Abstract
Reliable evaluation of medical image segmentation remains an open challenge. Although standardized overlap- and surface-based metrics are widely used, prevailing practices often rely on fixed aggregate summaries that obscure case-specific failure modes, robustness to perturbations, predictive uncertainty, and inconsistencies across anatomical structures or patient subgroups. Consequently, pipelines with similar overall scores may behave markedly differently in clinical or research settings. We propose a two-step framework for comprehensive and adaptive evaluation of segmentation pipelines. First, we introduce MedSegEvaluator, a modular evaluation library that extends conventional metrics with population-level and case-specific analyses. The framework supports perturbation-based robustness assessment, uncertainty quantification, inter-structure consistency analysis, and flexible aggregation strategies, enabling systematic and reproducible interrogation of segmentation outputs beyond leaderboard-style reporting. Second, we develop an agentic evaluation layer, AgenticMedSegEvaluator, built atop this library to enable adaptive evaluation. The agent dynamically orchestrates evaluation tools based on observed metric patterns, robustness signals, and case characteristics, generating structured, evidence-grounded reports. Experiments across 2D and 3D models and datasets demonstrate that agent-assisted evaluation uncovers failure modes and instabilities not evident from standard metrics alone, substantially influencing conclusions at both case and cohort levels.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_005.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=Z0OoyAj4gT
BibTex
@InProceedings{DatDar_From_MICCAISAT2026,
author = { Dathiya, Darshan AND Brunt, Keith AND Namdar, Khashayar},
title = { { From Metrics to Insight: Agent-Based Comprehensive Evaluation of Medical Image Segmentation } },
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}
}