MedGemma 1.5 Technical Report

Andrew Sellergren; Chufan Gao; Fereshteh Mahvar; Timo Kohlberger; Fayaz Jamil; Madeleine Traverse; Alberto Tono; Bashir Sadjad; Lin Yang; Charles Lau; Liron Yatziv; Tiffany Chen; Bram Sterling; Kenneth Philbrick; Richa Tiwari; Yun Liu; Madhuram Jajoo; Chandrashekar Sankarapu; Swapnil Vispute; Harshad Purandare; Abhishek Bijay Mishra; Sam Schmidgall; Tao Tu; Anil Palepu; Chunjong Park; Tim Strother; Rahul Thapa; Yong Cheng; Preeti Singh; Kat Black; Yossi Matias; Katherine Chou; Avinatan Hassidim; Kavi Goel; Joelle Barral; Tris Warkentin; Shravya Shetty; Dale Webster; Sunny Virmani; David F. Steiner; Can Kirmizibayrak; Daniel Golden

Source-linked AI summary

MedGemma 1.5 Technical Report

Andrew Sellergren, Chufan Gao, Fereshteh Mahvar, Timo Kohlberger, Fayaz Jamil, Madeleine Traverse, Alberto Tono, Bashir Sadjad, Lin Yang, Charles Lau, Liron Yatziv, Tiffany Chen, Bram Sterling, Kenneth Philbrick, Richa Tiwari, Yun Liu, Madhuram Jajoo, Chandrashekar Sankarapu, Swapnil Vispute, Harshad Purandare, Abhishek Bijay Mishra, Sam Schmidgall, Tao Tu, Anil Palepu, Chunjong Park, Tim Strother, Rahul Thapa, Yong Cheng, Preeti Singh, Kat Black, Yossi Matias, Katherine Chou, Avinatan Hassidim, Kavi Goel, Joelle Barral, Tris Warkentin, Shravya Shetty, Dale Webster, Sunny Virmani, David F. Steiner, Can Kirmizibayrak, Daniel Golden

arXiv:2604.05081v2Apr 6, 2026cs.AI

TL;DR

Open medical foundation models have limited coverage of complex volumetric, whole-slide, and longitudinal imaging tasks. MedGemma 1.5 integrates these modalities and document understanding in one 4B-parameter architecture, improving performance across diverse clinical benchmarks while retaining text-based competency.

Problem
Open-weight medical models have limited coverage of complex high-dimensional and longitudinal imaging beyond standard 2D tasks.
Method
MedGemma 1.5 adds curated training data, efficiency-oriented modeling updates, and expanded evaluations for volumetric imaging, whole-slide pathology, documents, and clinical reasoning.
Results
MedGemma 1.5 improves across diverse imaging and text benchmarks, including 5% accuracy on MedQA compared with MedGemma 1.0 and higher performance than Qwen3 VL 4B on all vision tasks.
Takeaways & Limitations
The open, efficient 4B model provides a versatile foundation for developer-built medical applications and data-processing tools rather than automated clinical decision-making.
Takeaways & Limitations
The model has minor regressions on specific legacy benchmarks such as SLAKE and VQA-RAD, which may require targeted fine-tuning for narrower tasks.

Abstract

from arXiv · show

We introduce MedGemma 1.5 4B, the latest model in the MedGemma collection. MedGemma 1.5 expands on MedGemma 1 by integrating additional capabilities: high-dimensional medical imaging (CT/MRI volumes and histopathology whole slide images), anatomical localization via bounding boxes, multi-timepoint chest X-ray analysis, and improved medical document understanding (lab reports, electronic health records). We detail the innovations required to enable these modalities within a single architecture, including new training data, long-context 3D volume slicing, and whole-slide pathology sampling. Compared to MedGemma 1 4B, MedGemma 1.5 4B demonstrates significant gains in these new areas, improving 3D MRI condition classification accuracy by 11% and 3D CT condition classification by 3% (absolute improvements). In whole slide pathology imaging, MedGemma 1.5 4B achieves a 47% macro F1 gain. Additionally, it improves anatomical localization with a 35% increase in Intersection over Union on chest X-rays and achieves a 4% macro accuracy for longitudinal (multi-timepoint) chest x-ray analysis. Beyond its improved multimodal performance over MedGemma 1, MedGemma 1.5 improves on text-based clinical knowledge and reasoning, improving by 5% on MedQA accuracy and 22% on EHRQA accuracy. It also achieves an average of 18% macro F1 on 4 different lab report information extraction datasets (EHR Datasets 2, 3, 4, and Mendeley Clinical Laboratory Test Reports). Taken together, MedGemma 1.5 serves as a robust, open resource for the community, designed as an improved foundation on which developers can create the next generation of medical AI systems. Resources and tutorials for building upon MedGemma 1.5 can be found at https://goo.gle/medgemma.

