1Macao Polytechnic University, Macao, China 2Netherlands Cancer Institute, Amsterdam, The Netherlands 3Zhejiang Cancer Hospital, Hangzhou, China 4The First People’s Hospital of Hangzhou, Affiliated Hangzhou Hospital of Nanjing Medical University 5Department of Radiology, University of Pittsburgh, PA, USA 6College of Information Science and Technology, Jinan University, Guangzhou, China 7Case Western Reserve University, USA 8School of Biomedical Engineering, Shenzhen University, Shenzhen, China 10University of Pittsburgh Medical Center, Pittsburgh, PA, USA
Abstract
Modern ultrasound systems are universal, economically cost-effective, and portable diagnostic tools capable of imaging the entire body. However, current artificial intelligence (AI) solutions remain fragmented into single-task, organ-specific tools. This critical gap between hardware versatility and software specificity limits workflow integration and the clinical utility of AI in general ultrasonography. To address this, we organized the Universal UltraSound Image Challenge 2025 (UUSIC25). Participants developed algorithms using a training set of 11,644 images aggregated from 12 discrete sources. Algorithms were evaluated on a fully independent, multi-center private test set of 2,479 images, composed of held-out internal samples and a cohort from an external center completely unseen during training to assess generalization. The top-ranking algorithm (SMART) achieved a macro-averaged DSC of 0.854 across 5 segmentation tasks and a macro-averaged AUC of 0.766 for binary classification tasks. While models demonstrated high capability in anatomical segmentation, performance variability was observed in complex diagnostic tasks subject to domain shift. General-purpose AI models can achieve high diagnostic accuracy and efficiency across multiple ultrasound tasks using a single network architecture. However, significant performance degradation on data from unseen institutions suggests that future development must prioritize domain generalization techniques before clinical deployment is feasible.
Challenge Design and Data
Figure 1. Study Flow Diagram. Data collection, source attribution, and stratification logic. The diagram illustrates the explicit separation of data streams: public datasets were utilized exclusively for training to promote generalization (n=10,010), while internal private data were stratified across all sets (n=5,499). Data from the external center (NKI, n=512) served as a strictly held-out test set.
Key Findings
Question: How effectively can a single general-purpose deep learning model handle multi-organ segmentation and classification tasks in clinical ultrasound?
Findings: In the UUSIC25 challenge involving 15 algorithms and 16,021 images across 7 anatomical regions, the winning query-driven Transformer model achieved high diagnostic accuracy (e.g., 0.942 Dice for fetal head segmentation, 0.837 AUC for breast malignancy) and efficiency. Notably, these unified models demonstrated robust generalization on a fully private, multi-center test set containing data from a completely unseen institution.
Meaning: These results suggest that developing high-performing, "all-in-one" clinical ultrasound AI systems is feasible, moving beyond the fragmented single-task paradigm; this paves the way for next-generation AI assistants that can streamline workflows and adapt to diverse clinical scenarios without manual intervention.
Results
Global Landscape and Performance Benchmarking
Figure 2. (a) Global Participation & Data Integration Network spanning five continents. (b) Methodological Configuration Matrix summarizing architectural choices for the top-10 teams. (c) Multi-Organ Diagnostic Versatility and Efficiency Pane showing segmentation (DSC) and classification (AUC) performance across anatomical regions. The winning model (SMART) demonstrates consistent coverage across tasks. (d) Diagnostic Precision (ROC Curves) for Breast Malignancy and Fatty Liver tasks on the private test set.
Publications
Diagnostic Performance of Universal-Learning Ultrasound AI Across Multiple Organs and Tasks: the UUSIC25 Challenge
Zehui Lin, Luyi Han, Xin Wang, Ying Zhou, Yanming Zhang, Tianyu Zhang, Lingyun Bao, Jiarui Zhou, Yue Sun, Jieyun Bai, Shuo Li, Shandong Wu, Dong Ni, Ritse Mann, Wendie Berg, Dong Xu, Tao Tan and the UUSIC25 Challenge Consortium
@article{lin2025diagnostic,
title={Diagnostic Performance of Universal-Learning Ultrasound AI Across Multiple Organs and Tasks: the UUSIC25 Challenge},
author={Lin, Zehui and Han, Luyi and Wang, Xin and Zhou, Ying and Zhang, Yanming and Zhang, Tianyu and Bao, Lingyun and Wu, Shandong and Xu, Dong and Tan, Tao and others},
journal={arXiv preprint arXiv:2512.17279},
year={2025}
}
Acknowledgements
This work was supported by the Science and Technology Development Fund, Macau SAR (File no. 0004/2025/ASJ) under the FDCT-FAPESP Joint Funding Scheme; the Shenzhen Medical Research Fund (Grant No. D2501013); and the Macao Polytechnic University Grant (Grant No. RP/FCA-17/2025). We thank the organizing committee of MICCAI 2025 for hosting the challenge.
