In portable prenatal ultrasound, the bottleneck is the frame, not the image

In low-cost obstetric ultrasound, the gating question is not whether the picture is clear but which single frame of a 20-second sweep is the one a fetal-growth measurement is supposed to come from. Operationally, a frontline operator with a portable scanner and no specialist on hand is choosing between a few good-looking frames, and that choice decides whether the measurement can be trusted.

The AnatoProto preprint attacks that choice directly. It adapts a frozen biomedical vision model with two cheap additions: an anatomy-weighted spatial prior that reweights where the model looks, and a within-case prototype pull that drags each frame embedding toward the mean of the sweep's positive frames. On ACOUSLIC-AI, the adapted model lands at F1 = 67.72, against 54.52 for the strongest foundation-model baseline. The 13-point gap is, more usefully, a sign that frame selection in this setting rewards spatial priors over bigger encoders.

The mechanism is reusable. Any domain where a portable device captures a noisy sweep and a downstream measurement needs one frame inherits the same bottleneck, and the same answer: anchor the model to anatomy, then let the sweep vote with itself. The authors' own ablation is the falsifier: prototype loss alone drops recall by 12 points. The gain depends on the anatomy prior being roughly right, and lives on a single public benchmark with no clinical-outcome validation. The portable scan got a more reliable second opinion. The specialist is still the specialist.