1. Introduction

MedGemma 1.5 expands open-weight medical foundation models to complex, high-dimensional medical imaging and improved medical document and clinical reasoning capabilities. It provides these diverse baseline capabilities within a single architecture as an open resource for healthcare AI development.

Introduction: Expanding open-weight medical foundation models to complex, high-dimensional modalities is described as essential for comprehensive healthcare AI development.The introduction notes that models and benchmarks for complicated imaging tasks remain limited relative to standard 2D imaging.
Introduction: MedGemma 1.5 introduces native support for 3D radiology interpretation of CT and MRI volumes and histopathology whole slide image interpretation.These capabilities expand beyond standard 2D imaging tasks.
Introduction: It adds fine-grained anatomical localization for X-rays via bounding boxes and multi-timepoint radiology analysis.These are two additional native medical imaging capabilities.
Introduction: Additional curated training datasets improve medical document PDF understanding and text-based clinical reasoning.The report positions these improvements alongside the expanded imaging capabilities.
Introduction: MedGemma 1.5 is presented as the first open model combining these diverse baseline capabilities in a single architecture.The model is designed as an improved foundation for the community.

2. Methods

MedGemma 1.5 retains the Gemma3-based architecture with a frozen MedSigLIP vision encoder while expanding training data, post-training, and evaluation for high-dimensional imaging and clinical understanding. Its preprocessing converts 3D volumes and whole-slide images into bounded sequences compatible with the 2D encoder and long-context model.

Model and training: MedGemma 1.5 is based on Gemma3 and uses a 400M MedSigLIP vision encoder.The architecture remains the same as MedGemma 1, with the vision encoder frozen during training.
Model and training: The methodological updates add training datasets for 3D CT and MRI volumes, histopathology whole-slide images, radiology, dermatology, and EHR/lab-report understanding.The updates also include modeling changes for training efficiency and expanded evaluations.
Model and training: Additional pretraining, distillation, and reinforcement learning combine original Gemma text-and-imaging data with new medical image-text pairs.Post-training uses Gemma 3 recipes augmented with medical data and domain-specific teachers for CT, MRI, and histopathology.
High-dimensional imaging preprocessing: 85 axial slices per query were capped for 3D CT and MR volumes, with each 896 × 896 slice processed as a 2D RGB image.The cap corresponds to 21,760 vision tokens and keeps total inputs below 32K tokens with text; CT uses multi-channel HU windowing, while MR uses per-volume min-max normalization with identical RGB channels.
High-dimensional imaging preprocessing: 126 non-overlapping 896 by 896 tissue patches per slide were capped for whole-slide images, producing 32,256 vision tokens.Patches were sampled from tissue-containing regions, with optical magnification stochastically selected at 5x, 10x, or 20x.

3. Evaluations · 3.1. Existing MedGemma Benchmarks

The evaluation suite covers original MedGemma benchmarks and new tasks, spanning medical imaging, visual question answering, report generation, and clinical text. Existing benchmarks assess classification, image-based and text-based multiple-choice question answering, medical-image VQA, and radiology report generation across diverse datasets.

3. Evaluations: Original and new-task evaluation results are reported separately in Table 3 and Table 4/Figure 2, respectively.Table 3 covers original MedGemma 1 evaluation tasks, while Table 4 and Figure 2 present new-task results.
3. Evaluations: MedGemma 1.5 evaluations generally use one inference run per example with temperature 0.0, while baseline models use their default temperature and sampling settings.Temperature 0.0 was also retained for new-dataset MedGemma 1 evaluations for consistency.
3. Evaluations: Benchmark performance can be sensitive to prompt changes, and further prompt optimization may improve results.Prompts were updated and standardized, with similar tasks using the same initial instruction prompt and manual optimization on training and validation splits.
3.1. Existing MedGemma Benchmarks: Chest X-ray classification uses F1-score across MIMIC-CXR, CheXpert, and ChestX-ray14, including five common lung conditions on MIMIC-CXR.MIMIC-CXR evaluation uses both a Med-Gemini test set and another test-set configuration described in the benchmark notes.
3.1. Existing MedGemma Benchmarks: Image-based MCQ accuracy is evaluated on DermMCQA, PathMCQA, and EyePACS.DermMCQA covers 136 skin conditions across 1996 patients; PathMCQA contains 450 patches from 354 whole slide images; EyePACS includes 3614 fundus images with 5-class diabetic retinopathy grades.
3.1. Existing MedGemma Benchmarks: Medical-image VQA is measured by average tokenized F1 on SLAKE and VQA-RAD.SLAKE contains CT, MRI, and X-Ray images spanning multiple body parts, while VQA-RAD covers radiology images of the head, chest, and abdomen.
3.1. Existing MedGemma Benchmarks: Radiology report generation is assessed with RadGraph F1 on a 912-image MIMIC-CXR test set.The test set is the one used in Tanno et al. (2024) and Yang et al. (2024).
3.1. Existing MedGemma Benchmarks: General medical text MCQ accuracy is evaluated on MedQA, MedMCQA, PubMedQA, MMLU medical subcategories, AfriMed-QA, and MedXpertQA.MedXpertQA is identified as out of distribution.

