MedGemma 1.5 Technical Report

A large collaborative team from Google and DeepMind has released MedGemma 1.5, a significant upgrade to their open-source medical AI model. The 4-billion parameter model expands beyond its predecessor by integrating high-dimensional medical imaging, including 3D CT and MRI volumes, and histopathology whole slide images. It also introduces anatomical localization via bounding boxes and multi-timepoint chest X-ray analysis, alongside improved understanding of complex medical documents like electronic health records (EHRs) and lab reports. The technical report details the architectural innovations required, such as long-context 3D volume slicing and specialized sampling for whole-slide images, to unify these diverse modalities within a single, efficient model.

Performance gains are substantial across the board. Compared to MedGemma 1, the new version improves 3D MRI condition classification accuracy by 11% and 3D CT classification by 3%. In the critical area of whole-slide pathology imaging, it achieves a dramatic 47% macro F1 score gain. For clinical text, it shows a 5% improvement on the MedQA benchmark and a 22% boost on EHRQA, demonstrating enhanced medical reasoning. The model also excels at information extraction from lab reports, scoring an average 18% macro F1 across four different datasets.

Designed as a robust, open resource, MedGemma 1.5 provides a powerful foundation model for the developer community. Its multimodal capabilities in imaging and text analysis position it to accelerate the creation of the next generation of AI-assisted diagnostic tools, clinical decision support systems, and automated medical data processing pipelines. The release includes resources and tutorials to help developers build upon this advanced platform.