API for clinics, labs, veterinary services and apps
LabReadAI API is a medical neural network you can plug into your product: it recognises lab report forms, reads scans (including DICOM) and symptom descriptions, asks clarifying questions and explains the result in plain language for a patient or a pet owner. Returns structured JSON, works in Russian and English, answers in minutes.
Integration is one POST request and an access key: your interface, texts and flow stay yours, the reading arrives as ready JSON. We open a test environment on your real forms before any contract.
Formats
File type is detected by content, not by extension. One request — up to 10 files of mixed types. Limits below are defaults: everything is negotiable for your integration.
Photo of a form, a screenshot from a lab portal, a photo of film, skin or a pet. We read handwriting, tables, charts and histograms.
up to 10 MB per file
Lab exports, reports, discharge summaries, referrals, scans. Multi-page documents; pages with images are turned into frames automatically.
up to 15 pages
Originals from MRI, CT, X-ray, mammography and ultrasound discs. We build windows from the full 16 bits, compute densities (HU) in code and read projection, side, contrast and slice thickness from the header. Personal tags are never read.
up to 50 MB per file, 150 MB per batch
Lab values, a study report, a discharge summary, a referral or a description of the situation in your own words — as a plain string, no file.
up to 2,000 characters
Cosmetic INCI or supplement facts as text or a label photo — for a "does it suit me" reading.
up to 4,000 characters
HEIC, TIFF, BMP, DOCX and other formats — enabled on request.
on request
Capabilities
Plug in the level you are ready to show your users. The layers are independent and stack.
Layer 1 · for systems
Any lab form becomes normalised JSON: marker, value, units, reference range, date, laboratory. No medical conclusions — for EHR systems, patient portals and storefronts. Disputed rows are re-read by a second model and an arbiter.
Layer 2 · for the user
A full reading: key takeaway, what is off, what it may mean, trends, what to do and which doctor to see. Lab tests, scans, symptoms, pregnancy, children, metabolic health, veterinary. The model asks clarifying questions itself.
Layer 3 · dialogue
Chat under the result: the user asks, the AI answers remembering everything it has read. Or a standalone medical AI assistant inside your interface — within our safety frame and reasoning layers.
What exactly it does
The reading type is a request parameter — the response structure stays the same.
Blood, biochemistry, hormones, urine, vitamins, tumour markers — forms from any laboratory, including handwritten ones and screen photos.
MRI, CT, ultrasound, X-ray, mammography, ECG, endoscopy: a photo of film, a PDF report or the original DICOM.
A description in your own words plus photos: rash, eyes, skin, a wound, a pet's paw. The answer: what it may be and whom to see.
The model asks 1 to 7 questions that cannot be inferred from the data, with example answers — and takes the answers into account.
The user asks about their reading; the model answers remembering the note, the answers and the history.
Three levels: by composition; by composition and lab tests; by composition, lab tests and a photo. The result — a score, safety, a recommendation.
From a story — a list of tests with the "why" and preparation for each.
Several dated forms in one upload — what went up, what went down. A combined view of the whole observation history — on request.
Whom we read
The result always refers to a doctor and contains no prescriptions or dosages.
For vet clinics and vet labs: an explanation for the owner with every form — without night calls to the clinic.
For doctors and vets
If you are a doctor, a vet or a student and did not come for an integration: the same neural network already works on the site. Upload a form or a scan — a second reading arrives in 5–7 minutes.
CBC, biochemistry, hormones, vitamins — double-check your own reading of a form before the visit.
Decode lab testsMRI, CT, ultrasound, X-ray, ECG — a photo of film, a PDF report or the original DICOM from a disc.
Decode scansRanges by species, breed and age: dogs, cats, horses, rodents, birds, exotics.
Decode pet testsA pet's X-ray, ultrasound, CT, MRI — in minutes, with a hint for the owner on what to do next.
Decode pet scansIntegration
Four steps from upload to ready JSON — the same for lab tests, scans, symptoms and veterinary.
A POST request with files and/or text, the reading type and the response language. Access key in the header.
Irrelevant uploads are rejected at once — you do not pay for junk. If the data lacks context, the model forms clarifying questions dynamically for that upload. Questions can be switched off.
Several independent readings, an arbiter and code checks where it is critical. Usually 5–7 minutes.
The same structure for all types. Synchronously, by webhook or by polling.
Example
Press "Send request" — you will see how the pipeline runs and what the JSON looks like.
POST /v1/decode Authorization: Bearer <api-key> Content-Type: multipart/form-data mode=lab # lab | scans | vet | vet-scans | symptoms | pregnancy | ... lang=en note="Female, 34. Fatigue, hair loss." files[]=@labs-2026-08-12.pdf files[]=@ferritin.jpg
{
"id": "6f1c…",
"status": "done",
"subject": { "kind": "human", "sex": "f", "age": 34 },
"recognized": [
{ "marker": "Ferritin", "value": 9, "unit": "ng/mL", "ref": "13–150", "flag": "low" },
{ "marker": "Hemoglobin", "value": 118, "unit": "g/L", "ref": "120–150", "flag": "low" }
],
"result": {
"conclusion": "…",
"findings": [ … ],
"dynamics": [ … ],
"actions": [ … ],
"recommendation": { "specialist": "GP / hematologist", "urgency": "planned" }
},
"questions": [],
"chat": { "available": true, "remaining": 5 }
}The example is illustrative: field names are agreed for your integration. Documentation and the OpenAPI spec — on request.
