Towards Fair Medical AI: Adversarial Debiasing of 3D CT Foundation Embeddings

TL;DR

The paper tackles demographic leakage in self-supervised 3D CT foundation embeddings (dimension $d=1408$) by proposing a VAE-based adversarial debiasing framework that maps embeddings to a demographic-free latent space of dimension $k=500$, preserving downstream lung cancer risk prediction. It introduces an encoder–decoder VAE with an adversarial head for age and sex, trained to minimize reconstruction and KL loss while thwarting demographic inference. Empirical results show the demographic signals become largely unrecoverable (e.g., sex predictor accuracy and AUC drop substantially) with negligible loss to 1-year and 2-year cancer prediction performance, and with improved fairness via lower EOD. The approach also demonstrates robustness to data-poisoning attacks targeting demographic groups, underscoring its potential for fairer and more secure deployment of 3D CT foundation models in clinical decision-making.$

Abstract

Self-supervised learning has revolutionized medical imaging by enabling efficient and generalizable feature extraction from large-scale unlabeled datasets. Recently, self-supervised foundation models have been extended to three-dimensional (3D) computed tomography (CT) data, generating compact, information-rich embeddings with 1408 features that achieve state-of-the-art performance on downstream tasks such as intracranial hemorrhage detection and lung cancer risk forecasting. However, these embeddings have been shown to encode demographic information, such as age, sex, and race, which poses a significant risk to the fairness of clinical applications. In this work, we propose a Variation Autoencoder (VAE) based adversarial debiasing framework to transform these embeddings into a new latent space where demographic information is no longer encoded, while maintaining the performance of critical downstream tasks. We validated our approach on the NLST lung cancer screening dataset, demonstrating that the debiased embeddings effectively eliminate multiple encoded demographic information and improve fairness without compromising predictive accuracy for lung cancer risk at 1-year and 2-year intervals. Additionally, our approach ensures the embeddings are robust against adversarial bias attacks. These results highlight the potential of adversarial debiasing techniques to ensure fairness and equity in clinical applications of self-supervised 3D CT embeddings, paving the way for their broader adoption in unbiased medical decision-making.

Figures (6)

• Figure 1: Conceptual illustration of our framework. The original 3D CT Foundation Model embedding encodes demographic information, which can lead to bias in downstream tasks. Our VAE can debias multiple demographic information while preserving downstream task performance.
• Figure 2: (a) ROC curve of linear classifier performance on sex prediction comparing original embedding (blue) vs. embedding after VAE reconstruction (red). (b) Scatter plot of linear classifier performance on age prediction comparing original embedding (blue) vs. embedding after VAE reconstruction (red).
• Figure 3: ROC curve of linear classifier performance on (a) cancer in 1 year, (b) cancer in 2 years prediction comparing original embedding (blue) vs. embedding after VAE reconstruction (red)
• Figure 4: EOD comparison for original embedding (blue) vs. VAE debiased embedding (red) for cancer in 1 year prediction when the (a) male, (b) female patient's cancer in 1 year label is poisoned by X percentage (X = 0, 25, 50, 75, 100).
• Figure 5: EOD comparison for original embedding (blue) vs. VAE debiased embedding (red) for cancer in 2 years prediction when the (a) male, (b) female patient's cancer in 2 years label is poisoned by X percentage (X = 0, 25, 50, 75, 100).