In an interventional suite, fluoroscopy runs continuously while a clinician threads a device through a patient's vasculature under live X-ray. The clinician's hands are near the field because that is where the work is. Over a career, the cumulative dose to those hands is a genuine occupational-health problem, and the conventional answer has been procedural — lead gloves, better positioning, discipline about where you put your fingers. An application published on 30 July 2026 and assigned to Canon Kabushiki Kaisha proposes that the machine simply notice.

The record is US20260221372A1, naming Marco Razeto, Yvonne Belton, Corne Hoogendoorn and James Matthews, classified under CPC H01J 35/14 and a cluster of G01T radiation-measurement codes.

A system receives captured images of the object and supplies these to an image recognition model to identify whether or not an unintended part of an operator (e.g. a hand) is present within the path of an X-ray beam.— System and Method for Adjusting Fluoroscopy Imaging Conditions, US20260221372A1

How it sees, and what it does

The detection input can come from either of two places. Claim 1 recites a series of fluoroscopy images captured by the X-ray apparatus itself, or a series of optical images captured over the same period — an ordinary camera watching the table. That alternative matters: the optical route can spot a hand approaching before it appears in the X-ray image, because a hand only shows up in a fluoroscopy frame once it is already being irradiated.

Those images go to at least one image recognition model, which claim 3 specifies may be a machine learning model and claim 4 narrows to a convolutional neural network. On a detection, the system reduces radiation by adjusting one or more of three things: the collimation of the beam, which limits the irradiated area; the presence of a filter between the source and the object; and the imaging parameters. Claim 2 spells the last of those out as reducing the tube current or reducing the frame rate.

Claim 5 is the more surgical version. The model outputs the location of what it detected, and the collimation is then adjusted to prevent at least part of that location from being irradiated. Rather than dimming the whole exposure, the system reshapes the beam around the hand. Claim 6 adds the necessary constraint — the collimation change is made on condition that the region of the object being examined continues to be irradiated, so the system does not protect the operator by blinding the clinician. Claims 7 and 8 identify that region using a further recognition model, one that finds it by identifying an instrument in the image.

Giving the picture back

Reducing dose costs image quality, and the second half of claim 1 is about undoing that. Once a predefined condition is met, the system further adjusts conditions to increase the radiation reaching the detector. Claim 15 lists what can satisfy that condition: a failure to detect the operator's part in subsequent images, a change in the capture angle, user input, or simply the expiry of a time limit. The default is that the machine restores itself when the hand leaves, without anyone touching a control.

Claims 9 and 10 handle the interval in between. Images captured while the dose is reduced are post-processed with a denoising filter or a noise-reduction machine learning model, and claim 10 gates that work on measurement — identify a region of interest, compute an image quality parameter, and apply post-processing to subsequent frames only if that parameter falls below a threshold. Cheap when the picture is fine, expensive only when it is not.

The cleverest claim in the set is claim 12, which distinguishes an ungloved hand from a gloved one and performs a different action for each, with claim 14 assigning a separate recognition model to each case. Claim 13 says what the gloved-hand action is: preventing an automatic exposure control system from increasing the intensity of the X-rays. That addresses a genuinely counterintuitive failure mode. Automatic exposure control raises output when the detector signal drops, and a dense object in the beam — such as a leaded glove — makes the signal drop. Left alone, the safety equipment provokes the machine into producing more radiation. This claim stops that loop rather than dimming the beam. It is also a neat illustration of why detection alone is not the invention: knowing a hand is present is only useful if the system knows which of several competing control behaviours to suppress, and a gloved hand and a bare one call for opposite responses.

Two notes on reading the record precisely. The abstract and the claims do not use the same vocabulary: the abstract speaks of "an unintended part of an operator (e.g. a hand)" while every claim recites "a subject of a user". They should not be blended, and only claim 12 ties that subject specifically to a hand. And the published text carries several drafting slips reproduced here rather than repaired — claims 1, 16 and 17 read "adjust imaging conditions including one or more:" with no "of"; claim 1 refers back to "the series of fluoroscopy frames" having introduced "a series of fluoroscopy images"; claim 7 uses the permissive "may be" inside a claim; claim 14 reads "the least one image recognition model"; and claim 16 opens with "a series of captured images of a object".

The application is a published application, not a granted patent. It confers nothing enforceable in this form and its claims may narrow before any grant issues — but as a description of where dose management is heading, it is unusually concrete: not a shield, a perception problem.