Medical image analysis is the use of software, increasingly AI, to find, measure and describe features in scans such as X-rays, CT, MRI, ultrasound, mammograms and pathology slides. In 2026 the FDA lists over 1,600 AI-enabled devices, about three-quarters of them in radiology. Large randomised evidence now exists for some uses, such as AI-supported mammography screening. Deploying these tools safely depends on regulation, human oversight and where the model runs.
Updated October 2026
This article is for technology and operations teams. It is not medical advice, and no AI tool described here replaces a qualified clinician.
What changed in 2026
- The FDA updated its AI-enabled medical devices list on 4 Sep 2026. An independent count puts it at 1,614 devices with decisions through 29 Jun 2026, 76.2% of them in radiology (Complizen). The FDA says it will explore tagging devices that use foundation models.
- The MASAI randomised trial of about 106,000 women in Sweden reported 12% fewer interval cancers with AI-supported mammography screening than with standard double reading (82 vs 93), in The Lancet, 2026 (Lund University).
- The EU's Digital Omnibus on AI entered into force on 27 Jul 2026. High-risk AI Act obligations for AI in regulated products, including medical devices, now apply from 2 Aug 2028; MDR and IVDR duties are unchanged (Lewis Silkin, Pure Global).
- Open medical foundation models are available for research and development, such as Google's MedGemma and MedSigLIP (2025). Their licence does not allow direct diagnosis or treatment without regulatory approval (The Decoder).
- At RSNA 2025, foundation models, agentic AI and automated preliminary reports were among the most-discussed topics (AuntMinnie).
What is medical image analysis?
Medical image analysis turns pixels from a scan into information a clinician can use: a measurement, an outline, a flag or a draft finding. It has three common task types:
| Task | What it does | Example |
|---|---|---|
| Classification or triage | Labels or prioritises a whole image or study | Flagging a suspected finding so it is read sooner |
| Detection | Finds and marks regions of interest | Marking areas on a mammogram for a radiologist to review |
| Segmentation and measurement | Outlines structures and measures them | Outlining an organ or lesion to track size over time |
Most imaging data is stored in the DICOM format and viewed in a PACS (picture archiving and communication system) or a DICOM viewer.
How AI medical image analysis works
- Acquire the study from the scanner or PACS in DICOM format.
- De-identify where data leaves the clinical system, removing patient identifiers from headers and burned-in text.
- Preprocess: resample, normalise intensity and select the relevant series.
- Run inference with a cleared model, for classification, detection or segmentation.
- Present results to a clinician in the PACS, worklist or report, never as a final diagnosis on its own.
- Record the clinician's decision and any disagreement with the AI.
- Monitor performance over time, since scanner changes, protocols and patient mix can shift accuracy.
Where the evidence is strongest: mammography screening
The MASAI trial is the clearest large-scale evidence so far. AI (ScreenPoint Medical's Transpara) triaged screening mammograms, sending low-risk exams to a single reader and high-risk exams to two readers.
| Outcome | Result | Source |
|---|---|---|
| Cancer detection | 29% higher, without more false positives | Lancet Digital Health, Feb 2025 |
| Screen-reading workload | 44% lower | Same |
| Interval cancers (missed between screens) | 12% fewer (82 vs 93) over two-year follow-up | The Lancet, 2026 |
These results apply to one AI product, one screening programme and one workflow. They do not show that every imaging AI tool performs this way.
Foundation models for radiology and pathology
Narrow tools are trained for one finding on one modality. Foundation models are pretrained on very large image sets, often without labels, and then adapted to many tasks with less labelled data. Many use vision transformer backbones.
| Model | Area | What is reported |
|---|---|---|
| Prov-GigaPath (Microsoft, Providence, UW) | Digital pathology | Trained on about 1.3 billion tiles from 171,189 slides of 30,000+ patients; best results on 25 of 26 benchmark tasks (Nature, May 2024) |
| MedSigLIP (Google) | Medical image encoder | 400M parameters, 448 x 448 input, trained on over 33 million image-text pairs |
| MedGemma (Google) | Image and text model | 4B and 27B sizes; research and development use, not direct diagnosis without approval |
Benchmark gains do not always carry into clinical practice. Any foundation-model product used for patient care still needs its own validation and regulatory clearance.
