Veterinary practices have few radiologists and clinical pathologists per animal patient, so images are often read by the treating general practitioner or sent to a teleradiology service with turnaround of hours to days. AI reads radiographs and microscope images in minutes, flags abnormal findings and drafts reports; the veterinarian confirms. Unlike human medicine, the EU medical-device regime does not apply, which makes market entry faster but shifts quality assurance to the buyer.
Problem & value
- Today: small-animal and equine GPs interpret most radiographs themselves; difficult cases go to specialist teleradiology. Cytology and blood smears are read by in-house staff or sent to reference labs. Fecal parasite checks are manual microscopy.
- AI value: faster second opinion, triage of urgent cases, consistent reading of routine samples, less staff time on microscopy, better client communication through quick preliminary findings.
- Evidence status: vendors publish sensitivity/specificity for specific findings, but independent peer-reviewed validation across breeds and equipment is thinner than in human radiology; treat vendor figures as claims to verify on your own cases. No benefit numbers are quoted here because none could be verified from primary sources in this review.
Variants
- Small-animal radiography: thoracic and abdominal radiographs of dogs and cats (cardiac size, pleural/pulmonary patterns, fractures, foreign bodies).
- Digital cytology and haematology: AI-assisted smear and slide reading, comparable to HLTH-002 / HLTH-026.
- Fecal and urine analysis: automated parasite egg detection and sediment classification.
- Equine and large animal: lameness imaging, ultrasound; far fewer products, often custom.
- Dental imaging: compare HLTH-014 (human-focused, transferable methods).
- Clinical documentation: scribes and discharge-note drafting for vets, see HLTH-003, HLTH-011.
- Livestock herd level: sensors and cameras rather than diagnostics, see VIS-014, PRED-044.
How it works
- image/sample acquisition (DX/CR plate, ultrasound, slide scanner, fecal imager)
- DICOM/PACS routing (study + species/breed metadata)
- image quality & positioning check (view, exposure, species)
- AI inference (classification / detection / segmentation per finding)
- overlay + preliminary findings in the viewer or practice software
- veterinarian review (accept / edit / reject)
- final report
- optional specialist teleradiology or lab escalation for uncertain or urgent cases
- write-back to practice management system, feedback loop for model monitoring
- Models: convolutional or transformer image classifiers and detectors trained on species-specific data; species and breed matter because anatomy varies strongly (deep-chested vs brachycephalic dogs).
- Language layer: an LLM or VLM drafts report text from findings plus history; it never decides alone.
- Human-in-the-loop: the veterinarian owns the diagnosis; urgent findings escalate to a specialist.
Data preparation
The data are radiographs, ultrasound images and lab results, usually in the practice management system (PIMS) and an imaging workstation, sometimes as DICOM in a small PACS, sometimes as JPEG exports or even printouts. Findings are often free text in the patient record.
Typical gaps: missing species, breed and age metadata, images sent without DICOM headers, inconsistent view labels, free-text findings without codes, duplicate patient or owner records and unclear data protection status, because owner data are personal data.
Preparation steps: - Check the DICOM or export interface of the imaging system and PIMS (PREP-005). - Complete species, breed and view metadata at the point of capture (PREP-008, PREP-010). - Convert legacy JPEG and paper findings into a consistent format (PREP-006, PREP-007). - Build a test set of 100-200 cases with confirmed diagnoses to compare vendors on your own cases (PREP-018), using existing reports as labels where reliable (PREP-020). - Remove owner names before sending images for development (PREP-022).
For a bought service, preparation takes days to weeks and mostly concerns metadata discipline; for own models it is the dominant effort.
Implementation paths
Which path for whom
- smb (veterinary practices): Path A, because integrated AI reading of radiographs and lab services exists at per-study prices and plugs into common practice software; the process is standard.
- enterprise (clinic groups, laboratories): Path A, because commodity image reading is best bought; consider negotiated volume pricing and EU hosting rather than building.
- developer: Path C, because open veterinary and general medical imaging models can be fine-tuned locally on species-specific data, which commercial products do not allow.
Path A — Buy (recommended)
- Radiograph AI services: Antech (with its imaging offering), SignalPET, Vetology, Radimal (mostly US vendors delivering AI plus optional specialist review), MetronMind (Italy, EU-based).
- Lab and microscopy: IDEXX (analysers and digital cytology, e.g. inCite and SediVue families), Zoetis Vetscan Imagyst (fecal and cytology with cloud AI), Heska (Antech group), Scil (Germany, Heska) analysers. Check which are EU-hosted and whether images leave the practice.
- Choose A when: a practice or small clinic group wants results without ML staff; ask for validation by species, EU data residency and a data processing agreement.
Path B — Build on cloud APIs
- VLMs (Claude, Gemini, GPT, Mistral) can describe veterinary images and draft reports, but general models are not validated for radiographic diagnosis; use only for report drafting from veterinarian-entered findings.
- Cloud ML platforms with EU regions (Azure, Google Cloud, AWS) to host custom models; DICOM stores (Azure DICOM service, Google Cloud Healthcare API, AWS HealthImaging) with EU regions.
- Choose B when: you are a software vendor integrating AI into a practice management product.
