Ultrasound strengthens routine diagnosis and guidance across specialties, but results can vary with operator skill and image quality. A new approach brings large-scale collaboration to ultrasound AI without moving patient data. UltraFedFM is a privacy-preserving foundation model trained across 16 institutions in 9 countries on more than one million ultrasound images spanning 19 organs and 10 modalities. By learning broad image patterns first, then adapting to specific clinical tasks, the model aims to deliver consistent decision support while respecting data protection rules. Reported performance indicates high diagnostic accuracy and precise lesion delineation across diverse tasks. In direct comparisons on multiple systemic diseases, the model exceeded the accuracy of mid-level ultrasonographers and performed on par with expert readers, highlighting potential value for clinical workflows that rely on fast, accessible imaging.
Collaborative Training Without Data Sharing
UltraFedFM uses federated learning so that participating sites train locally on their own datasets and share model updates rather than raw images. This structure allows cross-institution training on varied organs and modalities while keeping patient data on site. The development process combines large-scale, self-supervised pretraining with later task-specific adaptation for screening, disease classification, maternal-fetal assessment and segmentation. The training pipeline accounts for ultrasound’s quirks, including noise and artefacts and incorporates design choices that help the model learn from different probes and acquisition styles. The federated setup supports regular updates as new data become available, reflecting the changing mix of scans encountered in practice. In settings where privacy, governance and interoperability limit centralised pooling, this approach enables broad collaboration without weakening data safeguards.
Must Read: AI mpUS Maps Clinically Significant Prostate Cancer
Strong Results Across Diagnosis and Segmentation
Across diagnostic tasks, UltraFedFM achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.927, indicating high discriminative accuracy over multiple organs and modalities. In disease classification scenarios, it maintained strong performance even when the amount of data available for fine-tuning was reduced. Specific task benchmarks illustrate the breadth of capability: fetal plane recognition reached high accuracy, gallbladder cancer classification achieved high performance, and joint multi-organ diagnosis remained consistent where human accuracy can vary by organ. When compared with clinicians on eight systemic malignant diseases, the model outperformed ultrasonographers with 4–8 years of experience and matched those with more than 10 years.
Segmentation results were similarly strong. The model produced precise contours for breast and thyroid nodules, with reported dice similarity coefficient (DSC) values that align with robust pixel-level performance. It also delivered accurate measurements in obstetric assessment, including the angle of progression. Notably, segmentation quality remained competitive when only a fraction of fine-tuning data was used, pointing to efficient use of labels. These results indicate utility for tasks that demand consistent boundaries and measurements, where variability can affect downstream decisions such as monitoring, intervention planning and follow-up comparisons.
Robustness, Generalisation and Scale
UltraFedFM demonstrated performance that carried over to new institutions and datasets not involved in development, including scenarios with different imaging characteristics. Reported metrics on external sets showed high AUROC, with strong results on high-frequency ultrasound despite textural and colour differences. Analyses under acquisition imbalances and noise perturbations suggested stable predictions with reduced dispersion, indicating resilience to issues such as motion, operator technique and artefacts. These properties are important for ultrasound, where variability is common and can hinder model reliability.
Scaling analyses linked larger pretraining datasets to gains on downstream tasks, especially for pixel-level segmentation. Model size also contributed to improvements on more challenging tasks. In federated environments where data distributions differ by site, convergence and fairness were examined, and weighting strategies helped limit performance gaps across clients. Together, these findings suggest that expanding the collaborative pool of images and refining aggregation methods can further improve consistency across organs, modalities and clinical settings. At the same time, reported limitations included the need for broader domain coverage, larger clinician evaluation cohorts, and future deployment in fully decentralised clinical pathways.
A privacy-preserving, federated approach to ultrasound AI can deliver expert-level support across diagnosis and segmentation while adapting to diverse scans, organs and clinical uses. UltraFedFM was trained across multiple institutions and countries without centralising data, achieved high AUROC for disease diagnosis and produced accurate contours and measurements that held up under reduced labels and external validation. Comparative testing showed performance on par with expert readers and above mid-level practitioners, and analyses indicated stability under noise and acquisition differences. Further work to extend clinical domains, assess at larger scale and embed within live, decentralised workflows will help define how such systems can bolster ultrasound practice, supporting timely decisions and consistent care without compromising patient privacy.
Source: npj digital medicine
Image Credit: iStock
