Abstract
Accurate 3D pelvic tilt estimation matters in hip and spine surgery, where small errors can alter functional acetabular orientation. However, a single anteroposterior (AP) radiograph does not provide explicit depth information, which makes 3D tilt estimation challenging. To address this lack of depth information, we built a geometry-aware pipeline that predicts pelvic depth maps and anterior superior iliac spine (ASIS)/pubic tubercle (PT) heatmaps. Although it improves mean absolute error over direct regression from // to // (sagittal/coronal/axial), large sagittal tilt errors remain. Detecting these high-error cases requires assessing whether uncertainty is spatially aligned with depth error, which mean uncertainty (Unc. Mean) does not capture. We therefore revisit the area under the sparsification error (AUSE), a dataset-level metric that evaluates uncertainty estimators by measuring whether uncertainty ranks pixel-wise errors. Instead of averaging AUSE over cases, we compute case-wise ROI-AUSE by treating pixels in each pelvic depth ROI as samples. This score captures the mismatch between uncertainty and depth-error rankings, thereby serving as a reliability score for downstream pelvic tilt estimation. On 2290 AP radiographs, case-wise AUSE-RMSE detected sagittal angle-error outliers with AUROC 0.861/0.896/0.916 at IQR thresholds 1.0/1.5/2.0, whereas Unc. Mean reached only 0.529/0.534/0.556. At the IQR-1.5 threshold, a preliminary reference-free linear surrogate predicted AUSE from landmark-neighborhood uncertainty features and reached AUROC up to 0.669.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/UNSURE2026_070.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=~Daisuke_Watanabe2
BibTex
@InProceedings{WatDai_Beyond_MICCAISAT2026,
author = { Watanabe, Daisuke AND Gu, Yi AND Uemura, Keisuke AND Takao, Masaki AND Sugano, Nobuhiko AND Otake, Yoshito},
title = { { Beyond Mean Uncertainty: Case-wise AUSE for Detecting Failures in 3D Pelvic Tilt Estimation from a Single AP Radiograph } },
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}
}