This work presents EchoCare, novel ultrasound foundation model for generalist clinical use, developed via self-supervised learning on our curated, publicly available, large-scale unlabeled dataset EchoAtlas. EchoAtlas comprises 4.5 million ultrasound images, sourced from 23 clinical centers across 5 continents, and acquired using 38 distinct imaging devices, thus encompassing multi-center, multi-device, and multi-ethnic global cohorts.
- EchoAtlas:
- Models:
- Results:
EchoAtlas covers 9 major regions and 52 anatomical organs of the human body, supporting models pretrained on it to generalize effectively across comprehensive whole-body ultrasound clinical application.
Our data curation process commenced with a systematic search of open academic repositories：
- Figshare: An online repository where researchers can share, manage, and showcase research outputs with DOIs for citation.
- Github: The world's leading platform for hosting and collaborating on code projects.
- Grand-challenge: A platform for hosting medical imaging challenges and datasets.
- Kaggle: One of the largest AI & ML community.
- Mendeley: A reference manager and academic social network for researchers.
- Zenodo: An open-access repository for research outputs and datasets.
Before training or inference, load the pre-trained weights. This gives you a well-initialized model that can be fine-tuned for your own dataset or task.
import torch
import argparse
from monai.networks.nets.swin_unetr import SwinTransformer
parser = argparse.ArgumentParser(description="Swin Transformer")
parser.add_argument("--feature_size", default=128, type=int, help="feature size")
parser.add_argument("--in_channels", default=3, type=int, help="number of input channels")
parser.add_argument("--pretrained_checkpoint", default=None, help="encoder pretrained checkpoint")
parser.add_argument("--use_checkpoint", default=True, help="use gradient checkpointing to save memory")
args = parser.parse_args()
encoder = SwinTransformer(
in_chans=args.in_channels,
embed_dim=args.feature_size,
window_size=[8] * 2,
patch_size=[2] * 2,
depths=[2, 2, 18, 2],
num_heads=[4, 8, 16, 32],
mlp_ratio=4.0,
qkv_bias=True,
use_checkpoint=args.use_checkpoint,
spatial_dims=2,
use_v2=True)
if args.pretrained_checkpoint is not None:
model_dict = torch.load(args.pretrained_checkpoint, map_location=torch.device('cpu'))
state_dict = model_dict
state_dict.pop('mask_token')
encoder.load_state_dict(state_dict, strict=True)
print("Using pretrained self-supervised Swin Transformer backbone weights !")
# Test case: forward pass with dummy input
# Expected output feature map shapes:
# [1, 128, 128, 128], [1, 256, 64, 64], [1, 512, 32, 32], [1, 1024, 16, 16], [1, 2048, 8, 8]
x = torch.rand(1, 3, 256, 256)
x_outs = encoder(x)
print([x_out.shape for x_out in x_outs])
We thank MONAI for part of their codes.