Google Releases Open AI Model for Healthcare Developers

MedGemma 27B, according to the developers, will complement the 4B Multimodal and 27B text-only models previously presented in the MedGemma package. 4B Multimodal is a computer vision service. Earlier in the study, a board-certified physician from the United States rated 81% of chest X-ray reports prepared by 4B Multimodal as sufficiently accurate. MedGemma 27B text-only, as reported by Google, citing internal assessments, is among the “best-performing small open models in the MedQA medical knowledge and reasoning benchmark [standard for comparing performance. – Vademecum].” MedQA is a multilingual database of questions from US medical licensing exams. Google's language model scores 87.7% correct.

MedGemma is a collection of open-source AI models designed for use in various areas of medicine. The set of services is embedded in the Health AI Developer Foundations (HAI-DEF), a large project launched in November 2024 to help developers create and implement AI services in healthcare systems. Google develops many of the models embedded in HAI-DEF based on its Gemma family of lightweight AIs. Similar technology was used in the development of MedGemma 4B and MedSigLIP, which will allow new models to be adapted for work even on a mobile device. Also, preserving Gemma's non-medical capabilities allows new algorithms to effectively solve problems that combine specialized and non-specialized information, while maintaining the ability to follow instructions and work in languages ​​other than English.

MedSigLIP is described by Google as a lightweight image coder containing only 400 million parameters. The model was tuned based on several types of medical imaging: chest X-rays, patch test results, dermatological images, and fundus images. The developers identified the ability to create service-based models for classifying medical images, classifying images without specific training examples, and searching for visually or semantically similar images in large databases as three main uses of MedSigLIP.

"Because the MedGemma collection is open, its models can be downloaded, added to, and customized to meet the specific needs of developers," Google concluded.

In December 2024, Smart Engines, a Russian company specializing in AI developments, announced that it had received a patent in the United States for a technology that reduces the radiation load on a patient during computed tomography. The document was issued by the United States Patent and Trademark Office on December 17. Smart Engines employees began working on the project in 2018; an application for a patent in Russia was filed in 2021, but it is still under consideration. According to the developers, AI allows for an average reduction in radiation load by 15%.