Loading…
Ultrasound foundation model pretrained by federated learning across 16 institutions, transferring to disease diagnosis and lesion segmentation.
No providers recorded yet. Browse all providers
Ultrasound is the imaging modality a clinic reaches for first, and the one whose images are hardest to pool. Scans sit inside departments under privacy rules that forbid export, so pretraining corpora have stayed small and lopsided — the 3M-US collection behind USFM is 91% breast imaging, and a model trained on it carries that skew into questions about a gallbladder or a fetal plane.
UltraFedFM answers this by never moving the images. Sixteen client institutions across nine countries each pretrain a local masked autoencoder on their own scans; after every communication round only the local parameters travel to a central server, which averages them into one global model and sends it back. Six hundred rounds later the averaged encoder is kept and the decoder discarded, leaving one fixed vision transformer backbone that has seen 1,015,754 unlabeled images spanning 19 systemic organs and 10 imaging modalities, including contrast-enhanced, endorectal and high-frequency ultrasound — without any hospital revealing a pixel to another.
Built at The Chinese University of Hong Kong, Shenzhen with collaborators across China, Ireland and Saudi Arabia, it appeared as a preprint in November 2024 and in npj Digital Medicine in 2025.
The local model is a masked autoencoder whose encoder is a ViT-Base — 12 transformer blocks with 768-dimensional embeddings — and whose decoder is a lighter ViT-Small of 8 blocks and 512 dimensions. Inputs are 224 × 224 with a 0.75 masking ratio, and pretraining ran for 600 communication rounds at batch size 512 with AdamW at a 1.5e-4 base learning rate, one local epoch per round, on eight V100 GPUs. Of the 1,015,754 pretraining images, 782,513 come from public collections and the rest from the authors' partner hospitals; the pool is sharded into 16 clients by institution.
Evaluation covers 11 subtasks drawn from 19 annotated datasets. Averaged over eight diagnosis tasks the model reaches 0.927 AUROC against 0.894 for USFM (p = 0.002), and it reaches 0.878 DSC on lesion segmentation; on the multi-class pubic symphysis–fetal head task it scores 0.842 DSC with a mean absolute error of 8.80 on the angle of progression. Two cohorts withheld from pretraining — kidney disease and high-frequency skin ultrasound — yield 0.925 average AUROC. Against seven sonographers reading 80 images spanning eight diseases, it exceeds the two with 4–8 years of experience and matches the five with more than 10.
Adaptation is by fine-tuning the released checkpoint, so the practical audience is any ultrasound department with a small annotated set and a specific question: fetal plane classification (0.956 accuracy), gallbladder cancer detection (0.934), breast nodule segmentation (0.887 DSC) and thyroid nodule segmentation (0.882 DSC) are all reached this way. Label efficiency is the point — cut to 20% of the fine-tuning data, the model still averages 0.772 DSC on segmentation. The federated design also matters operationally: a hospital joining later can contribute a new organ or modality to the shared encoder without exporting its archive, which makes the approach usable where a data transfer agreement is not obtainable.
UltraFedFM shows that privacy regulation need not cap the size of a medical pretraining corpus, and that the resulting encoder transfers across organs, probe geometries and institutions. Code is Apache-2.0 and the pretrained and fine-tuned checkpoints are published, though on a university OneDrive share rather than a model hub; the private portion of the pretraining corpus is reachable only by controlled-access application. The authors are explicit about the geographic skew: 57.30% of pretraining images come from Chinese institutions, 15.94% from the United States and 10.14% from Germany, a bias they expect to surface as uneven performance on underrepresented populations. Every downstream path documented in the repository fine-tunes the encoder with a task head; there is no zero-shot mode.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.