Healthcare is more and more embracing AI to enhance workflow administration, affected person communication, and diagnostic and therapy assist. It’s crucial that these AI-based methods will not be solely high-performing, but in addition environment friendly and privacy-preserving. It’s with these issues in thoughts that we constructed and just lately launched Well being AI Developer Foundations (HAI-DEF). HAI-DEF is a set of light-weight open fashions designed to supply builders sturdy beginning factors for their very own well being analysis and software improvement. As a result of HAI-DEF fashions are open, builders retain full management over privateness, infrastructure and modifications to the fashions. In Might of this 12 months, we expanded the HAI-DEF assortment with MedGemma, a set of generative fashions based mostly on Gemma 3 which are designed to speed up healthcare and lifesciences AI improvement.
At present, we’re proud to announce two new fashions on this assortment. The primary is MedGemma 27B Multimodal, which enhances the previously-released 4B Multimodal and 27B text-only fashions by including assist for complicated multimodal and longitudinal digital well being report interpretation. The second new mannequin is MedSigLIP, a light-weight picture and textual content encoder for classification, search, and associated duties. MedSigLIP is predicated on the identical picture encoder that powers the 4B and 27B MedGemma fashions.
MedGemma and MedSigLIP are robust beginning factors for medical analysis and product improvement. MedGemma is helpful for medical textual content or imaging duties that require producing free textual content, like report era or visible query answering. MedSigLIP is advisable for imaging duties that contain structured outputs like classification or retrieval. All the above fashions could be run on a single GPU, and MedGemma 4B and MedSigLIP may even be tailored to run on cellular {hardware}.
Full particulars of MedGemma and MedSigLIP improvement and analysis could be discovered within the MedGemma technical report.
