St. Jude Family of Websites
Explore our cutting edge research, world-class patient care, career opportunities and more.
St. Jude Children's Research Hospital Home
St. Jude Family of Websites
Explore our cutting edge research, world-class patient care, career opportunities and more.
St. Jude Children's Research Hospital Home
Building and using trustworthy AI for medical imaging and multimodal data to advance discovery and care.
AI in medicine is advancing faster than our ability to decipher its trustworthiness and reliability. A model that performs well in a paper can still fail in real world clinical settings, and there is not a consensus in how to navigate this issue. The Yi lab works on both sides of that problem. First, we build methods that test medical AI’s trustworthiness, followed by building methods that make AI better: stress-testing AI robustness, measuring algorithmic fairness and studying clinician behavior amidst algorithmic advice. Second, we collaborate with scientists and clinicians to turn vetted AI tools into discoveries in pediatric cancers, sickle cell disease and other catastrophic diseases.
The Yi lab works in biomedical informatics. Our program spans two areas: 1) developing methods for safe and trustworthy AI for medical imaging and multimodal data and 2) applying AI tools to advance discovery and care in pediatric cancer and other catastrophic diseases, in collaboration with clinicians and scientists across the institution.
Benchmark accuracy is not evidence that an AI model is appropriate to use on patients. When a model provides inaccurate information pertaining to imaging, the consequence is not a lower score; it is a treatment decision made on bad information. We build methods that determine whether medical AI can be trusted and identify ways to make AI more accurate and efficient. The technologies may change, but the core questions remain the same. Our current work spans into the following areas:
Robustness and generalizability. Identifying when and how AI models break in real world clinical settings, e.g., different scanners, different protocols, and patients who do not resemble the training data.
Fairness and bias. Whether AI performs similarly for all patients. We develop the definitions, metrics and reporting standards the field currently lacks; in our audit of public medical imaging datasets, <5% reported patient age, race or ethnicity.
Trust in generative AI. Whether generative models can be trusted to counsel patients and support doctors and researchers and if a deepfake medical image can be differentiated from a real one.
Human–AI interaction. Designing AI that helps clinicians rather than hinders them. In our prospective study of 220 physicians, the AI explanations doctors accepted fastest were also the ones that led them to follow incorrect advice.
Children treated at St. Jude are imaged repeatedly over years, and these images are augmented by a rich record of labs, genomics and clinical notes. We collaborate with scientists and clinicians across the hospital to transform this information into research-ready data: quantitative phenotypes from imaging, structured variables pulled from clinical text and the endpoints and biomarkers that drive discovery research.
Pediatric cancer
Imaging biomarkers of treatment response and disease burden.
Survivorship
Quantitative markers of frailty and chronic disease in long-term childhood cancer survivor cohorts.
Sickle cell disease
Quantitative imaging of organ involvement and disease progression.
A discovery is only as good as the measurement behind it. Therefore, we make medical AI trustworthy first and then use it to answer questions we could not answer before.
Our research spans medicine and computer science, and both audiences need to see it to bridge the gap between code and clinic. Our work has been published in Radiology and Radiology: Artificial Intelligence (the flagship journals of the Radiological Society of North America), Nature Medicine and JAMA, as well as top computer science venues including NeurIPS and ICML.
Our lab is built from three kinds of people: clinical, technical and the ones who are becoming both. Medical students, residents and fellows work alongside research engineers and computer science trainees, and we collaborate with clinicians and scientists across medicine, basic sciences and computer sciences. Some members arrive knowing how to train an AI model and learn what a scan means; others arrive the other way around. The best work happens where those two paths meet.
Projects often pair a clinical trainee with a technical one: clinical context on one side and implementation on the other. New trainees usually start by contributing to an existing project alongside one of our engineers or postdocs, taking on more independence as they progress.
Technical work includes comfortability with Python, a deep learning framework, Git and the patience for complicated real-world data. Clinical work must include familiarity with clinical data, expertise in conceptualizing relevant biological or medical questions, literature searches and manuscript writing.
Email Dr. Yi at Paul.Yi@stjude.org with:
Paul Yi, MD
Associate Member
Department of Radiology
MS 220
St. Jude Children's Research Hospital
Follow Us