Ask a surgeon closing a mastectomy flap what she most wants to know, and the answer is not what the tissue looks like on the table. It is what that tissue will look like in three weeks — whether the skin she is stitching will heal cleanly, heal slowly, or die. A patent application published on July 16, 2026 proposes an unusual way to answer that question: stop scoring the tissue, and instead have a machine-learning model draw the photograph of the outcome before the outcome happens.

The application is US20260203900A1, titled METHODS AND SYSTEMS FOR CHARACTERIZING TISSUE OF A SUBJECT. It names Lina Gurevich as inventor and is assigned to Stryker Corporation. It is classified under G06T 7/0012 and A61B 5/7267 alongside G16H 50/20, a combination that places it squarely between medical image analysis and clinical decision support. It is an A1 — a published application, pending examination, not a granted patent — and everything below describes what it discloses and claims, not what any office has allowed.

Start with the problem. Fluorescence imaging is already routine in reconstructive and wound surgery: a fluorescent agent circulates, a near-infrared camera watches where it goes, and the resulting image encodes perfusion — which parts of the tissue are getting blood. The dependent claims here name exactly those inputs. Claim 8 narrows the input to a NIR image; claim 9 to an intraoperative perfusion image; claim 10 to a frame drawn from an intraoperative perfusion video. What the surgeon actually gets from that hardware is a grayscale or false-color brightness map, and what she needs is a prognosis. Bridging the two has traditionally meant experience, rules of thumb, or a numerical perfusion threshold.

A translator, not a classifier

Claim 1 of this application is canceled, so the first operative claim — and the independent method claim — is claim 2. It recites a four-step loop. Receive a fluorescence image of the subject's tissue. Provide that image to the generator of a trained generative adversarial model. Obtain back from that generator a simulated white-light image depicting the predicted future state of the tissue. Display it. Claims 20 and 21 mirror the same four steps as a system claim and a non-transitory computer-readable-medium claim.

The engineering pivot sits in step two, and it is worth being precise about. A generative adversarial network is trained as a pair: a generator that synthesizes images and a discriminator that tries to tell synthesized images from real ones. Once training ends, the discriminator has done its job and is usually discarded. Claim 2 recites the generator specifically, and recites it at inference time — the model is not being asked to segment the flap, to score it, or to output a probability of necrosis. It is being asked to perform image-to-image translation: take a perfusion image captured during the operation, and emit what an ordinary visible-light photograph of that same tissue would plausibly look like weeks later. The output is a picture, and the clinical content lives in what the picture depicts rather than in a number attached to it.

That framing is what makes the training data interesting, because the model can only learn that mapping if someone assembled the right corpus. Claim 12 recites training on a plurality of image pairs, and 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, US20260203900A1

In plain terms: photograph the same tissue twice, once under fluorescence while the patient is open and once under white light after recovery, and treat the two frames as an input-output example. That is a supervised setup, and claim 14 names the architecture that fits it — a pix2pix GAN model, the standard paired image-translation approach. The application also anticipates that such pairs are hard to accumulate at scale, since they require the same patient, the same tissue region, and a disciplined follow-up photograph. Claims 15 through 17 therefore cover training on unpaired image data — a set of intraoperative fluorescence images and a separate set of post-operation white-light images, collected across a plurality of patients, with no correspondence between them — and claim 18 names CycleGAN, the architecture built for exactly that unpaired case. Claim 19 generalizes further to first and second imaging modalities.

The scope of "future state" is defined in the dependents rather than left to interpretation. Claim 3 recites that the future state comprises necrosis, delayed healing, healing, or any combination thereof. Claim 11 lists the tissue types: breast tissue, burnt tissue, chronic wound tissue, acute wound tissue, or skin transplants — the reconstructive and wound-care territory where perfusion imaging is already deployed. Downstream, claim 4 adds providing a recommendation based on the predicted future state, claim 5 narrows that recommendation to administering a treatment, and claim 6 feeds the white-light image into a separate classification model to identify complications. So the disclosed system can be read as a two-stage pipeline: generate the predicted photograph first, then optionally let a conventional classifier read it.

What the drafting shows

Two features of the text as published are worth noting, because they are in the record and a careful reader will hit them. First, the phrase "displaying, on the display" appears in the abstract and in all three independent claims, but neither the abstract nor claim 2 ever introduces a display in the first place. There is no earlier antecedent for "the display" to refer back to — the kind of antecedent-basis issue that examiners raise under 35 U.S.C. 112(b). Second, the recited model is called a "generative adversarial (“GAN”) model" throughout, with the parenthetical expanding to a word, "network," that never actually appears in the sentence. Both are in the published text exactly as written, and both are the sort of thing routinely resolved by amendment during prosecution. Neither changes what the disclosure is about, but both are reasons the abstract of this application should not be read as a clean statement of claim scope. For scope, read claim 2.

Context helps here. This is not an isolated filing: Stryker's recent publication cohort includes US20260195867A1, on medical image enhancement using machine learning, and US20260195933A1, on real-time processing of medical imaging data using an external processing device — both published the week prior, on July 9, 2026, and both sitting in the same imaging-AI cluster. Published the same day as this one, US20260199018A1 is directed to visually guiding bone removal during joint surgery, another visualization-in-the-OR disclosure. Adjacent hardware work appears in US20260206113A1, on maximizing the output of surgical lights, and US20260185929A1, on characterizing fluids flowing through a conduit with optical emitters and detectors. Taken together, the cohort describes a company filing across the whole optical chain in the operating room — the light, the sensor, the real-time compute, and now the inference drawn from the image.

The short version is that this application is directed to a prediction problem dressed as a rendering problem. Rather than asking a network to output a risk score a surgeon has to trust, it asks a generator to produce the artifact a surgeon already knows how to read: a photograph of tissue. Whether that framing survives examination is a separate question, and one the published record cannot answer — as of publication on July 16, 2026, US20260203900A1 is a pending application, and its claims may be amended before any of it issues.