Improved performance for medical imaging use cases
MedGemma was designed from the ground up as a multimodal model, reflecting the multimodal nature of medicine. MedGemma 1 included support for the interpretation of two-dimensional medical images, including chest x-rays, dermatology images, fundus images, and histopathology patches.
With MedGemma 1.5, we expand support for high-dimensional medical imaging, starting with three-dimensional volume representations of CT Imaging And MRIas well as the entire slide histopathological imaging. Developers can create applications in which multiple slices (for CT or MRI) or multiple patches (for histopathology) are provided as input, along with a prompt describing the task.
On internal criteria, the absolute baseline accuracy of MedGemma 1.5 improved by 3% compared to MedGemma 1 (61% vs. 58%) in classifying disease-related CT findings and by 14% (65% vs. 51%) in classifying disease-related MRI findings, averaged across results. Additionally, based on a diverse internal benchmark of histopathology slides and associated results, the fidelity of MedGemma 1.5 predictions, based on RED-L score on cases with exactly one histopathology slide, improved by 0.47 compared to MedGemma 1 (0.49 vs. 0.02), corresponding to the score of 0.498 obtained by the task-specific test PolyPath model.
This new high-dimensional support is the natural evolution of CT Foundationour previous API-based tool for generating CT integrations. To our knowledge, MedGemma 1.5 is the first public release of a large open multimodal language model capable of interpreting high-dimensional medical data while retaining the ability to interpret general 2D data and text. Although these features are in their early stages and remain imperfect, developers will achieve better results by refining MedGemma models on their own data, and we hope to continually improve MedGemma models over time. We have published tutorial workbooks that illustrate how to use this high-dimensional imaging capability for CT (Cuddly face, Model garden) and histopathology (Cuddly face, Model garden).