Stryker Corporation has a published US patent application on file that would take a picture of tissue a surgeon is looking at right now and produce a picture of what that tissue is expected to look like weeks later. The application, US20260203900A1, titled METHODS AND SYSTEMS FOR CHARACTERIZING TISSUE OF A SUBJECT, published on July 16, 2026 and names Lina Gurevich as inventor. It is an A1 publication — a pending application, published and open to inspection, not a granted patent, and nothing in the record indicates it has been examined to allowance.

Read as a business document rather than a technical one, the interesting part is not that Stryker is using machine learning on medical images. Almost every device manufacturer with an imaging line is. The interesting part is where in the surgical workflow this particular application places the model. Claim 2 — the first operative claim, since claim 1 was canceled before publication — is a method claim with four steps: receive a fluorescence image of the subject's tissue; provide that image to a generator of a trained generative model; obtain from that generator a simulated white-light image depicting the tissue's predicted future state; and display it. The generator is the recited component, not the discriminator and not a classifier sitting downstream. The output is not a score, a heat map, or a segmentation mask. It is a photograph that does not exist yet.

The dependent claims are where the commercial shape becomes legible. Claim 3 defines the future states the method is meant to depict, and it is a short, blunt list: necrosis, delayed healing, healing. Claim 8 narrows the input to a near-infrared image; claims 9 and 10 narrow it further to an intraoperative perfusion image and to a frame drawn from an intraoperative perfusion video. That is a precise description of the data a fluorescence imaging platform already generates in the operating room during a procedure. The application is directed at making a second, predictive use of a signal that surgical teams are capturing anyway.

Claim 11 lists the tissue types: breast tissue, burnt tissue, chronic wound tissue, acute wound tissue, or skin transplants. Those are reconstructive, wound-care and burn-care settings — the clinical areas where perfusion assessment already drives an intraoperative decision, and where the cost of getting that decision wrong shows up later as a return to the operating room. Claims 4 and 5 add a recommendation step layered on the prediction, including a recommendation for administering a treatment. Claim 6 adds a separate classification model that consumes the generated white-light image to identify complications. Taken together, the dependents describe a decision-support stack, not a single image filter.

The training data is the part that is hard to copy

Claims 12 through 19 cover how the model gets trained, and this is where the application says the most about competitive position. Claim 12 recites training on a plurality of image pairs. Claim 13 defines what a pair is:

each image pair of the plurality of image pairs comprises a fluorescence image of a particular tissue during an operation and a white-light image of the particular tissue after the operation— claim 13, METHODS AND SYSTEMS FOR CHARACTERIZING TISSUE OF A SUBJECT, US20260203900A1

That single sentence describes an asset that is considerably harder to assemble than a model architecture. A paired corpus requires intraoperative fluorescence capture linked to the same patient's post-operative photography, at scale, across enough cases to train a generator. The architectures themselves are named in the claims and are public: claim 14 recites a pix2pix GAN model and claim 18 recites a CycleGAN model, both well-established published approaches. The application hedges against the data problem too — claims 15 through 17 cover training on unpaired image data collected from a plurality of patients, and claim 19 covers training across two imaging modalities. The claim set therefore reads as though it was drafted to survive whichever training regime the eventual product settles on, which is a practical drafting posture rather than a statement about scope.

One drafting detail is worth flagging in a neutral voice, because it bears on how this record reads today rather than how it may read after prosecution. Claim 2 and its two mirrored independents — claim 20, a system claim, and claim 21, a non-transitory computer-readable medium claim — each recite displaying, on the display, without any earlier step introducing a display. There is no antecedent for the definite article. The same phrasing appears in the abstract. It is the kind of indefiniteness issue that examiners routinely raise and applicants routinely amend, and it is one reason the operative claim language at publication is not a reliable guide to the claim language that would eventually issue.

The cohort context

A single application rarely tells you much on its own. The surrounding publications tell you more. Stryker has several other applications in the same recent publication window, and they cluster. US20260195867A1, SYSTEMS, DEVICES, AND METHODS FOR MEDICAL IMAGE ENHANCEMENT USING MACHINE LEARNING, published July 9, 2026, sits directly adjacent — the same imaging-AI lane, one step earlier in the pipeline. US20260195933A1, SYSTEMS AND METHODS FOR REAL-TIME PROCESSING OF MEDICAL IMAGING DATA UTILIZING AN EXTERNAL PROCESSING DEVICE, also published July 9, is directed at moving surgical image processing onto external compute — the infrastructure question that a generator-in-the-loop workflow raises immediately.

Alongside those, US20260199018A1, on visually guiding bone removal during a joint procedure, published the same day as the hero application; US20260206113A1 covers maximizing the output of surgical lights; and US20260185929A1 is directed to characterizing fluids flowing through a conduit using optical emitters and detectors. The mix is consistent: optical capture hardware on one side, software that interprets what the optics return on the other. US20260203900A1 is the far end of that arc — the point at which the interpretation stops describing the present and starts projecting forward.

The classification data supports the same read. The application is classified under G06T 7/0012 for medical image analysis, A61B 5/7267 and A61B 5/7275 for machine-learning-based diagnosis from physiological measurements, and G16H 50/20 for computer-aided diagnosis — health-informatics territory rather than device-mechanics territory. For a company whose revenue base is instruments and implants, a claim set that lands in G16H is a signal about which direction the value is being built in. What it does not tell you is timing: a published application is a disclosure, and disclosure is not a product, a clearance, or a launch.