After OpenAI and Anthropic, Google enters Medical AI race with MedGemma 1.5 and MedASR

New Delhi: Google has stepped up the race in medical AI with the launch of MedGemma 1.5 and a new speech-to-text system called MedASR. The company says the new models are designed to help developers build tools that can understand medical images, clinical documents, and spoken dictation. The move comes as big tech firms push deeper into healthcare, following similar efforts by OpenAI and Anthropic.

MedGemma is part of Google’s Health AI Developer Foundations (HAI-DEF) programme, which offers open models for healthcare research and product development. Since its first release last year, MedGemma has seen millions of downloads and hundreds of community-built versions on Hugging Face. With version 1.5, Google is expanding what the model can do, especially in complex medical imaging and clinical text analysis.

The biggest upgrade in MedGemma 1.5 is its ability to understand high-dimensional medical images. This includes CT scans, MRI volumes, and full histopathology slides. Developers can now feed the model multiple slices or image patches and ask it to analyse disease patterns, anatomical structures, or changes over time.

Google says internal tests show strong gains over the previous version. The model improved disease detection accuracy on CT scans and MRI images, and it also showed much better performance in reading pathology slides. These upgrades allow MedGemma 1.5 to move beyond simple 2D images like X-rays and into more advanced radiology and lab workflows.

MedGemma 1.5 is also more effective in long-run data, i.e., comparing the chest X-rays of a week or months. This will aid in monitoring disease progress or recovery. The model is now also able to identify organs and structures in X-rays much more accurately, and this is useful in clinical and research tools.

To top this, it is more effective in reading laboratory reports. It has the ability to harvest the key values, units and test names of unstructured documents. It is easier to convert scanned or typed reports to useful medical data.

Google has also enhanced the capability of MedGemma in dealing with clinical text and health records. The new version works more effectively on medical question-answering and queries based on electronic health records. This implies that it is able to access patient information, describe medical terms, and assist clinical reasoning more accurately than ever.

According to the company, these gains have been brought about by new training data and new training techniques. This is aimed at ensuring that MedGemma becomes more trusted in the healthcare-orientated language tasks and not in the image analysis.

In addition to MedGemma 1.5, Google has published MedASR, a healthcare speech-to-text model. It is taught medical vocabulary, accents and dictation styles. This enables it to transcribe doctor notes, radiology reports and clinical conversations with fewer errors as compared to general-purpose speech models.

MedASR in tests in Google had significantly fewer errors than Whisper, a popular speech-to-text system. One can also speak prompts into MedGemma, which makes hands-free AI medical device development easier.

MedGemma 1.5 and MedASR can be used on Hugging Face and Google Vertex AI, both as research and as a commercial offering. The smaller 4B model can be run locally by the developer or scaled on Google Cloud to bigger applications. Full DICOM support has also been added to Google, and this is important in managing medical image files.

Add TV9 English As A Trusted Source

To drive acceptance, Google has introduced the MedGemma Impact Challenge, a Kaggle hackathon, and a prize of $100,000. It is aimed at promoting the construction of real-world healthcare tools based on these models by developers.

Google is strongly positioning itself as a big player in medical AI with MedGemma 1.5 and MedASR. The company is placing a bet that open, multimodal models will take the next generation of healthcare software.