This repository contains the code implementation for US-1M, which establishes a new benchmark in medical imaging research, comprising 1.28 million rigorously annotated ultrasound images with precise anatomical structure delineations and comprehensive demographic metadata.
- May 23, 2025
- Demo released on Hugging Face spaces (view demo).
- Apr 13, 2025
- Annotations and code scripts for preparing the US-1M pretraining dataset are released.
- US-1M training and inference code are released, along with pretrained checkpoints.
Abstract: *Foundation models have made impressive strides in medical AI research, providing valuable features for downstream tasks and enabling task-specific models to adapt to new settings with minimal or no additional examples. However, their application in ultrasound (US) domain remains limited due to the lack of large-scale and openly accessible US medical datasets. To fill this gap, we introduce US-1M, a comprehensive and unified ultrasound dataset featuring multi-organ, multi-center, multi-device, and multi-race coverage, comprising over 1 million images. Building on US-1M, we demonstrate its versatility through three representative tasks across three critical medical applications: organ segmentation, disease classification, and image enhancement. Our experiments reveal substantial performance gains in multiple ultrasound imaging tasks, establishing new benchmarks for ultrasound-based AI models. The unified and hierarchical architecture of US-1M opens new research frontiers in medical imaging, particularly for ultrasound analysis. *
< 78D5 /blockquote>Main contributions of our work are:
- We present US-1M, the first large-scale, openly accessible, and hierarchical ultrasound imaging dataset comprising 1.28 million clinically validated images collected worldwide. This dataset serves as a cornerstone for foundation model research in medical US imaging.
- We demonstrate US-1M clinical versatility through foundation model pretraining, with experimental results confirming its utility in developing robust, generalizable models across diverse medical US tasks.
- By conducting case study on accuracy analysis we shed new insights into US foundation model architetures.
Table 1: Quantitative comparison results on organ segmentation (mean ± std).
Method Paper Pretrained on ImageNet Pretrained on US-1M DSC (%) ↑ NSD (%) ↑ DSC (%) ↑ NSD (%) ↑ U-Net Link 72.69 ± 1.89 76.10 ± 1.79 75.23 ± 1.47 78.40 ± 1.26 DeepLabv3 Link 76.65 ± 1.64 80.29 ± 1.68 78.72 ± 1.56 82.77 ± 1.07 SwinUNETR Link 78.62 ± 1.92 82.62 ± 1.39 80.49 ± 1.67 83.57 ± 1.39 Table 2: Quantitative comparison results on disease classification (mean ± std).
Method Paper Pretrained on ImageNet Pretrained on US-1M Macro-F1 (%) ↑ Precision (%) ↑ Recall (%) ↑ Macro-F1 (%) ↑ Precision (%) ↑ Recall (%) ↑ ResNet50 Link 58.21 ± 2.23 59.10 ± 3.99 58.53 ± 2.18 59.21 ± 2.16 60.69 ± 3.44 59.62 ± 2.18 Swin-Transformer v2 Link 60.15 ± 1.56 59.30 ± 1.97 60.78 ± 1.76 62.53 ± 1.35 65.27 ± 2.88 61.58 ± 1.50 Table 3: Quantitative comparison results on image enhancement (mean ± std).
Method Paper Pretrained on ImageNet Pretrained on US-1M NIQE (%) ↓ BRISQUE (%) ↓ PIQE (%) ↑ FID (%) ↓ NIQE (%) ↓ BRISQUE (%) ↓ PIQE (%) ↑ FID (%) ↓ CycleGAN Link 7.12 ± 2.04 31.68 ± 3.92 29.24 ± 3.31 118.6 ± 9.18 6.85 ± 1.59 20.74 ± 2.97 30.26 ± 3.61 62.53 ± 6.78 CUT Link 7.16 ± 1.86 37.91 ± 4.34 43.26 ± 3.78 143.5 ± 9.59 6.77 ± 1.42 21.8 ± 2.46 30.26 ± 2.84 61.81 ± 6.11 EnlightenGAN Link 6.76 ± 1.86 22.83 ± 2.43 28.6 ± 2.36 60.24 ± 3.08 6.55 ± 1.20 17.68 ± 2.50 30.18 ± 2.84 58.71 ± 4.25 Before using US-1M for training and inference, please refer to the installation instructions described at INSTALL.md
We provide 2 model weights for US-1M as listed in the table below.
model_namepretrained_weightsRes. GPUs Avg. score on 21 datasets ResNet50 resnet50256 8 x A100 (40G) 61.63 SwinTransformer-B-16 swintransformer256 8 x A100 (40G) 62.09 For preparing US-1M pretraining dataset, we provide instructions for i) Downloading raw datasets from publicly available sources and ii) downloading processed annotations and merging with raw-datasets to build US-1M Refer to the detailed instructions described in US1M-DATA.md.