Abstract
Interactive image segmentation incorporates sparse user inputs, such as clicks, and is often used to correct an existing segmentation. In such corrective workflows, the underlying model should respond where the user intends to intervene while behaving predictably in regions that were not targeted. Yet these interaction-specific properties are rarely evaluated directly. In this work, we study corrective interactive segmentation from a behavioral perspective and focus on two complementary questions: how well a click improves the intended target region, and how strongly the model alters remote parts of the segmentation that were not directly indicated by the user. To this end, we define a set of behavior-oriented evaluation measures that quantify local responsiveness and distant preservation beyond standard overlap metrics. We use these measures to compare several Dice-based training objectives with different spatial emphasis around click locations, and include a simple preservation-aware objective as a test case for explicitly controlling remote edits. Across multiple analyses on cardiac CT, stronger local weighting alone is found to produce little change in model behavior compared with standard Dice training. In contrast, the preservation-aware objective leads to markedly more localized updates and greatly reduces remote mask modifications. Further analysis shows that distant edits are more often beneficial than harmful, but that their benefit decreases as the initial segmentation quality improves, revealing a practical trade-off between broader automatic correction and conservative, trustworthy user interaction behavior.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/HAIC26_004.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=iEblzWmvnN
BibTex
@InProceedings{KovDan_Characterizing_MICCAISAT2026,
author = { Kovacs-Deak, Daniel AND Meyer, Maria Ines AND Lambrechts, Adriaan AND Wirix-Speetjens, Roel AND Maes, Frederik},
title = { { Characterizing Global and Local Model Behavior in Interactive Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17279},
month = {pending},
page = {pending}
}