AI Models for Medicine: Google Releases MedGemma 1.5 and MedASR

Google's research division, Google Research, has released MedGemma 1.5, the latest version of its AI model specialized in medical understanding and imaging. The predecessor to the multimodal, large language model was released in May 2025 as one of the first open-source models specifically trained on medical texts and images.

Version 1.5 has been optimized for higher performance in various areas and features significantly expanded capabilities in image processing. While the first version could already process two-dimensional medical images, MedGemma 1.5 can now also analyze three-dimensional imaging and temporal progressions.

Versatile Diagnostic Capabilities

The AI's strengths include the assessment and interpretation of radiological images such as computed tomography (CT) or magnetic resonance imaging (MRT). For lung X-rays, the model can also compare current findings with previous scans to describe disease progression and healing processes over time. However, MedGemma 1.5's capabilities extend far beyond radiology. The model also interprets histopathological tissue samples, dermatological findings, or retinal images.

In internal Google tests, MedGemma 1.5 shows numerous improvements over the previous version. This includes, among other things, better classification of CT findings (accuracy of 65 percent compared to 51 percent), better general image interpretation (accuracy of 62 percent compared to 59 percent), and improved extraction of data from laboratory reports (F1-score of 78 percent compared to 70 percent). Despite these versatile capabilities, the new model version has a manageable size of 4 billion parameters. This allows for local operation on computers within a hospital or doctor's office (“On-Premise”). This aspect is crucial for practical application, as sensitive health data does not necessarily have to be processed via a cloud or the internet.

MedASR for Medical Speech Recognition

In addition to MedGemma 1.5, Google Research has also released a second AI model called MedASR. MedASR is optimized for medical speech recognition and can be used for dictating medical reports or recording case discussions. MedASR can be directly combined with MedGemma, thus serving as a complete system for voice recordings in addition to medical texts and images.

With the open-source release of its medical AI models, Google is banking on their use and further development by developers. Through further AI training on their data (finetuning), the performance of the models can be further improved for specific applications, such as use in a particular medical specialty. Google provides extensive tutorials and tools with the release. In addition, $100,000 in prizes will be awarded for the “MedGemma Impact Challenge,” where developer teams will use the models for innovative AI applications in the healthcare system.

The release of MedGemma 1.5 and MedASR is of great importance for the integration of AI systems in medicine. The model has been used worldwide since its initial release. For example, Taiwan's national health administration used MedGemma to systematically assess the health status of lung cancer patients before surgery.

(mki)