Runs in the room
Inference is local by design, using compact hardware inside the hospital. The current system is planned around two PoE cameras and a Jetson Orin NX, so identifiable operating-room video does not need to leave the premises.
Remove the instruments the surgeon never uses.
ORite tracks instrument use across procedures, then gives operating-room and sterile-processing teams the evidence to build smarter surgical trays.
Scroll to follow the loop
We observe
A tray-facing camera records instrument departures and returns while local inference turns them into event data. Across cases, recollection becomes a ranked picture of what the team actually reaches for.
OR Managers reduce
The manager selects a usage threshold and sees the trade-off: a smaller everyday tray, rescue instruments kept nearby, and named preference pouches. The result is a proposal, never an automatic deletion list.
Surgeons approve
Surgeons review proposed changes instrument by instrument. They can approve the suggested destination or keep an item. Nothing routinely used is removed, and clinical authority stays with the team.
SPD implements
The approved set passes through sterilisation before instruments are prepared as an everyday tray, satellite tray, or individual pouches. The team receives a grouped count sheet, revisioned label, and explicit stock changes.
You save
The labelled everyday tray travels back to the operating theatre and resolves onto the instrument table. The next case measures the approved configuration again, so tray design remains a governed, evidence-led loop.
Built for the constraints an operating theatre actually has.
A new instrument needs no retraining
Identity comes from matching against a small gallery enrolled for that case, not from a fixed-class classifier. Add an instrument type and it works. This is what lets a new hospital onboard without a model rebuild.
Events, not footage
Storing video is wasted data and a privacy liability. Storing the event (instrument, action, timestamp) compresses the useful signal by orders of magnitude and sidesteps most of the exposure.
Vendor-neutral by construction
An instrument manufacturer has weak incentive to help a hospital buy fewer instruments. We have no catalogue to protect, and the whole stack is built on permissively licensed components.
Nothing leaves the theatre but events.
The camera faces the instrument table, below the sterile field. No footage is retained, no patient is in frame, and what persists is a list of instruments and timestamps. The privacy position is a consequence of the data model, not a policy bolted on afterwards.
How it works, and what it does not claim.
Does it prove an instrument was used on the patient?
No, and we will not claim otherwise. A tray-facing camera sees that an instrument left the tray and for how long. Picked up and set down looks similar to used. We approximate with a duration and count threshold, and the errors average out across cases. For set-trimming decisions that is the right level of rigour; for anything clinical it is not.
What does theatre staff have to do differently?
One photograph of the laid-out tray at setup, confirmed against the documented set list. That is the enrolment step, and it is the whole operational burden. If it takes longer than about a minute it will not happen, so that is the budget we design to.
Where does the data go?
Inference runs on a machine in the room. What persists is an event log: instrument, action, timestamp. No video is written to disk. Aggregation across cases happens on those events, which are already de-identified.
Can it tell near-identical instruments apart?
That is the pivotal technical question and we treat it as one. Fine-grained discrimination runs once per case on a clean setup photograph, not continuously on a cluttered pile. The vision stack is open-set by design, so new instrument families can be enrolled without rebuilding the model.
What happens to instruments that get removed from a set?
Nothing automatically. The output is evidence (usage frequency per instrument per set), and the decision stays with the people who own the tray. Every removal keeps the case count behind it on the record.