In a platform built around clinical cases and interconsultations, the medical image is the central piece of evidence. That makes it a target: when a case carries real value, the motivation to falsify it appears.
VeriMed is the system we use to verify that every image entering Teral is authentic and matches the case it accompanies. It doesn't rely on a single method — it applies six complementary verification layers that produce independent signals.
Those signals don't block anything on their own. They prioritize cases so a human can review them with all the evidence in front of them. The final decision is always made by a person.
Why the clinical image is a target
A well-documented clinical case has real value inside Teral: it builds professional reputation, serves as training material for other physicians, and feeds the platform's knowledge base. That value is what makes the community work.
That same value creates an incentive to deceive. This isn't a hypothetical problem, nor one unique to Teral: any system where visual evidence supports a claim faces the same pressure. And as generative tools have matured, the cost of producing a convincing image has dropped considerably.
The types of falsification we're concerned about fall into three distinct categories:
- Appropriation. Passing off as your own an image that belongs to another case, another physician, or an external source.
- Manipulation. Altering an image to obscure its origin, or generating it with AI tools.
- Clinical incoherence. Accompanying a real image with a clinical history that doesn't match it.
The important point is that these three categories can't be detected the same way. Checking whether an image already existed on the platform is a different problem from evaluating whether it matches the text describing it. A system that only did one of the two would leave the other door open.
Defense in layers
VeriMed doesn't rely on a single trick. It applies a set of complementary layers ranging from the fastest and cheapest to the most thorough. Each layer examines the image from a different angle and produces its own signal.
Evading one layer is possible. Evading all of them simultaneously is considerably harder.
That is the principle underlying the architecture. Robustness doesn't come from any single layer being infallible — it comes from the layers being heterogeneous: they analyze different things, with different techniques, and fail in different ways.
Layer 01 — Provenance and metadata
Reads the file's technical record: capture data, device, timestamps. Detects inconsistencies in the material's origin without looking at a single pixel.
Layer 02 — Exact digital fingerprint
Identifies files that are identical to others already on the platform, regardless of the filename or accompanying data.
Layer 03 — Perceptual hash
Recognizes when it's the same image even if the file is different, looking at visual appearance rather than bytes. Recompression, cropping, or brightness adjustments don't fool it.
Layer 04 — Visual understanding with AI
A vision model interprets the content of the image, not just its form. This makes it possible to find matches between images that look different but show the same thing.
Layer 05 — Specialized analyzers
A set of detectors dedicated to specific signals: recapture indicators, synthetic generation artifacts, and — using a multimodal model — whether the image is consistent with the case's clinical history. A chest X-ray described as a wrist fracture is an immediately obvious discrepancy for this analysis.
Layer 06 — Precision geometric verification
A fine-grained analysis that confirms whether two images correspond to the same physical specimen, even when orientation or framing differ.
Each layer contributes a signal. A scoring engine combines them into a risk level associated with the case, which determines whether it's worth having a person review it.
The principle that governs the whole system
No image is automatically rejected.
VeriMed doesn't block. It prioritizes. When signals indicate risk, the case reaches a reviewer with all the evidence organized and a report explaining why the system flagged it. The decision is made by the person.
This decision isn't technical — it's ethical, and it's worth explaining clearly. The cost of wrongly accusing a legitimate professional is high: it affects their reputation, their relationship with the platform, and their trust in the system. An automated false positive isn't a minor software error; it's harm done to someone who did nothing wrong.
That's why the architecture is built to assist human judgment, not replace it. The detectors produce signals. The triage report, written by AI, contextualizes those signals in plain language. The reviewer decides in a click, with the full evidence in front of them. What AI contributes is speed and context — not verdicts.
The other principles
- Heterogeneous layers, not redundant ones. There is no single point of failure. Robustness comes from the layers being different from one another, not from any one being perfect.
- Provenance ages better than detection. Against generative AI, analyzing an image's content is an endless race. Verifying its origin is a more stable long-term position. VeriMed is designed to incorporate verifiable provenance as those technologies mature.
- Learns from human decisions. Each reviewer's verdict is information. The system refines itself with those decisions rather than operating on fixed rules.
- Integration without duplicating data. VeriMed connects to the existing case system and works on references. It doesn't create parallel copies of images or a second storage infrastructure.
What AI does here
Three pieces of artificial intelligence work at different points in the pipeline, each with a bounded function.
Vision models. They represent each image as a vector capturing its visual meaning. This enables semantic similarity search: finding images that show the same thing, even if they don't resemble each other as files. It's what makes comparing a new image against the full historical archive possible in milliseconds.
Multimodal model. It evaluates the relationship between the image and the clinical text that accompanies it, and drafts the triage report that the reviewer reads. It's the component that translates technical signals into plain-language explanations.
Vector search at scale. The infrastructure that underpins the comparisons. Without it, matching each new image against the entire history wouldn't be feasible at upload time.
One detail that matters: the detection is calibrated with real clinical images from the platform, not general photography datasets. An X-ray doesn't behave like a vacation photo. Compression artifacts, contrast ranges, and capture characteristics are different — a detector trained on natural photos produces noise when confronted with medical material.
Privacy by design
Verifying clinical images requires processing them, and that means being explicit about what is kept and what isn't.
- Anonymization. Patient identifiers are protected before being stored. VeriMed retains no personally identifiable patient data.
- Minimization. The system references images where they already live on the platform. It doesn't generate unnecessary copies.
- Data governance. Designed to comply with applicable healthcare and data protection regulations, with access auditing.
- Transparency toward uploaders. Anyone uploading an image receives a general status on the process. Internal verification methods are not exposed — for the same reason this article describes the layers without detailing their thresholds.
Where VeriMed stands today
It's worth being precise about the system's status, because there's a real difference between a working pipeline and a fully operational system.
As of July 2026:
- Complete pipeline running and integrated with the case system, deployed in the cloud.
- First calibration completed with real clinical images from the platform.
- In controlled testing with the internal team, before expanding coverage to all users.
VeriMed is a system in progress and in evolution. We don't present it as infallible or as a certified medical device, because it's neither. It's an integrity layer that grows with the platform and adjusts with every decision a reviewer makes.
If you want to see the system in action, you can try it directly at verimed-api-1064932368438.us-central1.run.app.
Where it's going
Synthetic content detection is an arms race: every improvement in generation demands a response in detection, and that dynamic has no clear endpoint. Betting everything on detecting fakes is betting on winning an endless race.
The long-term defense isn't detecting fake images better. It's being able to verify the origin of the real ones.
That's where we want to take VeriMed. An image captured inside the application, with content credentials accompanying the file from the moment of capture, doesn't need to be analyzed to determine authenticity — its origin is verifiable by construction. The detection layers will still be needed for everything that enters through other channels, but verifiable provenance is the foundation you can build on for the next ten years.
While those technologies mature, the architecture is ready to incorporate them without being rebuilt.
This article describes VeriMed's layers and the principles that govern them, but deliberately does not publish thresholds, decision parameters, specific model names, or detection metrics. Detailing that information would make it easier for someone trying to circumvent the system — which would defeat the purpose of having it.
