Skip to collection list
Skip to video grid
Cost effective deployment of GE's CT axial classification algorithm on Intel server platform using Intel Deep Learning Deployment Toolkit
David Chevalier, Principal Engineer, GE Healthcare
Prashant Shah, Principal Engineer, Intel Corporation
Radiology is the cornerstone for applying deep learning techniques to help streamline the busy radiologist’s workflows and provide tools for accurate and consistent interpretation of medical images. Radiologists and clinical researchers spend considerable effort in searching for CT slices by anatomical regions in the PACS and VNA systems. GE Healthcare has developed a deep learning model in Tensorflow that can classify axial CT slices into six anatomical regions GE Healthcare needed a ﬂexible, high-performance deployment architecture so it can deliver its artificial intelligence (AI) innovations at the point of care, without driving up costs. Deployment of trained models require optimized hardware and software stack so that AI models can infer on a variety of deployment architectures, also allowing the workloads to co-exist with other workloads. GE and Intel collaborated on a core-scaling study to establish the inference throughput of the CT Axial Classification model when deployed on Intel Xeon based platform. We optimized GE’s model using Intel Deep Learning Deployment Toolkit (part of Computer Vision SDK) and deployed on Intel Xeon Servers to achieve the desired inference throughput. We will present the results of the analysis.