多种子评估策略在多中心胸片结核病检测中的域泛化性能
Rigorous multi-seed evaluation of domain generalization strategies for multi-center tuberculosis detection in chest radiography
该研究评估了七种域泛化策略在四个公共胸片队列中的性能,每种配置使用五个种子进行训练,共140次运行。结果显示,各方法的平均AUROC差异不大,AdaMix方法在概率校准方面表现更佳。多种子测试和校准报告应成为域泛化声明的最低标准。
- Article
- Open access
- Published: 20 November 2026
Scientific Reports (2026) Cite this article
We’re sharing this article early to provide faster access to peer-reviewed, accepted research. It is citable and carries a permanent DOI. This version is subject to further edits and will be replaced automatically by the final Version of Record. All legal disclaimers apply.
Abstract
Deep-learning models for tuberculosis (TB) screening on chest radiographs often lose accuracy at new sites, yet domain generalization (DG) methods are usually compared using a single training run. We built a multi-seed leave-one-domain-out benchmark of seven training objectives (empirical risk minimization (ERM), MixStyle, CORAL, a feature-mixing variant (AdaMix), IRM, GroupDRO and a proposed hyperspherical prototype entropic assignment regularizer (HP-EAR)) across four public chest-radiograph cohorts (1,280 images after near-duplicate removal). Each configuration was trained with five seeds (140 runs), giving 20 matched evaluations per method. Mean AUROC ranged only from 0.752 (ERM, CORAL, HP-EAR) to 0.767 (MixStyle). No method differed from ERM in paired tests or in mixed-effects models accounting for repeated seeds and cohorts (likelihood-ratio p = 0.59). Rankings were unstable: three different methods ranked first across the five individual seeds. AdaMix produced markedly better-calibrated probabilities (expected calibration error 0.232 versus 0.286–0.327; all Holm-adjusted p ≤ 0.048), driven by less over-confident predictions. Extending training from 12 to 30 epochs changed no conclusion. At the default threshold, mean sensitivity was 0.55–0.60 at unseen cohorts. Multi-seed testing, calibration and operating-point reporting should be minimum standards for DG claims in medical imaging.
Acknowledgements
The authors thank the contributors and maintainers of the Shenzhen, Montgomery, TB Chest X-ray Database and GlobalSubset repositories, and the reviewers and editor, whose comments substantially improved this work.
Funding
This research received no specific external grant from any funding agency in the public, commercial, or not-for-profit sectors.
Author information
Authors and Affiliations
Department of Information Systems, King Saud University, Riyadh, Saudi Arabia
Faisal Abdulaziz Almisned
Department of Nursing Faculty of Health Sciences, Bahçeşehir Cyprus University, K.K.T.C Mersin-10, Türkiye
Ibrahim Ibrahim Shuaibu
Authors
- Faisal Abdulaziz Almisned
- Ibrahim Ibrahim Shuaibu
Corresponding author
Correspondence to Ibrahim Ibrahim Shuaibu.
Ethics declarations
Competing interests
The authors declare no competing interests.
Ethical approval
This study used only pre-existing, publicly available, de-identified chest radiograph datasets, did not involve new participant recruitment, and was therefore exempt from institutional ethics review under standard policy for secondary analysis of anonymised public data.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Almisned, F.A., Shuaibu, I.I. Rigorous multi-seed evaluation of domain generalization strategies for multi-center tuberculosis detection in chest radiography. Sci Rep (2026). https://doi.org/10.1038/s41598-026-81727-z
Received: 18 August 2026
Accepted: 17 November 2026
Published: 20 November 2026
DOI: https://doi.org/10.1038/s41598-026-81727-z
Keywords
来源:Nature 健康护理主题 · nature.com