Hong Kong’s Centre for Artificial Intelligence and Robotics (CAIR) has unveiled an Ultrasound Large AI Model. The Centre believes that use of “EchoCare” in routine hospital examinations will significantly reduce reliance on scarce medical specialists, making ultrasound exams more widely available for a range of medical conditions. It will also assist doctors in making diagnoses more efficiently and accurately.

The EchoCare Ultrasound Large Model was trained on the first ultrasound image dataset known to exceed 4 million images.

The model introduces a “Structured Contrast Self-Supervised Learning Framework”, which leverages hierarchical tree labels derived from medical priors to enable multi-label semantic relational structured learning and implicit encoding. Through techniques such as Masked Image Modelling (MIM), Adaptive Hard Patch Mining, and Progressive Training, the model effectively enhances its ability to model the deep semantic features of ultrasound images and improves generalisation performance.

Test results show that “EchoCare” achieves state-of-the-art (SOTA) performance across seven medical tasks, including image segmentation, classification, detection, regression, and enhancement, as well as in over ten downstream applications. On average, it delivers a 3%-5% improvement compared with current SOTA methods.

At a press conference unveiling the new AI model, Deputy Director of CAIR, Professor Gaofeng Meng said that, unlike conventional large models, EchoCare adopts a structural self-supervised learning approach. It removes the need for extensive data annotation, enables feature learning, and decouples downstream tasks, thereby internalising prior knowledge in ultrasound and facilitating cross-task knowledge transfer.

He emphasised the model’s technical highlights, data advantages, and application results. Specific case validations included 1,556 ovarian tumour ultrasound cases at Qilu Hospital of Shandong University and more than 1,000 thyroid ultrasound examinations at Xiangya Hospital of Central South University, where EchoCare significantly outperformed existing state-of-the-art methods.