At ReadYourLab.com, we use MedGemma 1.5—an open-source medical AI model developed by Google—to help you understand your CT and MRI scans. It is one of the most advanced medical imaging AI models available today.
But no AI knows everything. MedGemma was trained on specific body parts and specific types of scans. It tends to perform well on body parts it has studied and less reliably on body parts it has never seen. This page explains exactly what MedGemma was trained on, so you know what to expect.
1. How MedGemma Learned
A CT or MRI scan is not a flat picture—it is a "volume" made up of hundreds of slices that together form a 3D image of your body. Unlike most AI models that only look at flat 2D images, MedGemma 1.5 was trained to analyze full 3D volumes, meaning it can look at your entire scan at once, just like a radiologist would.
One important detail: MedGemma was trained exclusively on axial (cross-section) images—the view where the body is sliced horizontally, like looking down from above. It was not trained on sagittal (side view) or coronal (front view) orientations. This matters because many common scan protocols—such as spine MRI or certain knee MRI sequences—rely on sagittal or coronal views.
Google trained MedGemma using a combination of:
Real Medical Scans
Thousands of de-identified CT and MRI volumes from real patients, covering specific body parts (detailed below).
Medical Knowledge
Millions of medical textbooks, research papers, and clinical Q&A datasets, giving it broad medical knowledge beyond just images.
Clinical Records
Real Electronic Health Records (EHRs), lab reports, and discharge summaries, so it understands how findings connect to real patient care.
2. Body Parts and Scan Types It Was Trained On
This is the most important section on this page. MedGemma was trained on specific body parts using <strong class="text-slate-800">axial (cross-section) images only</strong>. The AI is most reliable when analyzing scans that match its training data.
Head / Brain
CT and MRI (axial)
Axial scans of the head, including brain tissue, fluid spaces, skull structure, and the upper neck (upper cervical spine). Standard head scans typically extend down to about the C1–C3 vertebrae.
Chest
CT (axial) and X-ray
Axial CT scans of the chest area, including lungs, breast tissue, and thoracic spine visible in the field of view. Also trained extensively on chest X-rays, including comparing current vs. past X-rays to track changes over time.
Abdomen
CT and MRI (axial)
Axial scans covering abdominal organs such as the liver, kidneys, spleen, and surrounding structures. The lumbar spine and upper pelvis are also visible in these scans.
Knee
MRI (axial) and X-ray
Axial MRI of the knee joint, including ligaments, cartilage, and bone structures. Also trained on knee X-rays.
3. Body Parts It Was NOT Trained On
MedGemma was not trained on scans of the following body parts, nor on sagittal or coronal image orientations. If you upload a scan of one of these areas, the AI may still attempt to describe what it sees, but its analysis will be significantly less reliable.
Sagittal or coronal scans (any body part)
Lower neck / Thyroid (below C3)
Lower pelvis and Hip joints
Shoulder
Wrist, Hand, and Forearm
Ankle, Foot, and Lower Leg
Heart (dedicated cardiac imaging)
A note on overlap: Chest and abdomen axial scans do capture parts of the spine (thoracic and lumbar vertebrae) and breast tissue in the field of view, so the AI has some exposure to these structures. Head scans include the upper neck down to about C1–C3. However, it was not trained on dedicated spine imaging (sagittal MRI for disc herniations), dedicated breast imaging (mammography), or any non-axial orientations. Incidental visibility in axial slices is not the same as focused training.
4. Scan Types (Modalities) Supported
MedGemma was trained on these types of medical imaging:
CT Scans (Computed Tomography)
Axial 3D volumetric CT of head, chest, and abdomen only.
MRI Scans (Magnetic Resonance Imaging)
Axial 3D volumetric MRI of head, abdomen, and knee only.
X-rays
Chest X-rays and knee X-rays (2D images).
The model was not trained on sagittal or coronal image orientations, ultrasound, PET scans, mammography, DEXA bone density scans, or angiography. Many common scan protocols use sagittal or coronal views (especially spine and joint imaging), which fall outside the model's training.
5. Accuracy Numbers: What They Mean
Google tested MedGemma on specific benchmarks. These scores tell you how well the AI performs on the exact types of scans it was trained on. They do not apply to body parts or scan types outside its training data.
MRI (Trained Areas)
64.7%
Accuracy across 10 categories of MRI findings for head, abdomen, and knee. This means it correctly identifies the finding about 2 out of 3 times.
CT (Trained Areas)
~61%
Accuracy for CT findings in head, chest, and abdomen. Similar to MRI, roughly 3 out of 5 correct identifications.
Chest X-ray
89.5%
For the top 5 most common chest X-ray findings, MedGemma performs very well. This is one of its strongest areas.
Medical Knowledge
~70%
Scores nearly 70% on US Medical Licensing Exam style questions, showing strong general medical reasoning.
Important: These accuracy scores were measured on the body parts and scan types listed above. If you upload a scan of a body part the AI was not trained on (e.g. a spine MRI or a shoulder CT), the real accuracy will be significantly lower than these numbers suggest. Always discuss results with your doctor.
6. What It Does Well
When analyzing scans that fall within its training data (head, chest, abdomen, knee), MedGemma 1.5 is particularly good at:
Spotting Masses and Lesions
Detecting abnormal growths that span across multiple layers of a 3D scan, since it was trained to understand full volumes, not just flat slices.
Identifying Where Things Are
Precisely locating organs and structures, and noting when something looks abnormal in size, shape, or position.
Tracking Changes Over Time
Comparing a current chest X-ray to an older one to determine whether a condition is improving, stable, or getting worse.
Explaining Findings in Plain Language
Thanks to its medical text training, MedGemma can describe what it sees in terms that are easier to understand than a typical radiology report.
7. Limitations You Should Know
Being transparent about limitations is just as important as highlighting capabilities:
It Is Not a Doctor
MedGemma is a foundation model designed to assist, not replace, a radiologist. Its output is a starting point, not a diagnosis. Always consult a medical professional.
Axial Images Only
MedGemma was trained exclusively on axial (cross-section) images. Sagittal (side view) and coronal (front view) scans were not part of its training. If your scan uses these orientations, the AI is working outside what it has learned.
Training Gaps = Knowledge Gaps
Body parts outside its training data (dedicated spine imaging, pelvis, shoulder, etc.) will produce less reliable results. The AI may sound confident even when it is wrong about areas it has not studied.
Rare Conditions Are Harder
Even within trained body parts, rare diseases or unusual presentations may be missed. The AI learns from patterns in the data, so conditions it has seen rarely will be harder to detect.
Image Quality Matters
Motion artifacts (blurring from patient movement), low-resolution scans, or unusual scan protocols can reduce the AI's accuracy.
No Personal Context
The AI does not know your medical history, medications, or symptoms. A radiologist interprets your scan in the context of your full health picture—the AI cannot.
English-Only Evaluation
MedGemma was primarily evaluated using English-language prompts. Performance in other languages has not been formally tested.
Our Commitment to You
We believe in giving you honest, transparent information. When you upload a scan, we tell you upfront whether it falls within MedGemma's training data. Our goal is to help you walk into your next doctor's appointment feeling informed, prepared, and ready to ask the right questions—not to replace the expertise of your medical team.