3.2. New Multi-Modal Evaluations

MedGemma 1.5 was evaluated across multimodal medical tasks spanning 3D CT/MR condition classification, pathology whole-slide report generation, longitudinal chest X-ray reasoning, anatomical localization, and laboratory-report extraction. These evaluations used task-specific datasets, prompts, and metrics to assess clinical image and document understanding.

3D radiology: CT and MR evaluations measured binary classification of common conditions across internal head, chest, abdominal/pelvis, brain, knee, and abdomen acquisitions.Internal datasets were balanced by condition, with labels extracted through automated screening, Gemini-based extraction, and manual radiologist review.
3D radiology: CT performance was also evaluated on 1,564 non-contrast chest CT acquisitions from 1,304 patients in the public CT-RATE validation split.Macro accuracy was measured across binary predictions for 18 conditions and abnormalities using the dataset’s original labels.
Pathology: Whole-slide pathology evaluation compared generated text with original reports’ final diagnosis sections using ROUGE for single WSI-text pairs.Each whole-slide image was processed with patch extraction and provided with its specimen label.
Longitudinal imaging: Longitudinal chest X-ray evaluation required determining trajectories of five cardiopulmonary pathologies from paired prior and current studies.The evaluated pathologies were consolidation, edema, pleural effusion, pneumonia, and pneumothorax.
Anatomical localization: Anatomical localization on Chest ImaGenome used bounding-box generation for queried structures and evaluated performance with Intersection over Union.The model received a frontal chest X-ray and was prompted to output normalized 2D box coordinates in JSON format.
Laboratory documents: Laboratory-report evaluation converted document images and rendered PDFs into structured JSON containing clinical test attributes.The task covered pathology reports from EHR datasets 2–4 and Mendeley Clinical Laboratory Test Reports, supporting downstream standardization and interoperability.

3.3. New Text-based Evaluations

The evaluation used EHRNoteQA to assess clinical reasoning over real-world electronic health records. It focused on multiple-choice questions answered from patient discharge notes.

EHRNoteQA: EHRNoteQA contains 962 question-answer pairs derived from MIMIC-IV discharge summaries, covering treatment plans, diagnostics, and patient history.Each instance requires analyzing accumulated discharge summaries for a specific patient.
EHRNoteQA: The evaluation assessed clinical reasoning by asking questions about a specific patient’s accumulated discharge summaries.The benchmark covers diverse clinical topics, including treatment plans, diagnostics, and patient history.
Evaluation format: The evaluation used only the multiple-choice format, presenting each query with patient discharge notes, a question, and five answer choices.Overall accuracy was assessed across answer choices A through E.

4. Discussion

MedGemma 1.5 is presented as an efficient, open-source 4B-parameter medical AI foundation that retains text competency while extending to complex volumetric, pathology, and longitudinal imaging tasks. Its broader capabilities improve several text benchmarks and support practical medical data-processing applications, alongside documented comparison and legacy-benchmark limitations.

Core contribution: A single efficient 4B-parameter model retains established text competency while generalizing to high-dimensional and longitudinal imaging tasks.The discussion highlights evidence synthesis across volumetric scans, whole-slide pathology, and longitudinal imaging.
Core contribution: 5% accuracy on MedQA and 8% on MedXperQA (multimodal) improve over MedGemma 1.0.The gains are attributed to improved distillation and reinforcement learning.
Practical foundation: Enhanced anatomical localization, multi-timepoint analysis, EHR parsing, and lab-report parsing make MedGemma 1.5 a practical foundation for medical-specific data processing.These out-of-the-box functionalities are described as foundational tools rather than automated clinical decision-making.
Comparisons: Qwen3 VL 4B performs better on general text-based biomedical knowledge tasks such as MedQA, whereas MedGemma 1.5 excels on specialized clinical vision tasks.The comparison frames the models as embodying distinct design philosophies.
Comparisons: Some comparative evaluations could not be performed with Qwen3 because of logistical constraints in the internal evaluation framework.Future work may extend the framework to accommodate Qwen’s distinct inference protocols.
Limitations: MedGemma 1.5 shows minor regressions on legacy benchmarks including SLAKE and VQA-RAD as its capabilities expand toward a medical generalist.The discussion notes that these evaluations rely on token overlap and non-standardized ground-truth answers.

