Abstract
Bone age assessment from pediatric hand radiographs is widely used to evaluate skeletal maturity, but most deep learning methods provide only point estimates without indicating reliability. This limitation is critical in clinical and forensic settings, where incorrect but overconfident estimates can affect downstream decisions. In this work, we present a unified benchmark of uncertainty quantification methods for bone age regression on the RSNA Pediatric Bone Age dataset. We train multi-input models that combine hand radiographs with sex information and evaluate three backbones: EfficientNet-B3, ConvNeXtV2-Tiny, and ViT-B/16, allowing us to assess whether uncertainty quantification performance is model-agnostic. Across ten random seeds, we compare six uncertainty quantification strategies: vanilla conformal prediction (CP), k-Nearest Neighbor-based CP (kNN-CP), Random-Forest CP (RF-CP), Conformalized Quantile Regression (CQR), Monte Carlo Dropout (MCD), and Bayesian Neural Network (BNN). Performance is assessed using mean absolute error, prediction interval coverage probability, prediction interval normalized average width, and coverage-width criterion at 90%, 95%, and 99% desired coverage. EfficientNet-B3 and ConvNeXtV2-Tiny provide the strongest point prediction accuracy, while conformal prediction methods offer the most reliable coverage-width trade-off. RF-CP provides the best coverage–width trade-off by producing adaptive prediction intervals with reliable empirical coverage. Vanilla CP and kNN-CP remain competitive, whereas MCD tends to under-cover, and BNNs produce overly conservative intervals. The source code is available at https://github.com/bilgedogan/boneage_uncertainty
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/UNSURE2026_077.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=~Mümin_Emir_Soydal1
BibTex
@InProceedings{DoğNad_UncertaintyAware_MICCAISAT2026,
author = { Doğan, Nadide Bilge AND Işıklan, Fuat AND Soydal, Mümin Emir AND Kumbasar, Tufan AND Öksüz, İlkay AND Ekenel, Hazım Kemal AND Töreyin, Behçet Uğur},
title = { { Uncertainty-Aware Bone Age Assessment: Analysis of Conformal and Bayesian Methods } },
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}
}