will use Denoising Diffusion

Probabilistic Models (DDPMs) to create synthetic data and augment the dataset with more diverse images. The research team expects that engineered image synthesis will train DL models that reliably detect chest pathologies without being biased on race or sex. To achieve this goal, they have proposed three aims: 1)

Establish a baseline for pathology detection in chest radiographs; 2) Augment the real radiographs with synthetic chest radiographs representing minorities; and 3) Assess the impact of synthetic data on model fairness. The proposed research leverages the power of DL image generation algorithms to potentially improve the

accuracy and fairness of pathology detection in chest radiographs. Additionally, the generative model will be released publicly as a foundational model for researchers without access to the required computational resources to train such models. This approach has the potential to improve healthcare outcomes for

underserved populations and advance the field of fairness in AI research.