Nvidia DGX A100 Aids VinBrain to Develop AI-Based Healthcare Solutions

Driven by its mission to infuse artificial intelligence (AI) into Healthcare, VinBrain has developed AI-based healthcare solutions that help doctors make faster and better-informed decisions, improving the quality and consistency of healthcare to the public.

Developed with the help of the powerful NVIDIA DGX A100 supercomputer, the AI-based DrAid doctor assistant providing automated chest X-ray diagnosis and screening to radiologists has been deployed in 63 health facilities in Vietnam and will be deployed in 150 more health facilities in the country and seven hospitals in Myanmar this year.

VinBrain is based in Hanoi, Vietnam and funded by Vingroup, the country’s largest conglomerate with a market capitalisation of around US$ 16 billion. Its development efforts are focused on applying AI in areas such as medical diagnostics, treatment, prediction and prevention, management and operational efficiency to enrich the quality of life of people everywhere.

By using AI advanced technologies such as deep learning, it aims to help radiologists detect diseases and abnormalities more accurately and consistently and at a faster speed. This is particularly impactful when 4.7 billion people in the world do not have adequate access to radiologists.

The firm has attracted a number of talented applied scientists with expertise in machine learning, computer vision, NLP and large-scale product and service development. Working in teams, these scientists reside in different parts of the world, including Vietnam, Australia, South Korea, and the United States.

Solving tougher problems

After using the previous generation NVIDIA DGX systems built on NVIDIA V100 GPUs for about a year to develop its solutions, the AI firm was among the first in the region to upgrade to the next generation NVIDIA DGX A100 toward the end of 2020.

NVIDIA DGX A100 is the world’s first five petaFLOPS AI system built on NVIDIA A100 GPUs. Featuring the NVIDIA A100 Tensor Core GPU based on the NVIDIA Ampere architecture, it lets enterprises consolidate training, inference and analytics into a unified, easy-to-deploy AI infrastructure.

“We want to utilise the DGX A100 to build more sophisticated models with faster training time to solve tougher problems in relation to medical diagnosis, treatment and prevention in healthcare. Using the NVIDIA DGX A100 helps us to accelerate the development process. The system also provides us with more GPU memory and computation capacity than the previous generation NVIDIA AI system,” said Steven QH Truong, CEO, VinBrain.

Developing solutions faster

Its applied scientists are working on a number of healthcare and smart city/smart home projects using the new system. These include DrAid Diagnosis, DrAid Treatment, DrAid Smart Assistant, Smart City, and Smart Home.

The DGX A100 provides the immense processing power needed to develop AI models for:

• Chest X-ray image classification and segmentation problems for DrAid Diagnosis

• MRI registration and segmentation problems for DrAid Treatment

• Speech recognition and text-to-speech problems to perform voice command and/or editing medical reports for DrAid Smart Assistant

“So far, our applied scientists have found that training and inference workloads for computer vision and natural language processing tasks on the NVIDIA DGX A100 server is about 30 to 50 percent faster compared to the prior generation. The eight NVIDIA A100 GPUs with 320GB total GPU memory enables them to perform large-scale training of speech recognition tasks. This helps greatly in speeding up the model training convergence, resulting in shorter turnaround time for model development,” said Truong.

Faster training and evaluation are extremely helpful when VinBrain’s applied scientists experiment with new ideas, enabling them to develop a Proof of Concept (PoC) in a short time.

“For example, when we need to do a PoC within a few days or explore new directions to improve a model’s performance, our applied scientists need fast GPU with large amounts of memory where they can run multiprocessing in parallel. The system helps to save model development time for them” he explained.

Leveraging existing platforms

Beyond the NVIDIA DGX A100, the Vietnam developer is also leveraging the NVIDIA Clara healthcare application framework for AI-powered imaging and speech-to-text streaming inference. This platform provides full-stack GPU-accelerated libraries, SDKs and reference applications for developers, data scientists and researchers to create real-time, secure and scalable solutions.

For its smart city/smart home projects, VinBrain is trying out NVIDIA Jarvis, a fully accelerated application framework for building multimodal conversational AI services that use an end-to-end deep learning pipeline. With Jarvis, developers can fine-tune state-of-art-models on their data to achieve a deeper understanding of their specific context. They can also optimise for inference to offer end-to-end real-time services that run in under 300 milliseconds and deliver 7x higher throughput on GPUs compared with CPUs.

“VinBrain is a classic example of a forward-looking company that believes in leveraging advanced technologies to drive breakthroughs. The NVIDIA DGX A100 supercomputer is designed to help researchers expedite scientific discoveries and develop AI solutions that advance healthcare and make the world a better place,” said Raymond Teh, Vice President, Worldwide Field Operations (Asia Pacific), NVIDIA

Utilizing these platforms from NVIDIA helps VinBrain to reduce development time as well as give it access to more intuitive solutions from the NVIDIA community.

Forging ahead

Having experienced success in its use of the NVIDIA DGX A100, VinBrain plans to use the system for upcoming projects involving conversational AI and advanced big models.

Another project in the pipeline is an AI-based system to provide prognosis recommendation that can be used by both doctors and non-doctors in healthcare.

“We have gained much productivity improvement by using the NVIDIA DGX A100 to build more sophisticated models with faster training. It is a worthwhile investment — powerful computing hardware at a reasonable cost” said Mr Steven QH Truong.