Google has released MedGemma 1.5 4B, an updated open-weight medical AI model that extends its predecessor's capabilities to include 3D imaging volumes, whole-slide pathology analysis, and improved clinical document understanding, according to a technical report published on arXiv.

Medical AI models have historically struggled to operate across the full range of data types clinicians encounter — from radiology scans and pathology slides to lab reports and electronic health records. MedGemma 1.5 is Google's attempt to bring these modalities under a single model architecture, building on the original MedGemma 1 4B released earlier. The technical report details the training innovations required to make this work, including long-context 3D volume slicing and whole-slide pathology sampling techniques.

What MedGemma 1.5 Can Now Do

The most significant additions in this version are support for high-dimensional imaging modalities — specifically CT and MRI volumes — along with histopathology whole slide images (WSIs). These are among the most data-intensive inputs in clinical medicine, and handling them within a single compact 4-billion parameter model requires architectural choices that the report describes in detail.

The model also adds anatomical localization via bounding boxes, allowing it to indicate where in an image a finding is located rather than simply classifying what the image contains. Multi-timepoint chest X-ray analysis — comparing scans taken at different points in time — is another new capability, which is relevant for monitoring disease progression.

MedGemma 1.5 4B achieves a 47% macro F1 gain in whole-slide pathology imaging compared to its predecessor — a substantial leap for a model of this size.

On text-based tasks, the model improves 5% on MedQA accuracy and 22% on EHRQA accuracy compared to MedGemma 1, according to the report. These benchmarks test clinical question answering and electronic health record question answering respectively. All performance figures cited are self-reported by Google in the technical report and have not been independently verified.

The Numbers Behind the Claims

The reported improvements are worth examining specifically. On 3D MRI condition classification, MedGemma 1.5 4B improves by 11 percentage points in absolute terms over MedGemma 1. On 3D CT classification, the gain is 3 percentage points. The gap between these two figures likely reflects how much harder CT volumes are to interpret — they are typically larger, noisier datasets requiring different handling.

The 35% increase in Intersection over Union (IoU) on chest X-ray anatomical localization is notable. IoU measures how well a predicted bounding box overlaps with a ground-truth region, so a 35% improvement suggests the model has become substantially more precise at identifying where in an image a relevant structure or finding appears.

For lab report information extraction, the model achieves an average of 18% macro F1 across four datasets. This figure appears modest in isolation, but information extraction from unstructured clinical documents is a notoriously difficult task with high variance across document formats and clinical settings.

How Google Built It

Enabling all these modalities in one model required several technical innovations. For 3D volumes like CT and MRI scans, Google used a long-context slicing approach — processing the volume as a sequence of 2D slices rather than attempting true volumetric processing. For whole-slide pathology images, which can be several gigabytes in size, a sampling strategy selects representative regions for the model to process.

These are practical engineering approaches, not limitations unique to this model. Most current medical AI systems that handle large imaging formats use similar strategies. The report's value is in documenting how these approaches were integrated into a single unified architecture rather than separate specialist systems.

The model is built on Google's Gemma family of open-weight models, which means developers can access the weights and adapt them. Google has published resources and tutorials at its dedicated MedGemma page, positioning the release explicitly as a foundation for third-party developers rather than a finished clinical product.

Who This Is For

MedGemma 1.5 is not a clinical product cleared for diagnostic use. Google frames it as a research foundation model — a starting point that developers and researchers can fine-tune for specific applications. This distinction matters: the benchmarks in the report describe what the base model can do, not what a fine-tuned derivative might achieve in a real clinical deployment.

That said, the breadth of modalities covered in a single open model of this size is uncommon. Most open medical AI models specialise — a chest X-ray model, a pathology classifier, a clinical NLP system. Combining these under one architecture with reasonable performance across all of them gives researchers a more flexible starting point.

The model's 4-billion parameter scale is also meaningful. Larger models often outperform on benchmarks but are expensive to run and difficult to deploy in resource-constrained clinical environments. A 4B model can run on a single modern GPU, which lowers the barrier for hospital systems and research groups without large compute budgets.

What This Means

MedGemma 1.5 gives the medical AI research community a more capable open foundation model that spans imaging, pathology, and clinical text — reducing the need to build and maintain separate specialist systems, and lowering the cost of entry for groups developing the next generation of clinical AI tools.