Regulation in 2026: FDA, EU AI Act and MDR
| Market | What applies | Key dates |
|---|---|---|
| United States | Premarket clearance (510(k), De Novo or PMA); final PCCP guidance lets makers pre-agree how a model will be updated | PCCP guidance finalised Dec 2024; lifecycle guidance still in draft (Jan 2025) |
| European Union | MDR or IVDR certification, plus the AI Act for AI that is a safety component of a regulated device | AI Act high-risk obligations for these devices from 2 Aug 2028 (Pure Global) |
For deployers, three practical points follow. Keep model versions traceable to the cleared version. Plan updates through an approved change plan. Log human review of AI outputs.
Why edge and on-premises inference matter for imaging
Imaging studies are large and sensitive. Running inference inside the hospital network or on a device next to the scanner helps in three ways:
- Privacy: patient images do not leave the site, which simplifies data-protection review.
- Latency and bandwidth: large CT, MRI and pathology files are processed without upload delays.
- Resilience: triage keeps working during internet outages.
Vendors build for this. NVIDIA's IGX and Clara Holoscan platform targets low-latency AI at the medical edge, and surgical video company Proximie planned to process operating-room video locally for performance and privacy (NVIDIA). The trade-off is operational: many sites mean many model copies to deploy, update, monitor and audit.
Where it's heading: 2026 → 2030
| When | What it looks like |
|---|---|
| Around 2020 | Our view: most cleared imaging AI is narrow, one finding on one modality, often a CNN, and little randomised evidence exists. |
| 2026 (now) | 1,614 FDA-listed AI-enabled devices, about 76% in radiology (FDA, Complizen). Randomised screening evidence from MASAI. Foundation models are mostly in research. |
| By 2028 (expected) | EU AI Act high-risk obligations apply to AI medical devices from 2 Aug 2028 (Pure Global). The FDA has said it will explore tagging foundation-model devices on its list. |
| By 2030 (expected) | Our view: more cleared products are built on foundation models and updated under pre-approved change plans, and hospitals run them on-prem with central monitoring and audit trails. |
Run imaging AI on-prem and at the edge with NexaStack
NexaStack deploys, updates and monitors AI models across hospital sites and edge devices from one control plane. Models stay inside the site network, every version change is recorded, and each AI-assisted decision can be traced to the model version and the human reviewer. NexaStack is infrastructure for running models; it is not a medical device and does not provide diagnoses.
Frequently asked questions
What is medical image analysis?
Medical image analysis is the use of software to extract information from medical scans: measuring structures, outlining regions, flagging possible findings or drafting descriptions for a clinician to review.
What are four common medical imaging techniques?
X-ray (including mammography), computed tomography (CT), magnetic resonance imaging (MRI) and ultrasound. Nuclear imaging such as PET and digital pathology slides are also widely analysed with AI.
Can AI analyze medical images?
Yes. The FDA lists over 1,600 AI-enabled devices, most in radiology. These tools support clinicians; they do not replace their judgement, and each is cleared for a specific use.
Can ChatGPT analyze an MRI or make a medical diagnosis?
General-purpose chatbots are not cleared medical devices and should not be used to diagnose. Diagnosis needs a qualified clinician using validated, approved tools.
What are the latest medical image analysis trends in 2026?
Foundation models for radiology and pathology, randomised trial evidence for AI-supported screening, automated draft reports, EU AI Act timelines for medical devices, and on-prem or edge inference to keep patient data on site.
Run imaging AI on-prem with full traceability
See how NexaStack keeps imaging models inside your network and links every AI-assisted output to its model version and human reviewer. NexaStack is infrastructure for running models, not a medical device. Book an on-prem imaging AI infrastructure demo →
