Google Releases MedGemma 1.5 and MedASR Open Models for Medical AI

Details

• Google AI Developers announced MedGemma 1.5, an updated open model from Google Research, small enough to run offline with improved performance on medical imaging, text, medical records, and 2D images.
• MedGemma 1.5 4B supports high-dimensional CT, longitudinal chest X-ray interpretation, medical lab report extraction, and expands to 3D CT/MRI volumes and histopathology imaging.
• Internal benchmarks show improvements: 3% accuracy gain on CT disease classification (61% vs 58%), 14% on MRI (65% vs 51%), 35% on chest X-ray anatomical localization, and 18% on lab report extraction.
• MedASR, a new open medical speech-to-text model, transcribes healthcare vocabulary and generates prompts for MedGemma, reducing word error rates to under 5% on specialized jargon.
• Models available on Vertex AI and Hugging Face; built on Gemma 3 with fine-tuning on datasets like MIMIC-CXR, achieving 92% image captioning accuracy.
• Real-world use includes Qmed Asia's askCPG for Malaysian clinical guidelines, enhancing decision support with multimodal features.

Impact

Google's MedGemma 1.5 and MedASR releases strengthen its position in healthcare AI, where adoption is accelerating twice as fast as the broader economy, by providing open, fine-tunable models that outperform prior versions on key benchmarks like MedQA and EHRQA. This pressures rivals like Microsoft's Nuance and IBM Watson Health, as the open-source approach via Hugging Face and Vertex AI lowers entry barriers, enabling startups and providers to build custom diagnostic tools amid a $45 billion medical imaging market projected by 2028. With HIPAA-compliant deployments and bias-reduced training on diverse datasets, these models address regulatory hurdles stalling 40% of AI adoption, while multimodal capabilities for text, images, and speech align with trends in agentic systems and on-device inference. Over the next 12-24 months, expect accelerated R&D in telemedicine and predictive analytics, potentially handling 70% of routine diagnostics by 2030 and shifting funding toward scalable, privacy-preserving healthcare AI.