Healthcare Computer Vision Development Services

Computer vision reads medical images and video the way a trained eye does: it finds, counts, measures and compares. Common uses are counting cells on slides and smears, measuring wounds and skin lesions from photos, flagging findings on X-ray, CT and MRI to prioritise a worklist, grading retinal photos and recognising instruments in surgical video. In each case the model supports the clinician or the technician, who makes the final call.

Medical image analysis turns an image into numbers a clinician or a lab can act on. It covers detection (where a structure is), segmentation (its exact outline), classification (what it is) and measurement (how big it is and how it changes). Medical image analysis software wraps those models in a viewer, a report and a workflow, so the result reaches the person who needs it.

It depends on how rare and how subtle the thing you look for is. A feasibility check can start with a few hundred labelled images and tells you whether the task is learnable and how much more data a production model needs. Data from more than one site matters more than sheer volume.

No one can promise a number in advance: it depends on the task, the images, the quality of the labels and how different the next site’s images are. What you should expect is a report per site, on patients the model never saw, that shows how many findings it caught, how many of its flags were right and where it fails. In Smear reader, finding red cells held at 96–99% in six labs the model never saw, while flagging infected cells varied far more from lab to lab.

We test it on data from sites it never saw, without retraining, and report the results per site. In our own research, finding cells carried over to six other hospital labs, while spotting infection carried over much less reliably. That is the kind of result you want to know before launch, not after.

Yes. We can work inside your environment, on your servers or cloud account with access you control, or with de-identified exports. The finished model can also run on your hardware, in the browser or on the device, so images never leave it. We sign an NDA before any data moves, and a BAA when protected health information is involved.

Yes. Small models can run in the browser with WebGPU or WebAssembly, in a mobile app or on the device itself, so images never leave it. Smear reader, our own research project, runs two models, 4.2 MB together, in a browser tab.

You do. Your images and labels stay yours, and after the project the trained model, its weights and the source code are transferred to you, as with any software we build.

It depends on whether labelled data already exists, how many sites it comes from and where the model has to run. Training itself is rarely the long part: in our Smear reader research, the detector trained in 12 minutes and the classifier in 3.5, on a laptop. Most of the time goes into labelling rules, data checks and validation on another site, and a feasibility check on your images turns that into a concrete plan.

It depends on the data you already have, how many sites the model must be validated on and whether the software will be a regulated medical device. We start with a research stage: for $20,000–$50,000 you get a preview model trained on a sample of your images, a proof of concept that shows whether the task is feasible and what performance is realistic. From there the work is growing the dataset and tuning accuracy.

A production model typically costs $90,000–$180,000+ for non-diagnostic uses such as wound tracking, and $150,000–$300,000+ for a medical imaging model in radiology, pathology or dermatology, with dataset curation and an internal validation study. A regulated diagnostic product costs $300,000–$500,000+. For a first estimate of your case, use our cost calculator.

Both. The methods are the same; what changes is the input. X-ray, computed tomography (CT) and magnetic resonance imaging (MRI) arrive as DICOM files, CT and MRI as 3D volumes, and we use tools built for them, such as MONAI. Our published research, Smear reader, is on microscopy; for any image type we start with a feasibility check on your own data.

If it informs a diagnosis or treatment, it usually is, and the intended use decides. Take our free SaMD Check for a first view of how the FDA may classify your software, or contact us to discuss your case.