Put AI in the hands of healthcare

Artificial intelligence (AI) promises to bring about a revolution in healthcare. The underlying combination of Machine Learning and analytics can process medical data sets that are so large and medical images that are so numerous that they fall outside the scale of researchers, doctors and employees. This AI duo therefore promises to help identify risk patients and prevent the development of diseases and medical conditions. For existing patients, the hope is that AI can identify hidden diseases, localize medical problems, and in the development and application of treatments that promote patient recovery.

Yet adoption has been delayed by the costs and complexity of building and owning the kind of high-quality systems that are needed. However, that is changing because the processors are optimized for AI training and inference and because they are combined with the more powerful software.

An example of this are the second generation Intel® Xeon® Scalable processors. They deliver up to 30 times more performance for AI inference compared to the previous generation Xeon®. Intel® Deep Learning Boost now includes specific x86 extensions that help accelerate convolutional, neural network-based algorithms. Performance has been further improved for both batch and real-time inference using the vector neural network instruction (VNNI) function that reduces the number and complexity of the convolution operations required for AI inference. VNNI also reduces the volume of required computing power and memory access, further reducing latency and increasing the performance of AI applications.

Running production-like AI on a scale goes further than just hardware – it also requires powerful software. Here, the industry seems to be merging around Google's open-source TensorFlow for Machine Learning – a framework for building and training the kind of large-scale numerical calculations that AI requires. TensorFlow uses Python to provide a front-end API for building applications, but executes it in C ++. It is very suitable for training and performing neural networks for the image recognition workload required in common medical environments such as radiology and CT scans.

Intel® has worked closely with Google to optimize the successive generations of Intel® Xeon® using the Intel® Math Kernel Library (MKL). It has also gone a step further with the second-generation Intel® Xeon®, adding support for 8-bit precision inference on models used for image classification, object detection and recommendation systems. These systems already used 32-bit floating point to increase the speed of arithmetic operations performed per second and to reduce system memory by around 75 percent, so MKL is a real step forward.

AI, not X-rays

The Intel® Xeon® platform is already being used for intensive analysis in healthcare.

One of the adopters is the LineSafe National Imaging Academy of Wales that uses AI to ensure proper placement of naso-gastric (NG) tubes in patients, to carry food and medication through the nasal cavity to the stomach. It is an established practice for doctors and nurses to use X-ray manual processes to assess whether they have properly inserted the tube into the esophagus rather than the trachea. It can be difficult for radiologically inexperienced personnel to separate these organs by looking at x-rays due to their proximity, and it can be life threatening to do this wrong.

LineSafe now trains a mechanical learning model for a perfect placement of the pipe. The model is fed with thousands of chest x-rays, stored in the National Imaging Academy Wales, so that the system can accurately see if the NG tubes are positioned correctly.

The Intel® UK Health and Life Sciences (HLS) team has been working with LineSafe on the project since the beginning of 2019. Their role has been to help identify the best hardware configuration and AI-optimized software for training and production of the application. The AI modeling involved uses systems with the second generation Intel® Xeon® Scalable processors to accelerate the performance of the algorithms that analyze high-resolution images. Intel® Xeon® therefore makes it possible to make comparisons faster and more accurately than before.

Camera detection

CorporateHealth International (CHI) is another healthcare organization that embraces AI with Intel®. The company, based in Denmark, offers a managed endoscopy service with intestinal capsules based on a small video camera that is no larger than a vitamin pill and swallowed by the patient. The camera records up to 400,000 images as it travels through the digestive system on its mission to help detect symptoms of gastric and intestinal diseases. The images are analyzed by a team of nurses.

Director Hagen Wenzek: “AI can be a very valuable tool. We train a neural network with data from previous procedures that our team already has, so that the neural network is used to help nurses illuminate all suspicious images”. CHI collaborated with Intel® AI Builders – a network of software producers, integrators and equipment manufacturers – to select the best hardware and software combination for the application. AI Builders also assisted with the implementation and integration within CHI's business systems to securely process patient data.

CHI uses two servers with Intel® Xeon®, one for data processing and the other for AI development. The data processing server processes and analyzes the original RAW video files, which is much faster with the new Intel® Xeon® architecture and cheaper than leasing infrastructure, with similar capabilities, from cloud service providers.

“If you rent that out via the cloud, it actually gets expensive pretty quickly,” Wenzek notes. Hiring AI specialists would also have been expensive for CHI. CHI used the competencies and capabilities of the Intel® AI Builders program instead.

Planning the radiotherapy

Intel is also working with the Velindre University NHS Trust and the Cardiff University School of Engineering on the ASPIRE project. The aim of the project is to develop deep-learning systems that are able to automate the planning of the treatment of esophageal cancer with radiotherapy.

Such a schedule usually depends on a trained oncologist who performs a diagnostic computed tomography (CT scan) to determine the position of a tumor – a process that can take several days. ASPIRE strives to shorten that prognosis with a machine model that is trained to accurately determine the location of a tumor. The model is taught using more than a thousand 3D scans with labeled structures.

Here too, Intel® has assisted the ASPIRE team in choosing the hardware and software configuration, including workshops for staff involved in training and optimizing the performance and accuracy of the machine model. As the accuracy of the ASPIRE is improved, it is assumed that the assignment can be extended to the planning of radiotherapy against tumors in other places in the human body.

Image recognition

Image recognition is a promising part of medicine and here Intel® has developed the Open Visual Inference & Neural Network Optimization, or OpenVINO, that can help. It is a toolkit to help developers quickly build and implement computer vision for cameras in IP-based devices. OpenVINO works with popular open-source frameworks such as TensorFlow and Caffe and with Intel® processors.

The Intel® distribution of OpenVINO ™ is optimized for Intel® Xeon®. One of the advantages is that the AI workload can be accelerated without the need for expensive GPUs on peripherals that would otherwise be relatively cheap to purchase and run. It helps companies such as MaxQ AI, specialist in medical diagnostics, which has used the Intel® distribution to triple the computing power of its Accipio intracranial hemorrhage (ICH) and stroke detection platform. Accipio uses vision algorithms that have been trained using Machine Learning and neural networks.

Another company that uses the Intel® platform in medicine is AI platform supplier JLK Inspection. The company has built 37 algorithms for use in medical inspections of different parts of the body. Many of these algorithms are used on low-cost, small Intel® NUC mini PCs with Intel® Distribution or OpenVINO ™ for image recognition. Intel® Distribution or OpenVINO ™ gives AI software developers such as JLK Inspection a ready-made set of convolutional neural network (CNN) -based deep-learning tools for visual inferencing tasks such as image classification and object detection.

The common thread here is – clearly – that AI-driven analysis is beginning to penetrate into some stressful areas of medicine. It helps to tackle a number of tricky problems thanks to the power of AI to crack challenging volumes of structured and unstructured data, in support of medical professionals. Behind these developments is an accelerated hardware and software stack from Intel® that delivers the computational efficiency needed for fast and accurate analysis at a price point that is crucial for AI.