Abstract
Prompt-guided segmentation models are commonly compared under a fixed localization protocol, although the amount and form of spatial information may vary across workflows. We investigate whether model rankings remain stable when this localization information changes. On 200 CT examinations from a retrospective cohort of immunocompromised patients, with one expert target-lesion annotation per scan and 30 dual-annotated cases, we compare zero-shot MedSAM, MedSAM+LoRA, crop-guided nnU-Net, and nnU-Net with a binary boxprior channel. Under dense reference-derived slice-wise boxes, nnU-Net with box prior achieved the best held-out performance: Dice 0.914 (95% bootstrap CI: 0.902–0.923), HD95 1.19 mm (1.06–1.39), and ASSD 0.38 mm (0.31–0.46). Its agreement with a second reader was of similar magnitude to reader-reader agreement (Dice 0.813 vs. 0.801). MedSAM+LoRA improved zero-shot Dice from 0.598 to 0.866 with 294,912 trainable adapter parameters. In 185-case development cross-validation, nnU-Net with box prior remained stronger under dense slice-wise boxes (0.894 vs. 0.840). However, with one global in-plane box over the known target z-extent, the ranking reversed (0.771 vs. 0.825), and the trained MedSAM+LoRA system was less affected by box-size and jitter shifts. Prompt representation and its alignment between training and evaluation are therefore part of the evaluated system, not neutral implementation details.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TIA_024.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{EboEl_PromptGuided_MICCAISAT2026,
author = { Ebou, El Haj Samitt AND Biswas, Suparna AND Mahiou, Yanni AND Duron, Loïc AND Fournier, Laure AND de Margerie-Mellon, Constance AND Martin, Garance},
title = { { Prompt-Guided Target-Lesion Segmentation of Infectious Pulmonary Nodules: Model Benchmarking and Localization-Effort Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17269},
month = {pending},
page = {pending}
}