This article is more than 1 year old
Putting AI in the hands of healthcare
How to overcome the complexity and cost
Sponsored Artificial intelligence (AI) promises to revolutionize healthcare. The underlying combination of Machine Learning and analytics can process medical data sets so large and medical images so numerous that they are beyond the scale of researchers, physicians and staff. In so doing, this AI duo promises to help identify patients at risk and prevent the onset of diseases and medical conditions. For existing patients, the hope is AI can identify hidden illnesses, pinpoint medical problems and in the development and application of treatments that assist patient recovery.
Yet adoption has been held back thanks to the cost and complexity of building and owning the kinds of high-performance systems needed. That is changing, though, as processors become optimised for AI training and inference and as they are combined with the more powerful software.
As an example of the former, the second generation Intel® Xeon® Scalable processors are a case in point. They deliver anything up to a 30 times increase in performance for AI inference compared to the previous generation of Xeon®. Intel® Deep Learning Boost, meanwhile, includes specific x86 extensions that help accelerate convolutional neural-network-based algorithms. Performance is further improved for both batch and real-time inference using the vector-neural network instruction (VNNI) function that reduces the number and the complexity of convolution operations required for AI inference. VNNI also cuts down the volume of compute power and memory access required, further reducing latency and increasing performance of AI applications.
Running production-grade AI at scale goes beyond just hardware – it requires powerful software, too. Here, the industry seems to be coalescing around Google’s open-source TensorFlow for Machine Learning – a framework for building and training the kinds of large-scale numerical computation demanded by AI. TensorFlow uses Python to deliver a front-end API for building applications but executes in C++. It is well suited to training and running neural networks for the image recognition workloads demanded in common medical environments such as radiology and CT scans.
Intel® has worked closely with Google to optimise successive generations of Intel® Xeon® using the Intel® Math Kernel Library (MKL). It has gone a step further with its second generation Intel® Xeon®, too, by adding support for 8-bit precision inference on deep-learning models used for image classification, object detection and recommendation systems. These systems were already using 32-bit floating point to increase the rate of arithmetic operations executed per second and to reduce system memory consumed by approximately 75 per cent so MKL is a real step forward.
AI, not X-rays
The Intel® Xeon® platform is already being picked up for intensive analytics in healthcare.
One adopter is the LineSafe National Imaging Academy of Wales that is using AI to ensure the correct placement of naso-gastric (NG) tubes in patients, to carry food and medicine to the stomach through the nasal cavity. It’s established practice for doctors and nurses to employ manual processes aided by X-rays to assess whether they have correctly fed the tube into the oesophagus rather than the windpipe. It can be hard for staff inexperienced in radiology to tell these organs apart by looking at X-rays due to their close proximity, and getting this wrong can be life threatening.
LineSafe is now training a machine-learning model to perfect placement of the tube. The model is being fed on thousands of chest X-ray images, stored in the National Imaging Academy Wales, so the system can accurately tell whether NG tubes have been inserted correctly.
The Intel® UK Health and Life Sciences (HLS) team has been working on the project with LineSafe since early 2019. Their role has been to help identify the best hardware configuration and AI-optimised software for the application’s training and production workloads. The AI modelling involved uses systems running the second generation of Intel® Xeon® Scalable processors to speed performance of the algorithms that are analysing high-resolution images. Intel® Xeon® is thereby allowing comparisons to be made more quickly and more accurately than before.
CorporateHealth International (CHI) is another healthcare organisation embracing AI with Intel®. The Denmark-based company provides a managed colon capsule endoscopy service in the UK that relies on a tiny video camera that’s no bigger than a vitamin pill and is swallowed by the patient. The camera takes up to 400,000 images as it travels through the digestive system on its mission to help spot symptoms of gastrointestinal diseases. The footage is analysed by a team of nurses.
Managing director Hagen Wenzek says: “AI can provide an extremely valuable tool. We are training a neural network with data from previous procedures that our team already has so the neural network is being used to help nurses highlight all images that are suspicious.” CHI worked with Intel® AI Builders – a network of software makers, integrators and equipment manufacturers – to select the best hardware and software combination for the application. AI Builders also helped deploy and integrate within CHI’s corporate systems, to process patients’ data securely.
CHI uses two servers running Intel® Xeon®, one for data processing and the other for AI development. The data processing server processes and analyses the original RAW video files, something that is much faster to do using the new Intel® Xeon® architecture and more cost effective than leasing infrastructure, with similar capabilities, from cloud-service providers.
“If you rent that out through the cloud it is actually going to get pretty expensive very quickly,” noted Wenzek. Hiring AI specialists would have been costly for CHI, too. Rather, CHI exploited the competencies and capabilities found within the Intel® AI Builders programme.
Intel is also working with the Velindre University NHS Trust and the Cardiff University School of Engineering on the ASPIRE project. The project’s goal is to develop deep-learning systems that are capable of automating the planning of radiotherapy treatment of oesophageal cancer.
Such planning typically relies on a trained oncologist performing a diagnostic computing tomography (CT) scan to identify the position of a tumor – a process that can take several days. ASPIRE aims to cut that prognosis time with a machine model trained to accurately determine a tumour’s location. The model is being taught using more than one thousand 3D scans featuring labelled structures.
Again, Intel® helped the ASPIRE team choose the hardware and software configuration, with workshops also provided for staff involved in training and optimising the performance and accuracy of the machine model. As the accuracy of the ASPIRE is improved, it’s believed the remit could be extended to plan radiotherapy against tumors in other locations within the human body.
Image recognition is a promising part of medicine and here Intel® has developed the Open Visual Inference & Neural Network Optimisation, or OpenVINO, that can help. It’s a tool kit to help developers quickly build and deploy computer vision for cameras in IP-based devices. OpenVINO works with popular open-source frameworks like TensorFlow and Caffe and with Intel® processors.
The Intel® Distribution of OpenVINO™ is optimised for Intel® Xeon®. Among its benefits, it allows AI workloads to be accelerated without need for expensive GPUs on edge devices that are otherwise relatively cheap to purchase and to run. It’s helping companies like medical diagnostics specialist MaxQ AI that used the Intel® distribution to triple the computational power of its Accipio intracranial hemorrhage (ICH) and stroke detection platform. Accipio uses vision algorithms trained using Machine Learning and neural networks.
Another firm capitalising on the Intel® platform in medicine is AI platform provider JLK Inspection. The company has built 37 algorithms for use in medical inspections of different parts of the body. Many are deployed on low-cost, small form factor Intel® NUC mini PCs that run Intel® Distribution of OpenVINO™ for image recognition. Intel® Distribution of OpenVINO™ gives AI software developers like JLK Inspection a ready-made set of convolutional neural network (CNN)-based deep learning tools for visual inferencing tasks like image classification and object detection.
The common thread here is – clearly – that AI-driven analytics is beginning to make inroads into some taxing areas of medicine. It’s helping tackle some tough problems thanks to AI’s power to crunch challenging volumes of structured and unstructured data, to support medical professionals. Behind these inroads is an accelerated hardware and software stack from Intel® that’s delivering the computational efficiency needed for speedy and accurate analysis at a price point that’s critical in AI.
Sponsored by Intel®