G-TERM-AI

Fraunhofer HHI, in collaboration with Fraunhofer USA, Inc. – Center Mid-Atlantic, is launching the G-TERM-AI project (“Generative Testbed for Enhancing Device- and Protocol-Agnostic Reliability of Multimodal Medical AI”). Running from July 2025 to January 2027, this initiative addresses one of the most pressing challenges in healthcare AI: ensuring robust and reliable performance of AI-enabled medical devices across the diversity encountered in real-world clinical environments.

The adoption of artificial intelligence in medical products is accelerating, but regulatory approval remains a significant hurdle. AI models must demonstrate consistent results across a range of variables, from patient demographics and physical characteristics to different imaging devices and clinical protocols. However, collecting the comprehensive and diverse datasets needed to train and validate such models is hindered by high costs, strict privacy regulations, and limited access to varied patient samples. Harmonising datasets from different sources and devices further complicates this challenge, and existing solutions often fall short of ensuring true generalisability and reliability.

G-TERM-AI aims to overcome these limitations by developing a generative AI-based framework for continuous evaluation and improvement of medical AI models. The project will focus on creating methods to generate realistic and diverse synthetic datasets that reflect variations in devices, lab protocols, and patient characteristics. Additionally, it will establish a pipeline for evaluating and refining both single- and multimodal AI models—those that integrate multiple types of medical data, such as imaging and textual reports. Initial efforts will target two application areas: histopathology datasets and AI models with protocol variations (e.g., breast, dermatological, or gastrointestinal analysis), and radiology datasets with images from multiple device types (e.g., X-ray, CT scans).

The expected outcome is a proof-of-concept generative framework that produces high-quality synthetic datasets, supporting faster and more robust evaluation and retraining of medical AI models. This approach offers clear advantages over traditional methods by addressing data scarcity, reducing development costs, and improving model performance and generalisability. Ultimately, G-TERM-AI lays the groundwork for a scalable, cost-effective solution for developing medical AI models that are ready for broader regulatory approval and real-world deployment.

Fraunhofer HHI & Fraunhofer USA, Inc. – Center Mid-Atlantic