Introducing MedGemma: Advanced AI Models for Healthcare

As the healthcare industry increasingly adopts AI to streamline workflow management, enhance patient communication, and support diagnostics, it is crucial that these AI systems maintain high performance while prioritizing efficiency and privacy. In response to these needs, we are thrilled to unveil Health AI Developer Foundations (HAI-DEF), a collection of lightweight open models designed to empower developers to create robust health research tools and applications.

This year, we expanded our offering with the medGemma collection, featuring advanced generative models built on Gemma 3, aimed at accelerating advancements in healthcare and life sciences AI. Today, we are excited to announce two new models within this collection: MedGemma 27B Multimodal and MedSigLIP.

Overview of New MedGemma Models

The MedGemma 27B Multimodal model enhances the capabilities of the previously released 4B Multimodal and 27B text-only models by providing support for complex multimodal and longitudinal electronic health record interpretation. Meanwhile, MedSigLIP serves as a lightweight image and text encoder for classification and search tasks, utilizing the same robust image encoding architecture present in the original MedGemma models.

Capabilities and Applications

These models open up new possibilities for medical research and product development:

• MedGemma: Suited for tasks requiring free-text generation such as report writing and visual question answering.
• MedSigLIP: Ideal for imaging tasks that involve structured outputs like classification or retrieval.

Both models are designed to run efficiently on a single GPU, with the 4B and MedSigLIP variants even adaptable for mobile hardware.

Performance Metrics

Based on evaluations, MedGemma 4B has demonstrated impressive capabilities with a score of 64.4% on the MedQA benchmark, and 81% of generated chest X-ray reports were assessed as sufficiently accurate by a board-certified radiologist. The 27B models excel on various benchmarks, achieving high performance while remaining cost-effective compared to larger models.

Developers’ Experience and Flexibility

The open nature of the MedGemma collection allows developers to download and build upon these models, customizing them for their specific healthcare applications. This environment encourages flexibility, as models can be run locally or on different infrastructures, elevating privacy and control.

Real-World Applications

The usability of MedGemma and MedSigLIP has garnered interest from various developers in the healthcare sector. Notably, teams at DeepHealth in the USA are using MedSigLIP to enhance chest X-ray triaging, while Chang Gung Memorial Hospital in Taiwan found the MedGemma models adept at handling traditional Chinese-language medical literature.

Getting Started

For developers prepared to explore these models, we have provided detailed resources on GitHub for instance creation and fine-tuning. When scaling up, the MedGemma and MedSigLIP models can seamlessly integrate within Vertex AI.

We invite you to check out the official documentation and explore the potential of MedGemma and MedSigLIP in your healthcare applications. We look forward to seeing the innovative solutions you develop using these advanced AI tools.