Abstract
Promptable foundation models like SAM, MedSAM, MedSAM-2 and SAM3 promise to make medical ultrasound simpler: draw a box, get a mask, with no training. But thyroid benchmarks usually hand the model an oracle box taken from the ground-truth mask, assuming the user already knows where the nodule is, the opposite of simplifying. We benchmark four foundation models against the three fully-supervised CNNs they are meant to replace (U-Net, nnU-Net, TransUNet) on three public thyroid ultrasound datasets, reading every score against a trivial box-fill floor that just returns the prompt box (DSC $0.765$ to $0.851$). Zero-shot, MedSAM-2 clears the floor on every dataset (up to $0.907$ DSC), while vanilla SAM ViT-H falls below it and even the newest model, SAM3, is not separable from it on TN3K. Given the same box, a box-conditioned CNN wins: the nnU-Net ceiling ($0.914$ to $0.950$ DSC) beats every zero-shot foundation score, and LoRA fine-tuning (best: SAM3, $0.898$ to $0.935$) narrows the gap without closing it while reversing the zero-shot ordering, in which the medically retrained models lead. Finally we remove the human-drawn box: a learned YOLO detector proposes it automatically (detection rate above $0.95$), turning the benchmark into a no-human-in-the-loop pipeline at a cost of $0.03$ to $0.13$ DSC, largest for the model that led with an oracle box. For thyroid ultrasound, foundation promptability buys convenience, not accuracy: what it simplifies is the prompt, once a detector replaces the human, not the segmentation itself.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ASMUS_021.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=YD4QcIMdM6
BibTex
@InProceedings{MedRod_From_MICCAISAT2026,
author = { Medellin-Robles, Rodrigo},
title = { { From Oracle Boxes to Automatic Prompts: Benchmarking Frontier Foundation Models for Thyroid Ultrasound Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}