Race and AI in Radiology

Dr. Gichoya had started by tackling the original problem of bias in diagnoses for chest X-rays, since it has always been difficult to tell whether something is a real diagnosis, or simply just a finding. Her team built a deep learning model; however, they saw that it did not work well for black patients. With further investigation, they discovered that their model had learned signals that correlated with self-identified race.

Intrigued by this finding, Dr. Gichoya and her team sought to identify the factors that the model used when making its race determination. Because AI is black box in nature, the methods by which the algorithm learns remains largely unknown. When tested in other imaging modalities (mammogram, chest CT, spine imaging), the model still showed high accuracy. Additionally, the model retained accuracy when different information was eliminated from the images (ex. age, disease distributions, bone densities). The model was also able to predict race in healthy patients, showing that it did not rely on patterns of disease prevalence in specific ethnic groups.

Next, we spoke about the implications of this research in developing risk scores. Deep learning models are able to look at factors that humans are not trained or able to see. Dr. Gichoya highlights the model’s potential effectiveness in predicting osteoarthritis risk in black patients. We also look at applications in opportunistic screening and information about social determinants of health. For example, most patients presenting with chest pain often get chest CTs. Dr. Gichoya thinks that these images can be used by the model to learn about patients’ environmental exposures, like pollution.

We finish the episode with a discussion on the changing landscape of IR and how AI can be used as an assistive technology. Interventional cardiologists are already using AI to dictate their procedural reports in real-time. In the interventional oncology space, AI could help integrate imaging and pathology findings to determine personalized treatment courses. All of these applications depend on researchers’ ability to market their findings to peers and the public, Dr. Gichoya gives tips on how to do this.