Know the resolution of the instrument you trust

Every vision model has a smallest unit of attention. On a panoramic it is about three millimetres. Most of what I call is smaller.

I measured it on 31 August, on the model I trained on my own reads. One board-certified oral and maxillofacial radiologist, doing arithmetic any clinician can do in a minute. A vision model does not see pixels. It cuts the image into square tiles and turns each tile into one unit of attention. This model's tile is 32 pixels wide. My panoramics are 0.096 millimetres per pixel. One tile is 3.07 millimetres.

Now hold that ruler against the image. The crest of the alveolar bone, where periodontal bone loss begins, is a line a fraction of a millimetre wide. The wall of the mandibular canal is a pair of thin lines. The mental foramen is a small round shadow. Each lives inside one tile, sometimes inside a corner of one. The model can attend to the tile. It cannot attend to the line inside it.

More pixels feel like the fix. They are not. The largest panoramic reader published so far fed the same images to its model at three resolutions and the results moved by under two points (Nature Biomedical Engineering, 2026). Outlining every tooth on the image before asking about disease moved diseased-tooth identification by more than a fifth. Resolution was not the lever. Telling the model where to look was.

So the question for any AI that reads images for you, before any accuracy figure: what is its smallest unit of attention, in millimetres, on your images. If the findings you are paid to catch are smaller than that unit, no dataset will teach the model to see them. It will learn to say them from context instead, which is worse than silence, because it sounds the same.

What I would do differently: run this arithmetic before the first training example, not after the third run.

The check

The one-minute arithmetic:

Find the pixel size of your images. It is in the image header or the scanner specification.
Find the model's patch size. Vision models publish it. Some merge patches two by two before attending, so use the merged size.
Multiply. That is the smallest unit the model attends to, in millimetres.
List the findings in your work that are smaller than that number. Test those first, and test them by hand, against your own cases.

Sources

Nature Biomedical Engineering, 2026, DOI 10.1038/s41551-026-01713-8.