Google Updates Open Medical AI Suite With New Imaging and Speech Models, Launches $100K Developer Hackathon
Google has released updates to its open-source medical AI suite with MedGemma 1.5 4B and a new speech-to-text model, MedASR, and introduced the MedGemma Impact Challenge, a hackathon hosted on Kaggle offering $100,000 in prizes. The new models are part of Google’s Health AI Developer Foundations (HAI-DEF) program and are available for free through Hugging Face and Google Cloud’s Vertex AI. Both models are designed as starting points for developers building medical applications in imaging, clinical documentation, and multimodal data analysis.
Summary of the MedGemma collection of models and their capabilities.
Image credit: Google
All HAI-DEF models are provided for research and development purposes and require adaptation before use in clinical workflows. Tutorials and code examples are available via Google’s GitHub and model hosting platforms.
Developers and health tech organizations are already applying MedGemma across a range of clinical and research contexts. In Malaysia, Qmed Asia integrated the model into askCPG, a conversational tool for navigating over 150 national clinical practice guidelines. In Taiwan, the National Health Insurance Administration used MedGemma to extract structured insights from more than 30,000 pathology reports related to lung cancer surgeries.
Since its launch, MedGemma has also been referenced widely in medical AI literature as a base model for tasks involving medical text, support for multidisciplinary clinical decisions, mammography analysis, and other healthcare applications.
MedGemma 1.5 Expands Medical Imaging Capabilities
MedGemma 1.5 adds support for high-dimensional imaging, including 3D modalities such as CT and MRI scans, and whole-slide histopathology. It also improves handling of time-series X-rays and anatomical feature localization in chest X-rays. Performance benchmarks show a 14% improvement in MRI disease classification and a 35% gain in chest X-ray anatomical localization compared to the previous version.
“MedGemma 1.5 4B improves support for medical imaging, exceeding the performance of MedGemma 1 4B on high-dimensional image interpretation, anatomy localization and longitudinal disease assessment in chest X-rays, general medical image interpretation and extraction of content from medical laboratory reports.”
Image credit: Google
MedGemma 1.5 is released in a compact 4B parameter version for offline and low-resource use, while the larger 27B version remains available for more complex tasks. Alongside imaging, MedGemma also shows improved text understanding, with 5% higher accuracy on medical QA benchmarks and 22% on EHR-based question answering.
MedASR Introduced for Medical Speech Recognition
The newly introduced MedASR model transcribes medical dictation and spoken prompts for clinical tasks. It was benchmarked against Whisper large-v3 and reported 58% fewer errors on chest X-ray dictations and 82% fewer on a general medical speech benchmark. MedASR is optimized for medical vocabulary and is designed to integrate with MedGemma in speech-based workflows.
Image credit: Google
The model is available for use and fine-tuning, along with tutorials that help developers build their own systems combining MedASR with other clinical AI tools. Further details can be found in the MedASR model card.
Topic: Tech Giants