5. Conclusion

MedGemma 1.5 expands beyond standard 2D tasks to support high-dimensional and spatiotemporal medical modalities while retaining the computational and cost efficiency of its 4B-parameter architecture.

Conclusion: MedGemma 1.5 advances beyond standard 2D tasks to address 3D radiology, pathology whole-slide imaging, and longitudinal imaging sequences.These are high-dimensional, spatiotemporal modalities.
Conclusion: Despite its complex new capabilities and added document-understanding features, MedGemma 1.5 retains the computational and cost efficiency of a 4B-parameter architecture.
Conclusion: MedGemma 1.5 significantly expands the utility of the original model through these multimodal capabilities.

7. Contributions and Acknowledgments

The section lists the contributors to the MedGemma 1.5 Technical Report and identifies author-order notation for the last and co-first authors.

Contributors: Contributors listed include Timo Kohlberger, Fayaz Jamil, Madeleine Traverse, Alberto Tono, Bashir Sadjad, Lin Yang, Charles Lau, and Liron Yatziv.
Contributors: Additional contributors include Tiffany Chen, Bram Sterling, Kenneth Philbrick, Richa Tiwari, Yun Liu, Madhuram Jajoo, Chandrashekar Sankarapu, and Swapnil Vispute.
Contributors: The contributor list also includes Harshad Purandare, Abhishek Bijay Mishra, Sam Schmidgall, Tao Tu, Anil Palepu, Chunjong Park, Tim Strother, Rahul Thapa, Yong Cheng, Preeti Singh, and Kat Black.
Contributors: Further listed contributors are Yossi Matias, Katherine Chou, Avinatan Hassidim, Kavi Goel, Joelle Barral, Tris Warkentin, Shravya Shetty, Dale Webster, Sunny Virmani, David F. Steiner, Can Kirmizibayrak, and Daniel Golden†.
Author notation: The notation states that † denotes the last author and ∗ denotes a co-first author.

Use of AI in Manuscript Preparation

Parts of the manuscript were drafted with Gemini 2.5 Pro and Gemini 3 Pro, then refined by human editors and manually checked for accuracy.

Use of AI in Manuscript Preparation: Sections of the manuscript were drafted using Gemini 2.5 Pro and Gemini 3 Pro, refined by human editors, and manually checked for content accuracy.The authors state that they take full responsibility for the content.

A. CT-RATE Evaluation

On CT-RATE, MedGemma 4B models required granular inference and showed stronger zero-shot generalization than Gemini 3.0 Flash despite the dataset being out-of-distribution. However, MedGemma 1.5 4B incurred degraded general knowledge reasoning relative to MedGemma 1 4B and Gemma 3 4B.

CT-RATE Evaluation: MedGemma 4B models required a more granular inference strategy for CT-RATE than specialized CT architectures using single-pass multilabel prediction.The framework processed CT-RATE without resampling and required querying the model multiple times.
CT-RATE Evaluation: MedGemma 1 4B and 1.5 4B achieved much higher Macro F1 scores than Gemini 3.0 Flash on the out-of-distribution CT-RATE dataset.The result highlights stronger zero-shot generalization by the MedGemma models than by the general-purpose Gemini model.
General Knowledge Tradeoff: MedGemma 1.5 4B showed degraded general knowledge reasoning compared with MedGemma 1 4B and Gemma 3 4B.The report attributes this decline to a tradeoff associated with intensive imaging-focused fine-tuning in the 4B parameter class.

B. Prompts

MedGemma models use no system instruction by default, while evaluations apply task-specific medical or radiology assistant prompts. MedGemma 1.5 additionally enables silent thinking for selected medical reasoning benchmarks.

System prompts: MedGemma models have no system instruction; radiology evaluations use “You are a helpful radiology assistant,” while other benchmarks use “You are a helpful medical assistant.”The radiology prompt applies to MS CXRT, SlakeVQA, VQA-Rad, and Chest ImaGenome Localization.
Reasoning prompts: MedGemma 1.5 enables thinking by appending “SYSTEM INSTRUCTION: think silently if needed.” to prompts for seven specified benchmarks.These include MedQA, MedMCQA, EHRNoteQA, PubMedQA, MMLU Med, MedXpertQA (Text Only), and AfriMed-QA.