Every Pixel Is an Option: A System One Model for Medical Ultrasound Biometry

Every Pixel Is an Option. A System One Model for Medical Ultrasound Biometry. By Moein Heidari and eight others, from the University of British Columbia and the Rutgers Cancer Institute. Published on the second of November.

Ultrasound biometry is how clinicians get quantitative measurements out of an ultrasound image, in prenatal and cardiac examinations among many others. The measurements come from calipers placed at anatomical landmarks. Because that placement is still largely manual, the numbers vary from one operator to the next. Automated approaches have their own limits. Conventional methods often build a specialized prediction head for each ultrasound task. Vision-language models are more flexible, since they can be steered with text instructions, but they tend to predict coordinates directly. Direct coordinate prediction does not explicitly model a spatial likelihood distribution, so it cannot represent how ambiguous a landmark's location might be.

Kev-Bio is a vision-language model that uses the clinical measurement question itself to guide landmark detection. The title gives the stance: every pixel is an option. Instead of naming one coordinate, the model produces a probability over the image for where each caliper endpoint should go. The design takes its inspiration from System One decision models.

Concretely, Kev-Bio takes decision representations from its language model and uses them to generate a query for each caliper endpoint. Those queries are matched against dense ultrasound features that are shared across the endpoints, and the matching produces spatial probability maps. Landmark coordinates are read off those maps, and the clinical measurements are derived from the coordinates. Training is end-to-end, and it uses two kinds of loss together: one for landmark localization and one for the clinical measurement.

Evaluation is on the FU/GU Biometry 2026 dataset, which covers nine views. Kev-Bio reaches a mean radial error of twenty-two point five zero pixels, and a mean absolute error on clinical measurements of fifteen point three eight. The abstract reports that this outperforms the compared baselines in overall performance.

The shift is from asking a model for an answer to asking it for a map. A measurement question goes in as text, a probability map comes out for each endpoint, and the coordinates and the measurement are both derived from those maps. The authors conclude that the approach is effective for both landmark detection and clinical measurement across different anatomical views, which is the flexibility a single text-guided model was meant to offer in the first place.