AI TL;DR
Google released MedGemma 1.5 4B with improved CT/MRI support, longitudinal imaging, and medical document understanding. Here's how to use this open-source healthcare AI.
Google just released MedGemma 1.5 4B, an open-source medical AI model that can analyze CT scans, MRI images, and X-rays with remarkable accuracy.
Released on January 13, 2026, this update brings significant improvements over the original MedGemma—including 14% better MRI classification and support for longitudinal imaging.
Let's explore what this means for healthcare AI.
What is MedGemma?
MedGemma is part of Google's Health AI Developer Foundations (HAI-DEF) program. It's a collection of open models built on the Gemma 3 architecture, optimized specifically for:
- Medical text understanding
- Medical image interpretation
- Healthcare application development
The key word is open. Unlike many medical AI tools, MedGemma is freely available for developers to use, fine-tune, and deploy.
What's New in MedGemma 1.5 4B
High-Dimensional Medical Imaging
The biggest upgrade is support for 3D imaging:
| Modality | Capability |
|---|---|
| CT scans | 3D volume analysis, 3% accuracy improvement |
| MRI scans | 3D interpretation, 14% accuracy improvement |
| Histopathology | Whole-slide imaging (WSI) support |
| X-rays | Enhanced with localization |
Longitudinal Analysis
MedGemma 1.5 can now analyze images over time:
- Compare current chest X-ray to prior images
- Track disease progression
- Identify changes that might be missed in isolation
- Support clinical decision-making with temporal context
Anatomical Localization
New bounding box capabilities:
- Locate specific anatomical features in X-rays
- Point to abnormalities with coordinates
- Enable more precise clinical reporting
- Support training data annotation
Medical Document Understanding
Beyond imaging, MedGemma 1.5 processes:
- Lab reports (extract values and units)
- Electronic Health Records (EHR)
- Clinical notes
- Unstructured medical text
Technical Specifications
| Specification | Detail |
|---|---|
| Parameters | 4 billion |
| Context length | 128K+ tokens |
| Input types | Text, images, multi-modal |
| Architecture | Based on Gemma 3 |
| License | Open (HAI-DEF program) |
| GPU requirement | Single GPU capable |
The 4B parameter size is deliberately chosen—powerful enough for medical tasks, efficient enough for practical deployment.
Accuracy Improvements
CT Classification
MedGemma 1.5 shows 3% improvement over MedGemma 1 for classifying disease-related CT findings:
- Tumor detection
- Organ abnormalities
- Structural anomalies
- Disease staging
MRI Classification
The bigger gains are in MRI, with 14% improvement:
- Brain pathology detection
- Soft tissue analysis
- Joint assessment
- Neurological conditions
These improvements come from better training data and architectural refinements.
Use Cases
For Hospitals
- Preliminary screening - AI-assisted triage for imaging
- Second opinion - Validate radiologist interpretations
- Longitudinal tracking - Monitor chronic conditions
- Documentation - Automated report generation
For Research
- Drug trials - Analyze imaging endpoints
- Population studies - Process large imaging datasets
- Biomarker discovery - Find patterns in medical images
- Model development - Fine-tune for specific conditions
For Startups
- Teleradiology - Remote imaging interpretation
- Point-of-care - Portable diagnostic tools
- EHR integration - Automated data extraction
- Clinical decision support - AI-powered recommendations
Getting Started
Access Methods
MedGemma 1.5 4B is available on:
- Hugging Face - Direct model download
- Google Cloud Vertex AI - Managed deployment
- Kaggle - For experimentation
- GitHub - Code and documentation
Hardware Requirements
The model is designed for accessibility:
| Deployment | Requirement |
|---|---|
| Inference | Single consumer GPU |
| Fine-tuning | Single professional GPU |
| Production | Standard cloud instance |
No need for expensive multi-GPU clusters.
Recommended Version
Google recommends the instruction-tuned version as the starting point:
from transformers import AutoModelForCausalLM, AutoProcessor
model = AutoModelForCausalLM.from_pretrained(
"google/medgemma-1.5-4b-it"
)
processor = AutoProcessor.from_pretrained(
"google/medgemma-1.5-4b-it"
)
MedGemma Impact Challenge
Google launched a Kaggle hackathon alongside the 1.5 release:
- Build applications using MedGemma
- Compete for prizes
- Showcase healthcare AI innovation
- Connect with the medical AI community
This is a great opportunity to experiment with the model.
Comparison with Alternatives
MedGemma vs Commercial Solutions
| Aspect | MedGemma 1.5 | Commercial (e.g., Viz.ai) |
|---|---|---|
| Cost | Free | Subscription |
| Customization | Full control | Limited |
| FDA clearance | No (research) | Often yes |
| Support | Community | Enterprise |
| Deployment | Self-managed | Managed |
MedGemma vs General LLMs
| Aspect | MedGemma 1.5 | GPT-5/Claude |
|---|---|---|
| Medical optimization | ✅ Native | ⚠️ General |
| Medical imaging | ✅ CT/MRI/X-ray | ❌ Limited |
| EHR understanding | ✅ Trained | ⚠️ Generic |
| Cost | Free | API fees |
| HIPAA compliance | Depends on deployment | Depends on deployment |
Regulatory Considerations
Important caveats:
Not FDA Cleared
MedGemma is a research tool, not a cleared medical device. It should:
- Support, not replace, clinical judgment
- Be validated before clinical use
- Follow institutional review processes
- Comply with local regulations
HIPAA and Data Privacy
When using MedGemma with patient data:
- Deploy on-premises or in compliant cloud
- Ensure proper de-identification
- Maintain audit trails
- Get appropriate approvals
Related Models
MedGemma is part of a broader ecosystem:
| Model | Focus |
|---|---|
| MedGemma | Medical imaging and text |
| MedASR | Medical speech-to-text |
| BioBERT | Biomedical literature |
| ClinicalBERT | Clinical notes |
MedASR Companion
Google also released MedASR, a medical speech-to-text model:
- Convert doctor dictations to text
- Pair with MedGemma for reasoning
- Support clinical workflows
- Available alongside MedGemma 1.5
Looking Forward
Google's Healthcare AI Roadmap
MedGemma fits into Google's broader health strategy:
- Health AI Developer Foundations - Tools for developers
- Google Cloud Healthcare API - Enterprise integration
- Fitbit AI - Consumer health insights
- Research partnerships - Academic collaboration
Expected Developments
Future updates may include:
- More imaging modalities (ultrasound, PET)
- Larger model variants (e.g., 27B)
- Pre-built clinical applications
- Deeper EHR integration
Key Takeaways
✅ Open source medical AI for CT, MRI, X-rays
✅ 14% MRI improvement over previous version
✅ Longitudinal imaging for tracking changes over time
✅ Single GPU deployment possible
✅ Research use - not FDA cleared for clinical decisions
Explore more healthcare AI: Read about the AI Stethoscope for Heart Disease and Claude for Healthcare.