IR-GPT: AI Foundation Models to Optimize Interventional Radiology.

Area of Science:

• Artificial Intelligence in Medicine
• Interventional Radiology

Background:

• Foundation AI models excel at diverse data tasks but lack specialization in procedural medicine.
• Current deep learning in interventional radiology (IR) faces generalization and scalability challenges.

Purpose of the Study:

• To review existing deep learning work in IR.
• To identify barriers hindering AI generalization and large-scale deployment in IR.
• To propose a design for an "IR-GPT" foundation model for IR.

Main Methods:

• Literature review of deep learning applications in interventional radiology.
• Analysis of challenges in AI data collection, annotation, and training for IR.
• Conceptual design of a unified AI platform for IR.

Main Results:

• Identified key barriers to AI generalization and deployment in IR.
• Outlined a framework for an "IR-GPT" foundation model.
• Highlighted potential downstream applications of AI in IR.

Conclusions:

• A specialized foundation model, "IR-GPT," can unify AI efforts in interventional radiology.
• Addressing data and training challenges is crucial for successful AI deployment in IR.
• AI holds significant potential to advance interventional radiology practices.