Reading quality
The same material is read several times independently, disputed rows go to an arbiter, and numbers are recomputed in code where a mistake is critical.
The material is read independently several times. Disagreements are not smoothed over: they are the signal that a row must be re-read.
A disputed marker goes to a second reading, the arbiter decides — the response gets the agreed value.
Where a mistake is unacceptable, values, units and densities (HU) are recomputed by code, not by the model.
An irrelevant upload is rejected before the reading: the user gets a clear refusal, you do not pay for junk.
The key takeaway always comes. "What exactly is off", "What to do", "Freshness of tests" and "What was analysed" are switched on and off in the request itself — separately for each reading.
Reference ranges follow sex and age for people, and species, breed and age for animals.
What changes
The same form, the same lab, the same scan — only what the person on the other side of the screen sees changes.
Build or plug in
A bare model reads text. Everything between a form and an answer you can show a person has to be built — or taken ready-made.
Use cases
The owner gets the form and a clear explanation with it. Scans are read in minutes, DICOM from the machine — directly.
The "recognise" layer for the patient portal and the "explain" layer next to the result: fewer "what does this mean" calls.
A clear summary for the patient before the visit; structured markers — into the record. Clinical decisions stay with the doctor.
The patient uploads tests before the consultation; the app shows the reading, chat and trends in its own interface.
"Does it suit me" on the product card: by composition, by composition and tests, by composition, tests and a photo.
Check-up explanations, a "what to test" list, clarifying questions — as one JSON.
How we start
The order is the same for a clinic, a lab and an app — only the volume and the set of layers change.
Step 1 · Material
Send your real forms, reports or scans and describe what your user should see.
Step 2 · Pilot
We open a test environment: you run your flow and compare the answers with what you expected.
Step 3 · Integration
We agree on response fields, volume, delivery and retention, sign the documents and issue a production key.
Onboarding
Everything below is tuned to your setup: the set of layers, volume, response delivery and retention.
A key for your real forms: check accuracy and response format before a contract.
Per request, in packs or by subscription. The "recognise" layer is cheaper than "explain"; terms are agreed individually.
Take only recognition, only explanation or everything together — the set changes without rewriting the integration.
Synchronously, by webhook or by polling — whatever fits your backend.
Examples in Python, JavaScript and PHP, documentation and a walkthrough of the first responses with your team.
A contract, a data processing agreement and a retention and deletion policy — to your rules.
FAQ
Recognise lab forms into structured JSON, interpret lab tests and scans (MRI, CT, ultrasound, X-ray, DICOM), analyse symptoms with photos, ask clarifying questions, run a chat in the context of the result and match cosmetics and supplements by composition and lab tests. For people and animals, in Russian and English.
JPG, PNG, WebP, PDF up to 15 pages and DICOM (.dcm) up to 50 MB. Text can be sent as a string without a file. Other formats are enabled on request.
Yes: marker values, a study report, a discharge summary or a description of the situation can be sent as text. Scans need a file or the original DICOM.
Yes, these are dedicated veterinary pipelines: dogs, cats, horses, rodents, birds, reptiles, livestock. Lab tests, X-ray, ultrasound, CT, MRI and symptoms with photos.
If the conclusion lacks context, the model asks one to seven questions with example answers. The answers are sent in a second request and taken into account; without answers the reading still completes.
After the reading the user can ask questions; the model answers remembering the result, the note and the answers to the questions. There is also a standalone health AI assistant for your interface.
JSON with the same structure for all types: key takeaway, findings, trends, what to do, which specialist and how urgently. For the "recognise" layer — a list of markers with values, units and reference ranges.
No. The result is informational, is not a diagnosis and contains no prescriptions. The API is not intended for emergencies or for clinical decisions without a specialist.
Files and results are used only for your reading: they are not passed to third parties for advertising. Retention is limited, and for an integration you set the retention and deletion terms in the contract. Processing complies with Russian personal data law (152-FZ).
The price depends on volume and layers. Press "Discuss API integration", describe your product and integration needs — we will reply by email within one or two business days, give test access and propose terms. The price depends on the set of layers, monthly volume and how you collect the response.
Russian or English — the language is set in the request. The input can be in either: an English form can be read in Russian and vice versa.
Such an upload is rejected by the gate before the reading: a clear response says what could not be read and what to send instead. The request is not billed.
A full reading usually takes 5–7 minutes; the "recognise" layer is faster. The response can be collected synchronously, by webhook or by polling — whatever suits your backend.
The key takeaway always comes. The other blocks — "what exactly is off", "what to do", "freshness of tests", "what was analysed" — are switched on and off right in the request, separately for each reading.
The response comes as text blocks and fields with values — you lay them out in your own design. There are no logos, ads or links to LabReadAI in the response.
The test environment is opened before a contract. Then — a contract and a data processing agreement, a tariff for your volume, payment by invoice; retention and deletion terms are fixed in the contract.
Onboarding
Tell us about your clinic, app or project and its integration needs — we will discuss test access and terms.
or write to info@labreadai.com
Readings are informational, are not a diagnosis and do not replace a consultation with a doctor or a veterinarian. The service is not intended for emergencies.