Powered by
OMNIBIOFEX.CLOUD
DEMOCRATIZING WORLD-CLASS MEDICAL AI
MEDGEMMA
WHAT IT MEANS
MedGemma: The medical-specialized Gemma family.
1.5: The updated generation extending far beyond single 2D interpretations.
4B: An efficient 4-billion-parameter model class.
IT: Instruction-tuned to follow complex clinical prompts and questions.
Multimodal: Integrates a SigLIP image encoder pretrained on de-identified medical data with a medically-trained language decoder.
TECHNICAL SPECS
- Architecture Decoder-only Transformer
- Attention Grouped-Query Attention (GQA)
- Inputs Text + Medical Images
- Context Length 128K Tokens
- Image Resolution Normalized to 896 × 896
- Image Encoding 256 Tokens per image
- Max Output 8,192 Tokens
- Release Version 1.5.0 (Jan 2026)
WHAT IS SPECIAL ABOUT 1.5?
Going considerably beyond ordinary medical question answering and single 2D medical-image interpretation. MedGemma 1.5 expands natively into 3D, temporal, and complex document reasoning.
EHR UNDERSTANDING
Interpret text-based Electronic Health Record information. A massive jump in EHRQA accuracy allows conversion of longitudinal records into summaries, event extraction, and Q&A over patient records.
MEDICAL DOCUMENT → JSON
Extract structured information from unstructured medical laboratory reports. The most compelling engineering use case: Lab Report → MedGemma → Structured JSON.
3D MEDICAL IMAGING
Works natively with three-dimensional representations from CT and MRI scans. An essential evolution beyond flat 2D slice interpretation.
WHOLE-SLIDE HISTOPATHOLOGY
Simultaneously interprets multiple high-resolution patches taken from a whole-slide histopathology image (WSI).
LONGITUDINAL IMAGING
Interpret a current chest X-ray in the strict context of previous images, powering applications focused on disease progression and temporal changes.
ANATOMICAL LOCALIZATION
Supports precise, bounding-box-based localization of medical features and anomalies directly within the text output.
PERFORMANCE BENCHMARKS
While the massive 27B model excels at generalized reasoning, the 1.5 4B model shows incredible, specialized improvements—especially in EHR comprehension and Document-to-JSON extraction.
EHR & DOCUMENT UNDERSTANDING
The jump from 67.6% → 89.6% on EHRQA relative to MedGemma 1 4B is particularly significant.
| Model | EHRQA Accuracy | EHRNoteQA Accuracy |
|---|---|---|
| Gemma 3 4B | 70.9% | 78.0% |
| MedGemma 1 4B | 67.6% | 79.4% |
| MedGemma 1.5 4B IT | 89.6% | 80.4% |
| MedGemma 1 27B | 90.5% | 90.7% |
DOCUMENT TO STRUCTURED JSON
On raw-PDF-to-JSON evaluation (EHR Dataset 2), MedGemma 1.5 4B outperforms even the previous 27B model.
| Metric | MedGemma 1 4B | MedGemma 1 27B | MedGemma 1.5 4B IT |
|---|---|---|---|
| Macro F1 | 78.0 | 76.0 | 91.0 |
| Micro F1 | 75.0 | 70.0 | 88.0 |
IMAGING PERFORMANCE (1.0 vs 1.5)
| Task / Modality | Metric | MedGemma 1 4B | MedGemma 1.5 4B IT |
|---|---|---|---|
| CT Dataset 1 | Macro Accuracy | 58.2% | 61.1% |
| MRI Dataset 1 | Macro Accuracy | 51.3% | 64.7% |
| MIMIC CXR | Macro F1 | 88.9% | 89.5% |
| WSI-Path (Histopath) | ROUGE | 2.2% | 49.4% |
| EyePACS (Fundus) | Accuracy | 64.9% | 76.8% |
| MS-CXR-T (Longitudinal) | Macro Accuracy | 61.1% | 65.7% |
| Anatomy Localization | Bounding-box IoU | 3.1% | 38.0% |
TRAINING DATA
The multimodal components receive deep medical-domain training using both public and licensed/de-identified private datasets.
Included Sources: MIMIC-CXR, Chest ImaGenome, SLAKE, PAD-UFES-20, SCIN, TCGA, CAMELYON, PMC-OA, CT-RATE, alongside internal CT, MRI, dermatology, and EHR datasets.
Note: All datasets are strictly anonymized/de-identified to protect participant and patient privacy.
CUSTOMIZATION
MedGemma is a developer foundation model designed for agentic orchestration and fine-tuning.
- In-Context Learning: Highly responsive to carefully designed prompts and few-shot examples.
- LoRA Fine-Tuning: Parameter-efficient adaptation for specialized clinical tasks.
- Agentic Orchestration: Engineered to serve as a reasoning component alongside FHIR generators, search APIs, and larger reasoning models.
THE MOST IMPORTANT LIMITATION
MedGemma 1.5 4B IT is NOT a clinical diagnostic system.
Google describes MedGemma as a starting point for developers. Its outputs are not intended to directly inform diagnosis, patient-management decisions, or treatment recommendations. Outputs should be regarded as preliminary and independently verified.
WRONG ARCHITECTURE: Patient → MedGemma → Diagnosis/Treatment
CORRECT ARCHITECTURE: Medical Data → MedGemma 1.5 → Extraction/Draft → Validated Application Logic and/or Qualified Human Review