Abstract
Medical image segmentation systems require not only high average accuracy but also robustness against rare prediction failures that can adversely affect downstream clinical decision-making. Test-time augmentation (TTA) is widely used during inference and is often assumed to improve prediction reliability; however, its practical impact on segmentation failure modes and deployment robustness remains poorly understood.
In this work, we perform a systematic evaluation of standard geometric TTA across three segmentation models spanning different performance regimes and two medical imaging tasks: brain tumor segmentation in MRI and polyp segmentation in colonoscopy images. We assess TTA using overlap (Dice), boundary (BF1), and surface-distance (HD95) metrics, complemented by statistical significance testing, view-count ablations, and qualitative analyses.
Our results show that TTA primarily suppresses extreme spatial errors rather than consistently improving average overlap performance. The most consistent benefits are observed in HD95, indicating reductions in rare boundary failures, while improvements in Dice are often limited for strong pretrained models. In contrast, larger gains are observed for lower-performing models, suggesting that TTA is most effective for models with greater residual error. Furthermore, performance improvements exhibit rapidly diminishing returns, with four-view TTA capturing nearly all measurable benefit.
These findings suggest that the practical value of geometric TTA lies in mitigating segmentation failures and improving prediction reliability rather than serving as a general-purpose accuracy enhancement technique. Our study provides empirical insight into when geometric TTA is most beneficial and offers practical guidance for its deployment in medical image segmentation systems.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/UNSURE2026_003.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=~Subhash_Kashyap1
BibTex
@InProceedings{KasSub_Beyond_MICCAISAT2026,
author = { Kashyap, Subhash},
title = { { Beyond Dice: A Reliability-Oriented Evaluation of Geometric Test-Time Augmentation in Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17260},
month = {pending},
page = {pending}
}