Abstract
Accurate image registration is central to image-guided interventions, yet the choice of evaluation metric can substantially influence how registration quality is interpreted, because region overlap, boundary alignment, and target localization capture different error dimensions. We use MR-TRUS prostate interventions as a concrete application to quantify the relationships, agreement, and residual disagreement among Dice similarity coefficient (DSC), target registration error (TRE), and 95th percentile Hausdorff distance (HD95). TRE reflects landmark-based local registration error that is relevant to target-directed tasks such as biopsy or focal therapy, but is often unavailable because landmark annotation is labor-intensive. This work addresses three questions: (1) how strongly DSC, TRE, and HD95 are correlated and agree at the case level; (2) whether metric choice changes the comparison of registration methods; and (3) whether DSC and HD95 can approximate landmark-based TRE when landmarks are unavailable. Seven deep learning-based registration methods were evaluated using a 108-case MR-TRUS prostate registration cohort. We analyzed case-wise metric relationships using Pearson and Spearman correlations and pairwise agreement with Cohen’s kappa. We further assessed whether metric choice changes the comparison of registration methods by ranking the seven methods using mean DSC, TRE, and HD95, and by comparing the best-performing method selected by each metric for each case. Linear regression models were also fitted to predict landmark-based TRE from DSC and HD95. The metrics showed significant associations across methods, with strong-to-very-strong Pearson correlation coefficients (PCCs) between all metric pairs (, all ). However, agreement was imperfect. Cohen’s kappa ranged from to , and metric pairs involving TRE showed weaker agreement than DSC-HD95. When methods were ranked by mean DSC, TRE, and HD95, the top two methods remained consistent across metrics, but lower-ranked method ordering and case-wise best-method selection changed with the metric. DSC and HD95 also approximated TRE with across individual methods and a pooled RMSE of mm over method-case predictions. These findings suggest that DSC, TRE, and HD95 are correlated but not interchangeable, supporting task-oriented multi-metric evaluation when landmark-based TRE is unavailable.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ASMUS_057.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=fmNpAP2UPJ
BibTex
@InProceedings{LiuMen_Medical_MICCAISAT2026,
author = { Liu, Mengting AND Yang, Mengting AND Ma, Shixing AND Hu, Yipeng AND Min, Zhe},
title = { { Medical Image Registration Metrics Examined for TRUS-Guided Prostate Interventions } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}