EchoCare - Large-scale ultrasound model launched by the Hong Kong Academy of Sciences

What is listening?

EchoCare is a large-scale ultrasound model developed by the Centre for Artificial Intelligence and Robotics Innovation (CAIR) at the Hong Kong Innovation Research Institute, Chinese Academy of Sciences. The model is trained on the EchoAtlas dataset, which contains 4.5 million ultrasound images. This dataset covers images acquired by 23 clinical centers across five continents using 38 different imaging devices, encompassing nine major regions of the human body and 52 anatomical organs, making it one of the largest ultrasound image datasets currently available. EchoCare employs a "structured contrastive self-supervised learning framework," significantly improving the model's ability to model the deep semantics of ultrasound images and its generalization performance through techniques such as image mask reconstruction and adaptive hard patch mining. In seven major medical tasks, including ultrasound image segmentation, classification, detection, regression, and enhancement, its performance surpasses current state-of-the-art methods.

The main function of listening

• Ultrasound image segmentationIt can accurately segment different tissues and organs in ultrasound images, helping doctors to more clearly identify the boundaries between lesion areas and normal tissues, and providing a more accurate basis for diagnosis.
• Lesion classificationIt can classify lesions in ultrasound images, such as distinguishing between benign and malignant tumors, to help doctors quickly determine the nature of the lesions and improve diagnostic efficiency.
• Organ detection and segmentationIt can not only detect the location of organs in images, but also perform precise segmentation of organs, providing detailed anatomical information for subsequent diagnosis and treatment.
• Image enhancementThe quality of ultrasound images is improved by enhancing their contrast and clarity, enabling doctors to observe subtle structures and lesions more clearly and reducing the possibility of misdiagnosis.
• Report generationIt automatically generates diagnostic reports based on ultrasound image analysis results, saving doctors time in writing reports, improving work efficiency, and ensuring the accuracy and consistency of reports.

The technical principles of listening

• Large-scale dataset constructionThe model was trained using the EchoAtlas dataset, which contains 4.5 million ultrasound images. This dataset covers a global cohort with multiple centers, devices, and ethnicities, encompassing nine regions of the human body and 52 anatomical organs, providing rich learning materials for the model.
• Structured contrastive self-supervised learning frameworkThe model introduces hierarchical tree-shaped labels based on medical priors to achieve structured learning and implicit encoding of multi-label semantic relationships, thereby improving the model's ability to model the deep semantics of ultrasound images.
• Image mask reconstruction technologyBy performing mask reconstruction on the image, the model's learning and understanding of local image features is enhanced, thereby improving the model's robustness and generalization performance.
• Adaptive hard patch mining techniqueIt automatically identifies and focuses on image regions that are difficult to learn, and strengthens learning in a targeted manner to improve the model's ability to process complex images.
• Progressive training strategyA progressive training method is adopted to gradually increase the difficulty and complexity of model training, helping the model to better adapt to different types of ultrasound images and improve overall performance.

The project address of Lingyin

• Project official websitehttps://echocare.cares-copilot.com/
• Github repositoryhttps://github.com/CAIR-HKISI/EchoCare
• arXiv technical paper: https://arxiv.org/pdf/2509.11752

Application scenarios of listening

• Routine hospital check-upIts application in routine ultrasound examinations in hospitals can significantly reduce reliance on professional personnel, assist doctors in making diagnoses more efficiently and accurately, effectively improve the efficiency of medical services, and provide more possibilities for the optimal allocation of medical resources.
• Disease diagnosis and screeningIt can be used for the diagnosis and screening of a variety of diseases. For example, specific cases in the Department of Obstetrics and Gynecology of Qilu Hospital of Shandong University, involving 1,556 cases of ovarian tumor ultrasound and more than 1,000 cases of thyroid ultrasound examination at Xiangya Hospital of Central South University, have verified that its performance is significantly better than the existing state-of-the-art methods.
• EchocardiographyIn the detection and analysis of aortic aneurysms using cardiac ultrasound, retrospective verification can be performed, providing strong support for the diagnosis of heart diseases.
• Ultrasound image processingIt includes seven major medical tasks such as ultrasound image segmentation, classification, detection, regression, and enhancement, as well as more than ten downstream applications, such as lesion classification, organ detection and segmentation, and image enhancement, to improve the quality and diagnostic value of ultrasound images.
• Clinical adaptation and researchThe base model is planned to be open-sourced for use by medical institutions. In the future, it will be promoted in stages for prospective research, application in emergency room scenarios, and hardware integration with ultrasound equipment companies.