The conversation has moved from “can AI read scans” to “which AI is cleared, how does it integrate, and who signs the report.”
Three years ago, the debate around AI in medical imaging was still about capability. Could a model match a radiologist in lung nodule detection? Could it find a stroke faster than the on-call attending? Those questions are mostly answered now. The FDA’s list of AI/ML-enabled medical devices has crossed 1,000 entries, and the modal radiology department in a US health system runs at least one AI tool in production every day.
The interesting questions are different now. Which models are FDA-cleared for your specific use case? How they integrate with your PACS, RIS, and reporting workflow. What the regulatory submission looks like when you build your own. Who’s accountable when the model misses a finding? And – for healthtech companies and medical device vendors building their own diagnostic AI – what the production architecture has to look like to clear regulatory review and survive in clinical use for the years that follow.
Here’s what we see working.
The model landscape isn’t what it was
The biggest shift since the “is this technically possible” era is that you rarely train a medical imaging model from scratch anymore. The foundation models have caught up, and the build math has changed.
For radiology, MAIRA-2 from Microsoft, Google’s MedGemma, and academic foundation models like RadFM provide strong starting points for chest X-ray, CT, and report-generation tasks. For pathology, Path Foundation (Google), Virchow (Paige), and CHIEF cover most whole-slide imaging tasks. For general segmentation, TotalSegmentator handles automated anatomical labeling across CT and MRI with surprising robustness, and MedSAM extends segment-anything to medical imagery. MONAI remains the de facto framework for healthcare-specific model development, with strong tooling for DICOM ingestion, augmentation, and deployment.
The practical implication is that your team’s effort moves from “design and train a model” to “fine-tune a foundation model for your specific clinical task, validate it across your patient population, and integrate it.” That’s a different competency profile from what most healthtech AI teams hired for three years ago.
The other shift: general multimodal models like GPT-5, Claude Opus, and Gemini 2.5 are good enough at image understanding to be useful for non-diagnostic clinical workflows – patient communication drafts, report summarization, documentation, prior-authorization writing. They aren’t medical devices. Don’t position them as ones.
Where computer vision actually pays off in production
Not every imaging task benefits equally from AI in the clinical workflow. The patterns we see working consistently:
Worklist triage and prioritization. AI flags scans likely to contain critical findings – intracranial hemorrhage, pulmonary embolism, aortic dissection – and pushes them to the top of the radiologist’s queue. The radiologist still reads the scan and makes the call. The AI just changes the order. This is the highest-ROI, lowest-regulatory-friction use case in most departments and a strong place to start.
Pre-read and second-read assistance. The model highlights suspected findings; the radiologist confirms, rejects, or adds. Mammography screening, lung CT screening, fracture detection, and brain MRI for MS lesions all have mature workflows for this pattern. The trust model is intact – the human always signs the report.
Quantification. Volumetric measurements of tumors, organs, lesion counts, and cardiac function metrics. These are tedious for a human to do precisely and well-suited to automation. The model proposes the measurement; the radiologist accepts it or adjusts it. The output is written to a structured report field.
Screening at scale. Population-level mammography, lung cancer screening, and diabetic retinopathy. Well-suited to AI because the volume justifies the validation work, and the workflow is standardized.
Digital pathology. Slide-level AI is mature enough now to handle prostate, breast, colorectal, and lung pathology workflows. The reading pattern is the same as in radiology: the model proposes regions of interest, and the pathologist confirms them.
Report drafting. Most radiology reporting is now AI-assisted at the draft level. The model produces structured findings based on the imaging and prior reports; the radiologist edits and signs. The productivity gain is real, and the regulatory profile is manageable because the human is in the loop.
Where we still recommend caution: anything positioned as an autonomous diagnosis without a clinician in the loop, anything making a recommendation that affects a treatment decision without an explainability artifact the clinician can see, and anything trained on a single site’s data without validation across patient populations, scanner manufacturers, and acquisition protocols.
The architecture that scales
Production CV in healthcare looks different from a production CV in general software. A few patterns we use repeatedly:
DICOM and HL7 FHIR integration are non-negotiable. Your AI service must receive DICOM studies via the modality worklist or PACS, return structured findings via FHIR Observations or DICOM Structured Reports, and trigger PACS notifications without disrupting existing workflow. If your model is great but the integration is bad, no department will deploy it.
On-premises is still the default for most hospital deployments. Cloud-only is a hard sell for production clinical use in most US health systems and nearly impossible in EU public health systems. Plan for containerized on-premises deployment behind the hospital firewall, with model updates pushed through a controlled release channel.
Edge inference for time-sensitive cases. Stroke detection, large vessel occlusion, hemorrhage triage – anything where seconds matter – runs at the scanner edge, not in a regional data center. The latency budget is in seconds, not minutes.
Federated learning when you need cross-site training data. Training data must remain within each institution’s network. Federated frameworks let you train across multiple sites without the data ever leaving. This is what makes models robust across scanner manufacturers and patient demographics.
Subgroup performance monitoring, not just aggregate accuracy. Performance has to be tracked across protected classes (age, sex, race, where available) and across scanner manufacturer, model, and acquisition protocol. Drift in subgroup performance is the most common pattern of post-deployment failure, and aggregate metrics won’t show it.
Audit trails for every inference. Which model version, which input study, what was returned, what the radiologist did with it, and what was in the final report. This is what your compliance team will need when a finding is missed, and you have to explain the system’s behavior years later.
The regulatory layer
The FDA’s 2024 finalization of the Predetermined Change Control Plan framework was the biggest practical shift for AI medical devices. You can now pre-specify a class of model updates – retraining on additional data, threshold adjustments, expansion to additional anatomies – and execute them without a new 510(k), as long as you stay within the PCCP. For any team planning to iterate models post-clearance (which is to say, everyone), the PCCP is now the most important part of your submission strategy.
Good Machine Learning Practice principles (jointly published by FDA, Health Canada, and MHRA) define what regulators expect across the lifecycle. In the EU, the AI Act classifies most diagnostic AI as high-risk, layering additional conformity assessment requirements on top of MDR. CE-marking timelines for AI medical devices have stretched as notified bodies work through the overlap with the AI Act.
Practical guidance: design for regulatory submission from day one, not retroactively. The eval suite, the subgroup analysis, the audit trails, the bias testing, the validation across sites – these are what your submission depends on. Trying to assemble them after the fact for a model that wasn’t built with them in mind is how clearance timelines slip by by 12 to 18 months.
How this connects to broader healthcare AI
Diagnostic computer vision is one piece of the larger healthcare AI picture. Our work on the AI-powered personalized healthcare platform – combining EHR data, wearables, and genomic signals to drive personalized care pathways – shares much of the same infrastructure: HL7 FHIR integration, HIPAA-grade data handling, audit-grade observability, and architecture that supports both inference at the edge and orchestrated reasoning on the broader patient record. The CV component is one input into a larger clinical AI surface. The teams getting the best results aren’t running these as separate stacks.
Where this leaves things
If you’re a medical device vendor, healthtech company, or health system considering deploying or building diagnostic computer vision, the work in 2026 is fundamentally about integration, validation, and regulatory architecture – not model accuracy. Models are mostly solved problems at the foundation layer. The hard parts are the parts most teams underinvest in.
If you’ve got a candidate use case or a deployment that’s stalled in regulatory review, that’s the conversation we like having. Contact us to find out more.
