Running summarizing tools on a new article
Check the paper below, then start its synthesis.
An open vision-language model for diverse medical applications
Andrew Sellergren (Google (United States)), Sahar Kazemzadeh (Google (United States)), Fereshteh Mahvar (Google (United States)), Atilla P. Kiraly (Google (United States)), Madeleine Traverse (Google (United States)), Timo Kohlberger (Google (United States)), Chufan Gao (Google (United States)), Ronnachai Jaroensri (Google (United States)), Shawn Xu (Google (United States)), Fayaz Jamil (Google (United States)), Cian Hughes (Google (United States)), Fereshteh Mahvar, Liron Yatziv (Google (United States)), Liron Yatziv (Google (United States)), Tiffany J. Chen (Google (United States)), Bram Sterling (Google (United States)), Sathaiah Baby (Google (United States)), Susanna Maria Baby (Google (United States)), Jeremy Lai (Google (United States)), Samuel Schmidgall (Google (United States)), Lu Yang (Google (United States)), Kejia Chen (Google (United States)), Per G. Björnsson (Google (United States)), Shashir Reddy (Google (United States)), Ryan Brush (Google (United States)), Kenneth Philbrick (Google (United States)), Bashir Sadjad (Google (United States)), Mercy Asiedu (Google (United States)), Ines Mezerreg (Google (United States)), Howard Hu (Google (United States)), Howard Yang (Google (United States)), Richa Tiwari (Google (United States)), Sunny Jansen (Google (United States)), Preeti Singh (Google (United States)), Aishwarya Kamath (Google (United States)), Johan Ferret (Google (United States)), Catherine Kozlowski, Rachelle Sico (Google (United States)), Aishwarya Kamath (Google (United States)), Johan Ferret (Google (United States)), Shreya Pathak (Google (United States)), Nino Vieillard (Google (United States)), Ramona Merhej (Google (United States)), Sarah Perrin (Google (United States)), Tatiana Matejovicova (Google (United States)), Alexandre Ramé (Google (United States)), Morgane Rivière (Google (United States)), Louis Rouillard (Google (United States)), Thomas Mesnard (Google (United States)), Geoffrey Cideron (Google (United States)), Jean-bastien Grill (Google (United States)), Jean-Baptiste Alayrac (Google (United States)), Edouard Yvinec (Google (United States)), Michelle Casbon (Google (United States)), Elena Buchatskaya (Google (United States)), Jean-Baptiste Alayrac (Google (United States)), Dmitry Lepikhin (Google (United States)), Vlad Feinberg (Google (United States)), Sebastian Borgeaud (Google (United States)), Alek Andreev (Google (United States)), Cassidy Hardin (Google (United States)), Robert Dadashi (Google (United States)), Léonard Hussenot (Google (United States)), Armand Joulin (Google (United States)), Olivier Bachem (Google (United States)), Clément Farabet (Google (United States)), Joelle K. Barral (Google (United States)), Tris Warkentin (Google (United States)), Jonathon Shlens (Google (United States)), David J. Fleet (Google (United States)), Victor Cotruta (Google (United States)), Victor Cotruta (Google (United States)), Omar Sanseviero (Google (United States)), Gus Martins (Google (United States)), Phoebe Kirk (Google (United States)), Dr. Anand Rao (Google (United States)), Yossi Matias (Google (United States)), Katherine Chou (Google (United States)), Avinatan Hassidim (Google (United States)), Dale R. Webster (Google (United States)), Shravya Shetty (Google (United States)), Kapil Goel (Google (United States)), Yun Liu (Google (United States)), Can Kirmizibayrak (Google (United States)), David F. Steiner (Google (United States)), Rory Pilgrim (Google (United States)), Daniel Golden (Google (United States)), Lin Yang (Google (United States))
Abstract: Artificial intelligence has high potential for impact in healthcare applications, but its training and deployment are challenging due to diverse data, a complex spectrum of possible tasks and important privacy needs. High-performing foundation models that enable data-efficient fine-tuning for diverse downstream tasks can meaningfully accelerate development in this domain. We introduce MedGemma, a collection of medical vision-language foundation models based on Gemma 3. MedGemma demonstrates advanced medical understanding and reasoning across images and text and multiple medical imaging domains, exceeding the performance of similarly sized generative models while maintaining the general capabilities of the Gemma base models. For out-of-distribution tasks, MedGemma achieves improvements of 2.6-10% in medical image question answering, 15.5-18.1% in chest X-ray finding classification and 10.8% in agentic evaluations compared with the base models. Our results show that fine-tuning MedGemma can be more effective than fine-tuning the base Gemma 3 model for medical tasks, particularly in the setting of limited training data. We additionally introduce MedSigLIP, a medically tuned vision encoder derived from SigLIP. MedSigLIP powers the visual understanding capabilities of MedGemma and, as an encoder, achieves performance comparable to or better than that of many specialized medical image encoders. Taken together, the MedGemma collection provides a strong foundation of medical image and text capabilities, with the potential to accelerate medical research and the development of downstream applications.