Abstract
Segment Anything Model 2 (SAM2) can segment 3D medical volumes without training by propagating a single prompt across slices, but accuracy degrades with distance from the prompted slice. Two approaches have been proposed independently: bidirectional propagation and periodic re-prompting. We ask how they interact, evaluating SAM2 zero-shot on the whole-tumor region of BraTS2021 ($n=150$) across a grid of propagation directions, seed criteria, and re-prompt intervals, under a pre-specified analysis with equivalence testing. At a moderate re-prompt interval, bidirectional and forward-only propagation are statistically \emph{equivalent}: once corrections are frequent, propagation direction is immaterial. But the equivalence holds only at frequent re-prompting. As re-prompts grow sparse – the regime that matters when annotation is costly – bidirectional propagation degrades far less, opening a gap of up to $0.14$ Dice. We further show that forward-only propagation from an interior seed forfeits accuracy to a structural coverage limit and, in a pilot ablation ($n=30$), that retaining SAM2’s memory bank across re-prompt boundaries silently negates re-prompting entirely. Bidirectional propagation does not add accuracy on top of frequent re-prompting; it makes infrequent re-prompting viable.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BrainWorks_017.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/profile?id=~Sheikh_Ziauddin1
BibTex
@InProceedings{ZiaShe_Bidirectional_MICCAISAT2026,
author = { Ziauddin, Sheikh AND Ahmed, Mirza Tahir},
title = { { Bidirectional Propagation Makes Sparse Re-Prompting Viable in Zero-Shot SAM2 for Whole-Tumor Glioma Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17254},
month = {pending},
page = {pending}
}