GE HEALTHCARE ACCELERATES AI INNOVATION WITH HEALTHCARE-SPECIFIC FOUNDATION MODELS POWERED BY NVIDIA

Building on a long-term artificial intelligence (AI) collaboration, GE HealthCare used NVIDIA technology to develop its recent research model SonoSAMTrack, which combines a promptable foundation model for segmenting objects on ultrasound images called SonoSAM. SonoSAMTrack focuses on segmenting anatomies, lesions, and other essential areas in ultrasound images. SonoSAMLite is a streamlined version of SonoSAMTrack.

In healthcare, leveraging AI to enhance patient care, streamline operational efficiencies, and make informed decisions has become increasingly important. Traditionally, the approach to integrating AI into healthcare systems required the retraining of models to accommodate the unique requirements of different patient populations and hospital settings. This conventional method can lead to heightened costs, complexity, and the need for specialized personnel, therefore hindering the broad adoption of AI technologies in healthcare domains. Foundation models have risen to prominence due to their ability to operate as human-in-the-loop AI systems, garnering significant attention.

Foundation and generative AI models could play a crucial role by enabling swift adaptation to various diseases, facilitating screening, early detection, tracking progression, and identifying non-invasive biomarkers with minimal training requirements, such as zero-shot or few-shot settings.

In a recent study conducted by GE HealthCare, its research project, SonoSAMTrack, showcased high performance across seven ultrasound datasets, encompassing a wide range of anatomies (adult heart and fetal head) and pathologies (breast lesions and musculoskeletal pathologies), as well as different scanning devices. Notably, it outperformed competing methods by a substantial margin. In addition, SonoSamTrack exhibited enhanced performance metrics in terms of speed and efficiency, requiring only 2-6 clicks for precise segmentation, thus minimizing user input2. This achievement was made possible through distillation and quantization techniques, utilizing the NVIDIA TensorRT software development kit and other capabilities for quantization-aware training.