Speaker: Abdurrahim Yılmaz

Abstract: Vision-language models (VLMs) have demonstrated strong potential for integrating visual and textual information, enabling unified multimodal reasoning in medical applications. More recently, the emergence of agentic AI systems, capable of planning, reasoning, and interacting with tools, has further expanded the scope of intelligent clinical decision support. However, deploying such systems in real-world healthcare settings remains challenging due to constraints related to computational resources, data availability, and patient privacy. In this talk, I will present a general framework for developing resource-efficient and privacy-preserving medical VLMs, and discuss how agentic capabilities can be incorporated into these systems to enhance adaptability and usability. The approach emphasizes modular design, the use of smaller models, and data-centric strategies to enable scalable, interpretable, and locally deployable solutions. I will highlight key considerations for integrating agentic AI with multimodal models in clinical contexts and outline broader implications for building accessible and trustworthy medical AI systems.

Bio: Abdurrahim Yılmaz is a PhD candidate in Artificial Intelligence at the Faculty of Medicine, Imperial College London. His research focuses on biomedical AI, spanning the full pipeline from data collection/preprocessing and model development to benchmark design and clinician-centered evaluation. His work covers multiple clinical domains with a particular emphasis on dermatological AI. He is also the co-founder of a digital health startup and is actively engaged in translating academic research into real-world, people-centered healthcare solutions. Prior to his work in AI, he has a background in mechatronics and microrobotics.

https://us06web.zoom.us/j/84636366540
MEETING ID: 84636366540
PASSCODE: 25862