SIR 2024
General IR
Jacqueline Brenner, Ms. (she/her/hers)
MD/MPH Student Class of 2026
Miller School of Medicine
Financial relationships: Full list of relationships is listed on the CME information page.
Lindsey Hazen, Clinical Nurse
Clinical Research Nurse
National Institutes of Health
Disclosure information not submitted.
Hannah Huth, BA
Research Fellow
Center for Interventional Oncology, Radiology, and Imaging Sciences, NIH Clinical Center, NIH, Bethesda, MD, USA
Financial relationships: Full list of relationships is listed on the CME information page.
Sheng Xu, PhD
Staff Scientist
National Institutes of Health Clinical Center
Financial relationships: Full list of relationships is listed on the CME information page.
Ifechi Ukeh, MD
Deputy Chief, Interventional Radiology
National Institutes of Health
Financial relationships: Full list of relationships is listed on the CME information page.
Tabea Borde, MD, PhD
Clinical Research Fellow
National Institutes of Health
Financial relationships: Full list of relationships is listed on the CME information page.
Laetitia Saccenti, MD (she/her/hers)
Research Fellow
National Institutes of Health
Financial relationships: Full list of relationships is listed on the CME information page.
Cali Lubrant, MD
IR staff
National Institutes of Health Clinical Center
Disclosure information not submitted.
Jeff Plum, BS
Clinical Information Specialist
NIH Clinical Center
Disclosure information not submitted.
Gregg Cohen, PhD
Staff Scientist/Manager PACS/RIS
NIH Clinical Center
Disclosure information not submitted.
Bradford J. Wood, MD, FSIR
Director NIH Center for IO, Chief IR
NIH
Financial relationships: Full list of relationships is listed on the CME information page.
James Anibal, BS
PhD Student, NIH, Oxford Cambridge Scholars Program
NIH Clinical Center
Disclosure information not submitted.
• Understand the Potential of Foundation Models in Interventional Radiology (IR):
o Learn about the role and potential impact of foundation models in IR.
• Evaluate the Challenges and Pathways to Data Integration in IR:
o Assess the current challenges associated with conventional Al methods in IR and explore pathways for integrating foundation models into IR practice.
• Discuss the Future of Al in Procedural Medicine:
o Explore the future challenges of deploying an IR-specific foundation model, considering physician and patient perspectives.
.
Background:
Foundation models, such as GPT-4, are capable of complex tasks that involve text and medical images, but have not been customized for procedural medicine, including image-guided minimally invasive procedures performed by interventional radiologists [1,2,3]. This study outlined the potential design of an “IR-GPT” model, contextualize challenges, and highlight applications. We introduce a simple app for collecting annotation data from live-procedures for training IR-GPT.
Clinical Findings/Procedure Details: IR-GPT data will be annotated via “procedural narration” where a physician describes the procedure in real-time, providing assessments and justifications for actions – like instructing a trainee. Data will also be collected for pre-procedure treatment decisions. The narrations, along with reports/outcomes, will be converted into instructional data for each step in the procedural sequence. GPT will generate questions, that would otherwise (hypothetically) be asked by a clinician. The narrative data is structured into the actions and outcomes of each step that the model aims to replicate. For each time point, instructional data in the form of a question (with context) will be input into the model, along with a representative summary of the outputs from preceding steps in the procedure. The IR-GPT model will output a response, which will be compared to corresponding annotations in the voice recordings. We have developed a mobile/web app for collecting procedural narration data to be paired with imaging and reports. This software will facilitate the rapid development of a large database of IR-specific multimodal data with annotations.
Conclusion and/or Teaching Points:
By prioritizing critical use cases and confronting the current limitations of clinical AI, IR-GPT aims to refine the decision-making process and diminish human error in procedural medicine. However, the path to integration presents challenges, including ensuring diversity in data, safeguarding patient privacy, and addressing biases within models.