8000 GitHub - dqdinh/USF-MAE · GitHub
[go: up one dir, main page]

Skip to content
dqdinhPublic
forked from Yusufii9/USF-MAE

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

 
 

Repository files navigation

🩺 USF-MAE: Ultrasound Self-Supervised Foundation Model with Masked Autoencoding

This repository accompanies the paper:

Paper Title: USF-MAE: Ultrasound Self-Supervised Foundation Model with Masked Autoencoding
Authors: Youssef Megahed, Robin Ducharme, Aylin Erman, Mark Walker, Steven Hawken, Adrian D.C. Chan Published in the Biomedical Signal Processing and Control Journal

arXiv paper


USF-MAE Architecture

🧠 Overview

USF-MAE is the first large-scale, self-supervised MAE foundation model trained exclusively on ultrasound (US) data.
The model leverages a Vision Transformer (ViT)–based masked autoencoder (MAE) architecture to learn rich, modality-specific representations directly from unlabeled ultrasound scans.

USF-MAE was pretrained on 370,000 2D and 3D ultrasound images curated from 46 open-source datasets, collectively named OpenUS-46.
This repository provides:

  • The USF-MAE pretrained checkpoint (Checkpoints section)
  • Access links and references for OpenUS-46 (OpenUS-46 Dataset Collection section) - Additional three datasets (a total of 49D "OpenUS-49") can be accessed using the following link: Link
  • Scripts for USF-MAE pretraining
  • Access the Facebook MAE scripts using the following link: Link

📦 Model Highlights

  • Architecture: Vision Transformer (ViT) encoder–decoder MAE
  • Pretraining strategy: Self-supervised masked patch reconstruction
  • Pretraining data: 360K unlabeled ultrasound images from 43 datasets (10k from 3 datasets were used in evaluation)
  • Downstream tasks:
    • Breast cancer classification (BUS-BRA)
    • Ovarian tumor subtype classification (MMOTU-2D)
    • Gastrointestinal stromal tumor detection (GIST514-DB)
  • Key results:
    • F1-scores of 81.6 %, 79.6 %, and 82.4 % on the above tasks
    • Outperformed CNN and ViT baselines
    • Comparable to the supervised UltraSam model on BUS-BRA and outperformed it on the MMOTU-2D and GIST514-DB datasets while using no labels

🌍 OpenUS-46 Dataset Collection

OpenUS-46 is a large, unified compilation of 46 publicly available ultrasound datasets from multiple anatomical regions and clinical tasks.
All datasets remain under their original licenses. We provide direct access links for research reproducibility and transparency.

Dataset Name Anatomy / Target Region Access Link
105US (1) Liver Link
AbdomenUS (2) Abdomen Link
ACOUSLIC (3) Abdomen Link
AUL (4) Liver Link
BLUES (5) Lung Link
BP (6) Neck Link
Brachial Plexus (7) Neck Link
BrEaST (8) Breast Link
BUS (Dataset B) (9) Breast Link
BUS_UC (10) Breast Link
BUS_UCML (11) Breast Link
BUS-BRA (12) Breast Link
Cactus Dataset (13) Heart Link
CAMUS_public (14) Echocardiography Link
CardiacUDC (15) Heart Link
CCAUI (16) Common Carotid Artery Link
DFHI (17) Fetal Head Link
EchoCP (18) Heart Link
EchoNet-Dynamic (19) Heart Link
EchoNet-Pediatric (20) Heart Link
FALLMUD (21) Leg Muscle Link
FASS (22) Fetal Abdominal Link
Fast-U-Net (23) Fetal Abdominal Link
FEFT (24) Fetal Echocardiography Link
FETAL_PLANES_DB (25) Fetal Anatomy Link
FH-PS-AOP (26) Fetal Head Link
GIST514-DB (27) Gastrointestinal Link
HC (28) Fetal Head Circumference Link
JNU-IFM (29) Pelvis and Fetal Head Link
kidneyUS (30) Kidney Link
LUSS_phantom (31) Lung Link
MicroSeg (32) Prostate Link
MMOTU-2D (33) Ovarian Tumor Link
MMOTU-3D (34) Ovarian Tumor Link
Pocus (35) Lung Link
Porcine (36) Spinal Cord Link
PSFHS (37) Pelvis and Fetal Head Link
regPro (38) Prostate Link
S1 (39) Breast Link
Segthy (40) Thyroid and Neck Link
STMUS_NDA (41) Musculoskeletal Link
STU-Hospital (42) N/A Link
Thyroid Dataset (43) Head and Neck Link
Thyroid US Cineclip (44) Thyroid Link
Ultrasound Fetus Dataset (45) Fetal Link
UPBD (46) Brachial Plexus Link

📘 Note: All datasets retain their original licenses and should be cited appropriately.


🧩 Checkpoints

Model Pretraining Data Epochs Download
USF-MAE (100 epochs - recommended) OpenUS-46 (370K images) 100 Download
USF-MAE (500 epochs) OpenUS-46 (370K images) 500 Download

🧪 Usage

# Clone repository
git clone https://github.com/Yusufii9/USF-MAE.git
cd USF-MAE

📚 References

If you find our work helpful for your research, please consider citing us using the following BibTeX entry:

@article{MEGAHED2026110313,
author = {Youssef Megahed and Robin Ducharme and Aylin Erman and Mark C. Walker and Steven Hawken and Adrian D.C. Chan},
title = {USF-MAE: Ultrasound self-supervised foundation model with masked autoencoding},
journal = {Biomedical Signal Processing and Control},
volume = {122},
pages = {110313},
year = {2026},
issn = {1746-8094},
doi = {https://doi.org/10.1016/j.bspc.2026.110313},
url = {https://www.sciencedirect.com/science/article/pii/S1746809426008670}
}

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0