Path C — Self-host (open source)
- Frameworks: MONAI, nnU-Net, PyTorch/timm; Orthanc as DICOM server; OHIF viewer; Label Studio or CVAT for annotation; YOLO-family detectors for parasite eggs and cells.
- Foundation models: self-supervised vision backbones (DINOv2-class) adapted by fine-tuning; open VLMs (Qwen-VL family) for report drafts.
- Hardware: HW2 is enough for inference and fine-tuning of small models; edge devices (HW1) for slide scanners.
Path D — Custom training
Worthwhile for universities, large clinic chains and laboratories with thousands of labelled, expert- confirmed studies. Needs board-certified radiologists or pathologists for ground truth, species- and equipment-diverse data, and a held-out external test set. Data sharing between clinics needs a legal basis where owner identifiers appear in metadata, so strip them at ingestion.
State of the art (2026)
- Commercial radiograph AI exists for common small-animal findings; detection of single clear findings is stronger than comprehensive reading of a whole study.
- Fecal and cytology AI is established in point-of-care systems with human confirmation of positives.
- Foundation-model and VLM approaches are spreading, but public veterinary benchmarks and datasets are small and fragmented compared with human imaging.
- Limits: domain shift across breeds, positioning and X-ray equipment; rare conditions; low-quality images; training data biased to dogs and cats; false reassurance when AI says "normal". Deployed models need ongoing monitoring against specialist reads.
Shared resources & combinations
- Same stack as HLTH-001 (radiology), HLTH-002 (pathology), HLTH-026 (lab diagnostics): DICOM server, annotation tooling, GPU inference server (HW2).
- Combines with HLTH-003 (ambient scribe) and HLTH-011 (reports), HLTH-004 (decision support), VIS-014 (herd-level monitoring) and PRED-044 (livestock sensors).
Effort, cost & KPIs
- Path A: days to weeks; per-study or per-subscription pricing; PACS/analyser integration is the main work.
- Path B/C: 3–9 months for a validated product including annotation; cost drivers are expert labelling and validation.
- Path D: 12+ months with specialist time.
- KPIs: agreement with specialist read per finding, sensitivity/specificity on local cases, time to report, share of studies escalated, override rate, client-reported turnaround.
People & change
Involve the practice owner or clinic head (budget, liability), the veterinarians and nurses who take and read images, the IT provider or PIMS vendor, and the data protection officer (or external DPO). The image vendor or teleradiology partner is a key supplier: clarify the AVV (Art. 28 GDPR) and where the data are stored. Owner data are personal data, but animal images themselves are not; a DPIA is usually not required unless large-scale owner profiling is added.
Works council: depends. Many practices have none. Where one exists, §87(1) no. 6 BetrVG may apply if the software logs individual users' actions; the BAG sees objective suitability as sufficient. §90(1) no. 3 requires timely information on AI-supported work processes. Agree that usage data are not used to evaluate individual employees.
Support AI literacy (Art. 4 AI Act) with a short briefing on species and breed limits, false negatives and automation bias (GUIDE-006). Choose one or two veterinarians as AI champions and run a shadow-mode pilot of four weeks, comparing AI hints with the final report (GUIDE-009, GUIDE-008). Typical concerns: loss of professional judgement and cost per study. Show that the veterinarian signs every report, that urgent cases still go to specialists, and track the cost per case (GUIDE-010).
Risks, regulation & compliance
- Medical devices: Regulation (EU) 2017/745 (MDR) and IVDR cover devices for humans only, so veterinary diagnostic software is generally outside CE medical-device conformity assessment; there is no harmonised EU pre-market approval for veterinary devices. Veterinary medicinal products fall under Regulation (EU) 2019/6. National veterinary professional rules (e.g. who may diagnose) still apply.
- EU AI Act (Reg. 2024/1689): high-risk classification via Annex I attaches to MDR/IVDR products, so typical veterinary diagnostics are minimal risk (AI literacy, Art. 4). Check if a product is also marketed for humans. Not legal advice; no Art. 5 prohibited practice is involved.
- GDPR: animal data itself is not personal data, but owner names, addresses and linked records are; DICOM headers and notes often contain them. Use DPAs and EU hosting; pseudonymise for training.
- Liability and trust: the veterinarian remains responsible; document AI use in the record and inform owners where it influences diagnosis or cost.
- Mitigations: local validation before go-live, double reading for urgent cases, drift monitoring.
Getting started
- Collect 100–200 recent local cases with final specialist diagnoses as a test set.
- Shortlist 2–3 vendors from Path A and compare on your cases, by species and equipment.
- Confirm EU data residency, DPA and DICOM/PACS integration.
- Define escalation rules and how AI output is recorded in the patient file.
- Review override rate and discordant cases monthly.
Sources
- Regulation (EU) 2017/745 on medical devices (scope: devices for human use) — https://eur-lex.europa.eu/eli/reg/2017/745/oj
- Regulation (EU) 2019/6 on veterinary medicinal products — https://eur-lex.europa.eu/eli/reg/2019/6/oj
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- MONAI project — https://github.com/Project-MONAI/MONAI
- nnU-Net — https://github.com/MIC-DKFZ/nnUNet
- Orthanc DICOM server — https://www.orthanc-server.com/
- IDEXX — https://www.idexx.com/
- Zoetis Vetscan Imagyst — https://www.zoetisus.com/
- SignalPET — https://www.signalpet.com/