Overview

pAt Siemens Healthineers, our purpose is to enable healthcare providers to increase value by empowering them on their journey toward expanding precision medicine, transforming care delivery, and improving patient experiences, all made possible by digitalizing healthcare. An estimated 5 million patients globally benefit every day from our innovative technologies and services in the areas of diagnostic and therapeutic imaging, laboratory diagnostics, and molecular medicine, as well as digital health and enterprise services. We are a leading medical technology company with over 120 years of experience and 18,000 patents globally. Through the dedication of more than 50,000 colleagues in 75 countries, we will continue to innovate and shape the future of healthcare./p

Supernova Award Category

AI and Augmented Humanity

The Problem

pOne-third of all deaths—34 deaths per minute and 18 million each year—are due to cardiovascular disease.² Cardiac MRI has established itself as a gold standard for evaluating heart function, heart chamber volumes, and myocardial tissue evaluation.³ To extract quantitative measurements from the CMR images, the cardiologists typically use manual or semi-automatic tools, a time-consuming step that is error prone and affected by the inter-user subjectivity in interpreting the images. Adding to the challenge is the wide range of available cardiovascular data types coupled with physician subjectivity in interpreting and measuring the data, which can lead to undiagnosed or misdiagnosed diseases. Siemens Healthineers is a pioneer in the use of AI for medical applications. They research novel AI use cases and incorporate their findings in their scanners and radiology and cardiology applications. These AI use cases need to be implemented seamlessly into the clinical workflow to save time and increase consistency and accuracy in measurements and diagnoses. However, AI cannot come at the expense of delays in the clinical workflow. The computation system powering AI needs to keep pace with the data being generated by the scanners, requiring the system to offer low latency for AI inference and high throughput. This allows healthcare systems to care for more patients in a day./p p /p

The Solution

pSiemens Healthineers and Intel have collaborated on optimizing Siemens Healthineers’ heart chamber detection and quantification model for 2nd Generation Intel Xeon Scalable processors.1 This AI model performs semantic segmentation of the left and right ventricles of the heart and can be extended to all four chambers. The input to the AI model is a stack of MRI images of the beating heart and the output identifies regions or structures of the heart, where each structure is color coded. This automates the laborintensive manual segmentation process, accelerating time  to results./p p“We can now develop multiple near-real-time, often critical medical imaging use cases, such as cardiac MRI and others, using Intel® Xeon® Scalable processors, without the added cost or complexity of accelerators.”  —Dr. Dorin Comaniciu, senior VP, Siemens Healthineersbr /  /p

The results

pAchieving 5.5x speedup of int8 over the baseline fp32 model on 2nd Generation Intel Xeon Scalable processors is a result of efficient, low-precision convolutions due to Intel Deep Learning Boost, efficient concatenations in int8, and resample operation optimizations.¹ The neural network is trained to identify regions of the heart. The weights and activations of the neural network are represented as fp numbers. The models are typically trained in fp32 precision. Once the desired level of accuracy is obtained, the model is ready to be incorporated in products. While many fp32 models are deployed in products today, quantizing the model to int8 can offer significant performance benefits. Typically, when quantization is performed correctly, there is little to no accuracy loss in the resulting model. Our goal was to limit the accuracy loss to 0.5%. Our tests indicate that the resulting accuracy loss obtained was 0.001%. Intel quantized the trained model from Siemens Healthineers from fp32 to int8 using the Intel Distribution of OpenVINO toolkit, ensuring that accuracy was not compromised. The resulting shift in accuracy on the validation set of images was a low .001%; essentially maintaining imaging results for analysis at higher computational speeds./p

Metrics

pThe optimization of the cardiac MRI segmentation model demonstrates the power of 2nd Generation Intel Xeon Scalable processors—allowing Siemens Healthineers to meet the growing needs of data-intensive AI applications for the health and life sciences industry. The optimization process indicates how solutions can be customized to meet specific real-world requirements for performance and accuracy. Siemens Healthineers continues to refine and evolve AI training models to improve accuracy and support evolving workloads and use cases. With Intel and Siemens Healthineers, the health and life sciences industry can leverage AI to integrate and analyze large amounts of data—helping to improve healthcare via more accurate diagnoses for better-targeted treatments./p pWith these optimizations, benchmarking the model on  2nd Generation Intel Xeon Scalable processors resulted in a 5.5x gain over the baseline model with overall throughput of 201.36 images/sec using 14 inference streams with two threads each, on a single 2nd Generation Intel Xeon Scalable processor socket.¹/p

The Technology

p5.5x speedup: Based on Siemens Healthineers and Intel analysis on 2nd Gen Intel® Xeon® Platinum 8280 Processor (28 Cores) with 192GB, DDR4-2933, using Intel® OpenVino™ 2019 R1. HT ON, Turbo ON. CentOS Linux release 7.6.1810, kernel 4.19.5-1.el7.elrepo.x86_64.Custom topology and dataset (image resolution 288x288). Comparing FP32 vs Int8 with Intel® DL Boost performance on the system./p

Disruptive Factor

pIntel Deep Learning Boost is built into 2nd Generation Intel Xeon Scalable processors to accelerate deep learning use cases. It extends the instruction set with a new Vector Neural Network Instruction (VNNI). Tasks such as convolutions, which typically required many instructions, can now be accomplished with just one instruction. Examples of these targeted workloads include image classification, image segmentation, speech recognition, language translation, object detection, and more. In order to take advantage of these instructions, models that are typically trained in floating point 32 (fp32) need to be quantized to int8. Through quantization, these workloads can be accelerated, but care must be taken to preserve the accuracy of the model. The team used the Intel Distribution of OpenVINO toolkit to optimize, quantize, and execute the model. The resulting solution achieved 5.5x speedup with almost no degradation  in accuracy.¹ Such accelerations enable future solutions that:/p p* Process cardiac MRI data with unprecedented efficiency; at 200 fps, a full cardiac MRI exam, short axis spatio-temporal stack can be analyzed in less than a second/p p* Open the possibility for near-real-time clinical applications of cardiac MRI, making the interpretation of data available right after its acquisition/p

Shining Moment

pThe optimization process indicates how solutions can be customized to meet specific real-world requirements for performance and accuracy. With Intel and Siemens Healthineers, the health and life sciences industry can leverage AI to integrate and analyze large amounts of data—helping to improve healthcare via more accurate diagnoses for better-targeted treatments./p