Platform to Deliver Enhanced Surgical Outcomes with AI-Powered Medical Devices
, Kaliber Labs
, Kaliber Labs
Building AI models to interpret video feeds requires a unique set of tools. This is especially true for real-time, surgical video feeds. In this talk, we describe how we tackled the challenge by building a proprietary end-to-end platform, MeDUSA, to ingest data, curate feeds, and generate datasets for training and testing models. We also dive into how we use MeDUSA to build applications to support minimally invasive surgical procedures. Minimally invasive surgeries are on the rise because of the myriad benefits they offer patients, but they are technically demanding for surgeons. To augment surgeons’ skills, we have developed over 41 deep learning models that enable accurate, real-time anatomy and pathology recognition, as well as precise structure annotation, pinless landmarking, and measurement. MeDUSA has been central to this work through its performance of critical tasks including image labeling and AI-augmented labeling, model training, model optimization and testing, packaging these models as a runtime application, and deployment to NVIDIA Clara AGX edge devices at scale. MeDUSA, in turn, draws on the power of NVIDIA’s AI hardware and software stacks to process video feeds at scale. While we are currently focused on building arthroscopic software solutions, we are making MeDUSA available with the KL Inference SDK to encourage other developers to build surgical AI solutions for other surgical specialties.