Abstract
Promptable SAM-family models are promising for interactive medical segmentation, but their robustness under ultrasound degradation remains under-characterized. We benchmark nine models on three public ultrasound datasets (BUSI, CAMUS, and DDTI) using 12 clinically motivated corruptions, clean L0 plus corrupted levels L1–L9, and three prompt modes: bounding box, point, and point+bbox. Corruption reduces mean Dice from 0.710 to 0.679 and increases mean HD95 from 51.41 to 54.91 pixels. MedSAM2 has the strongest absolute accuracy, yet in the dataset-balanced bbox table it drops from first on clean images to fourth under corruption; image-pooled rank stability assigns nonzero rank-1 probability only to SAM3 (0.77) and UltraSAM (0.23). Bbox prompting is the most reliable interaction mode, but perturbation analysis shows that its advantage depends strongly on box tightness. Pixelation and salt-and-pepper noise are most damaging. Overall, clean accuracy and corruption stability rank models differently, highlighting the need to evaluate model choice and prompt design under clinically plausible degradation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CLiMeM_011.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{NguTra_How_MICCAISAT2026,
author = { Nguyen, Tran Quoc Duy AND Nguyen, Minh Quan AND Nguyen, Trung Thanh AND Nguyen, Thanh-Huy AND Nguyen, Thien AND Liu, Xiaofeng AND Bagci, Ulas},
title = { { How Resilient are Foundation Segmentation Models to Noisy Ultrasound Artifacts? A Multi-Dataset Benchmark Across Prompts, Noise Types, and Severity Levels } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}
