Google’s MedGemma 1.5 & MedASR: Open AI for Medical Imaging and Dictation

Why It Matters

Open access to state‑of‑the‑art imaging and dictation models accelerates AI integration in healthcare, enabling faster diagnostics and cost‑effective workflow automation.

Key Takeaways

• Google releases MedGemma 1.5, a 4B multimodal medical model.
• Supports CT, MRI, longitudinal X-rays, and whole‑slide pathology images.
• Improves accuracy on text, lab reports, and 2D imaging tasks.
• Introduces MedASR, a speech‑to‑text model for clinical dictation.
• Models are open‑source, free for research and commercial use.

Summary

Google unveiled two new open‑source AI models aimed at accelerating medical imaging analysis and clinical documentation, expanding its MedGemma family with version 1.5 and launching MedASR for speech‑to‑text conversion.

MedGemma 1.5 is a 4‑billion‑parameter multimodal model trained on the MedMA dataset. It can ingest high‑dimensional modalities such as CT, MRI, longitudinal chest X‑rays, and whole‑slide histopathology, delivering higher accuracy on text, lab reports, and 2‑D images compared with its predecessor. MedASR is a fine‑tuned automatic speech recognition model optimized for physician dictation, reducing transcription errors and workflow friction.

The announcement highlighted that both models are freely available on Hugging Face and can be scaled through Vertex AI, allowing developers to embed them directly into electronic health record systems or research pipelines. Google emphasized the offline capability of MedGemma 1.5, a critical feature for hospitals with limited connectivity.

By open‑sourcing high‑performance, multimodal and voice AI, Google lowers the barrier for innovators to build diagnostic tools, decision‑support apps, and automated documentation, potentially speeding up adoption of AI‑driven care and creating new revenue streams for health‑tech firms.

Original Description

Google released MedGemma 1.5 and MedASR — open medical AI models that analyze CT/MRI scans and convert clinical speech into text, even offline.