Page 159 - ITU KALEIDOSCOPE, ATLANTA 2019

ICT for Health: Networks, standards and innovation

Processors are major computing units in AI for health laboratory test, medical treatment, monitor & recovery. The
systems. Performance could be evaluated by the metrics of mainstream applications of AI are listed as below.
calculation speed, data bandwidth and power consumption
per unit of time. Processors used in AI usually include central ➢ AI Virtual assistant: Logistics such as online
processing unit (CPU), graphics processing unit (GPU), reservation, intelligence triage, payment and
field-programmable gate array (FPGA), application specific repairment monitoring could be effectively completed
integrated circuit (ASIC) and system-on-a-chip (SoC) by AI virtual assistants. Information input could be in
accelerators, etc. The Google Brain was once based on the various formats, such as audio, pictures, EHR scan and
CPU, but the general calculators have limited abilities for questions answering. By speech/image recognition and
floating point calculations, and was unable to meet deep natural language processing technology, it could
learning requirements, especially in model training. Though understand the patient's description of symptoms,
GPU is currently the primary choice due to its high- automatically provides intelligent consultation, triage
bandwidth caches and strong parallel computing power, suggestions and payment assistance. An intelligent
customized chips have more potential than these general- voice product “Yun Medical Sound” has been applied
purpose chips. FPGA is very flexible to achieve a high in more than 40 hospitals in China, with a voice
degree of customization, and ASIC even has a better transcription accuracy rate of over 97%. Additionally,
performance, with a computing speed of over 5-10 times these kinds of products’ functions can also be expanded
faster than FPGA. High R&D costs and production cycles to service rating, doctor matching, in-hospital
are two main obstacles for customized processors. A scale navigation, medical insurance reimbursement, pre-
effect may reduce the cost in the long term. Tractica forecasts diagnosis data collection, post-diagnosis follow-up, re-
that the market for deep learning chipsets will increase from examination reminder, health knowledge teaching, etc.
$1.6 billion in 2017 to $66.3 billion by 2025, and ASIC The application forms of the virtual assistant are very
market will be the largest by 2025 [26]. flexible and adaptive to certain scenarios, including
APP, websites and embedded programs, etc.
Network architecture is also customized to support AI
services. Continuous health condition monitoring and ➢ Medical imaging aided diagnosis: The core steps in
complex health management scenarios require flexible health service, such as symptom check and image
computing abilities. Architecture with a combination of inspection are currently penetrated with AI in the form
cloud and edge computing will be increasingly suitable for of medical imaging aided diagnosis. Based on
growing health needs. Cloud-computing solutions offer a computer vision and pattern recognition technology, AI
pay-per-use model that provides on-demand access to could achieve image classification and retrieval, 3D
computing resources. The cloud platform for deep learning reconstruction, image segmentation, feature extraction,
can be customized on TensorFlow, Caffe, MXNet, Torch, lesion identification, target area delineation and
etc., and provides developers with common models to reduce automatic annotation, etc. Various application
R&D costs. Algorithm training, assessment, visualization scenarios include fundus screening [27], breast
tools and API services are also available for customization. pathology diagnosis [28], X-ray reading, brain CT
Because of the convenience and low-cost operation, AI segmentation, bone injury identification, bone age
training tasks are gradually deployed on the cloud instead of analysis, organ delineation, dermatological auxiliary
the device. Meanwhile, edge computing developed on the diagnosis, etc. Some research even shows a better
devices is designed to be adaptive to application scenarios. It performance and efficiency than that provided by
is a blue sea with diverse forms and low competitiveness. IoT humans [29].
or wearables such as intelligent watches, headphones and
wristbands, and mobile phones are currently major drivers of ➢ Clinical decision support system (CDSS): Key steps
the edge market. AI inference tasks are increasingly such as laboratory test judgement and medical
deployed on devices to support the diversified scenarios and treatment are integrated with CDSS. Traditional CDSS
needs. builds on a top-down approach, with expertise and rules
based on expert systems to simulate the clinical
4.3 Closely integrated algorithm with health process decision-making process. AI based CDSS, without the
reliance on predefined rules, could ensure the
With large databases, high-performance computing, AI timeliness of evidence updates. Advanced natural
algorithms could strongly support and achieve personalized language processing, cognitive computing, automatic
medicine. The close combination between AI algorithms and reasoning and deep learning, etc. are used. AI-based
traditional health processes is the key to success. As is shown CDSS could greatly take full advantage of digital
in Figure 1, the framework of the health industry consists of medical data accumulated on a large scale in clinical
service and management. The integration can also be seen work in recent years, and overcome the weakness of
from these two perspectives. inefficiency in knowledge construction and limited
information coverage for traditional decision making,
Service process usually includes reservation, check-in & thus eventually accelerating industry development.
triage, payment/ pre-pay, symptom check, image inspection,

– 139 –