Slynnpack Equine musculoskeletal imaging

Kentucky · Formed September 2026

Thirty-six views
per yearling.
Nobody reads
them at scale.

We are building a computer vision model for equine musculoskeletal radiographs, trained on the most rigorously standardised imaging archive in medicine.

Partner your archive Read the evidence

Decision support for licensed veterinarians. Not a diagnosis. Not a cleared device.

Study
K2609‑0001
Required views
36 yearling
Format
DICOM 3.0, single study
Taken within
21 days of selling
Filed by
4 days before selling
Reviewed by auction
None

§01The problem

The films exist. The reading time does not.

Every yearling catalogued at a major North American auction is radiographed to a fixed minimum standard, submitted as a single DICOM study and filed four days before the horse sells.

The auction house stores those studies and reviews none of them. It warrants nothing about their accuracy. Interpretation falls entirely to a small number of veterinarians, working through thousands of horses across a two-week window, on the busiest days of their year.

A missed finding is not a paperwork error. It is the difference between a horse that sells and a horse that walks back to the barn.

§02Why this archive

Three properties that almost never occur together.

  • 01 / Standardised at source

    The schema is already done

    Fixed view list, mandated study structure, controlled view naming, machine-readable patient identifiers, uppercase fields. The auction houses imposed a data standard years ago, for their own reasons, and enforce it on every consignor.

  • 02 / Priced by experts

    Every study carries a label

    Each horse receives a public hammer price, or a public buy-back. That figure is the aggregated judgement of the best equine veterinarians in the world, expressed in dollars. Human radiology never gets a label like this.

  • 03 / Followed for a decade

    The outcome is public too

    Racing records are published for the working life of the horse. Every image set can be joined to what actually happened on the track, years later, without a follow-up study or a single phone call.

§03What we build

A reading model, and the discipline around it.

Detection and grading for the findings that move yearling prices, beginning with the fetlock, delivered to the veterinarian who already holds the films.

  1. Localise, then grade

    Find the structure first, grade the finding second. The published work in this area shows the two-stage approach outperforms grading the whole film at once.

  2. Return a calibrated confidence, not a verdict

    A number a veterinarian can argue with is more useful than a label they cannot. Every output carries its confidence and the region it was drawn from.

  3. Report agreement, not just accuracy

    Two experienced veterinarians grade the same film differently more often than the industry admits. We measure against a panel, and publish where the panel disagrees.

  4. Stay on the seller's side of the fence

    We work from films a practice already holds, before submission, outside the auction repository entirely. We do not touch repository content and we claim no access to it.

§04The ask

We need films before we need customers.

If your practice holds an archive of equine musculoskeletal studies, that archive is the single scarcest input in this field.

We are looking for retrospective, de-identified studies under a written data licence. Not exclusivity. Not your client list. Not anything that leaves your control without your signature on it.

See the partnership terms

What a partner practice receives

Cost to the practice
None
Named collaboration
Yes, if wanted
Publication credit
Co-authorship
Early access
Before general release
Data leaves your control
Never without signature
Exclusivity demanded
None

Full terms, de-identification method and the licence itself are on the practices page. We will send the draft licence before you commit to anything.

§05Evidence

The hard part is smaller than it looks.

Published research has already graded one of these findings from a few hundred images, using radiographs from a central Kentucky practice.

That result is not ours. We cite it because it sets the floor: this problem does not need a million labelled studies to become useful, and a Kentucky practice has already shown it will share films for the purpose.

The full citation

Third-party published result

Deep learning model shows promise for detecting and grading sesamoiditis in horse radiographs

Radiographs used
255
Reported accuracy
92.7%
Mean average precision
81.8%
Journal
American Journal of Veterinary Research
Year
2023

Guo L, Yu X, Thair A, Rideout A, Collins A, Wang Z J, Hore M. Not affiliated with Slynnpack. Cited as independent published work.

§06Where we are

Early, and saying so.

The company was formed in September 2026. There is no product to demonstrate yet, and we would rather tell you that than stage a screenshot.

Roadmap

  1. Company formed
  2. First retrospective data licences signed
  3. Fetlock detection and grading baseline
  4. Pre-submission screening pilot with consignors
  5. Pre-purchase examination support beyond racing

Hold an equine radiograph archive?

Then you hold the part of this that cannot be bought, hired or scraped. We would like to talk.

Start a conversation hello@slynnpack.com