NVIDIA Addresses AI Compute Challenges for Healthcare Startups with Clara Platform and GPU-Accelerated Solutions

For AI companies, computing power has never been a low-cost expense. As the driving force behind algorithmic advancement, how to rationally utilize computing resources to accelerate algorithm iteration has become an unavoidable challenge for startups.

Although NVIDIA has launched AI chips tailored for different scenarios to accelerate algorithm execution, the exponential growth of data—from imaging to cytology and genomics—makes it difficult for medical AI companies to identify a suitable positioning and select chips that are universally applicable across the entire development workflow.

At the Emtech China 2019 Global Emerging Technology Summit, which recently opened in Beijing, NVIDIA Vice President Kimberly Powell delivered a keynote address themed around the development of artificial intelligence. While discussing AI advancements across various industries, she also highlighted NVIDIA’s computing power solutions for the healthcare sector.

NVIDIA Vice President Kimberly Powell

Medical artificial intelligence has relied on a variety of tools to collect health data from the human body since its inception. In 2017, NVIDIA combined a specific machine with novel detection technology to record protein data at the atomic level, collecting up to 3 terabytes (TB) of protein data per day, while the volume of genetic data is far greater.

The process of collecting receipt data is ongoing, yet processing this data proves exceptionally challenging. Kimberly Powell described the disarray of existing data as “chaotic,” arguing that under such circumstances, we must leverage AI to address these computational challenges.

NVIDIA’s three major partners—GE, Canon, and Siemens—have long achieved substantial research results in this area.

Kimberly Powell explained to VCBeat how these giants are leveraging AI to optimize device imaging: Canna has developed an AI real-time algorithm for CT scanners, helping to reduce imaging time and generate more safe and accurate real-time images.

GE’s hardware innovations are equally critical: in the event of an intracranial hemorrhage, GE’s imaging systems can assist radiologists in reorganizing their workflows and reprioritizing tasks. Based on the severity of patients’ symptoms, the system can reorder the work queue to guide clinical decision-making.

Siemens has also achieved numerous technological breakthroughs in the field of AI. It has launched a highly secure and standardized anthropometric platform that integrates AI into CT scanning, leveraging its knowledge graph to provide diagnostic recommendations for patients.

Beyond these medical device giants, startups are also carving out entirely new niches and leveraging AI technologies to address them.

Genomic research and new drug development represent significant applications of AI beyond medical imaging. Toptom has leveraged AI to compare 72,000 proteins, examining their mutual interactions. Additionally, the company employs Generative Adversarial Networks (GANs), a deep learning technology that assists researchers in designing compounds; to date, it has generated 5,000 such compounds.

Meanwhile, artificial intelligence technologies can leverage computer vision and screening techniques to gain a comprehensive understanding of intracellular compounds and their interrelationships, thereby elucidating how certain crystals are developed in the context of new drug discovery.

The aforementioned high-performance computing and artificial intelligence technologies are inseparable from computational power support. At the 2018 Radiological Society of North America (RSNA) annual meeting, NVIDIA launched the Clara medical imaging supercomputing platform, aiming to provide unified support services for all medical imaging applications.

Clara SDK provides medical application developers with a suite of GPU-accelerated libraries for computing, advanced visualization, and AI. As Clara SDK evolves, we will also provide containers that can be used to build hardware-abstracted applications. These containers enable medical image reconstruction, image processing, segmentation, classification, and 3D rendering.

By leveraging Docker and NVIDIA’s Kubernetes on GPUs, developers can deploy applications across multiple computing environments, including embedded, on-premises, or cloud-based systems.

When it comes to treatment and diagnosis, radiologists often need to spend hours carefully examining a patient’s 3D images. This is a tedious process in which radiologists must review CT or MRI scans slice by slice, manually delineating, annotating, and correcting the organs or abnormalities of interest, and then repeating this step for all 3D image slices of the specific organ or abnormality.

NVIDIA’s AI-assisted annotation SDK can accelerate this process by 10-fold, facilitating faster detection of abnormalities. This is achieved by enabling application developers and data scientists to integrate the AI-assisted annotation SDK into their existing applications and apply AI-assisted workflows to radiography.

The AI-assisted annotation SDK leverages NVIDIA’s Transfer Learning Toolkit to continuously self-learn, enabling each newly annotated image to serve as training data and further enhance the accuracy of the provided pre-trained deep learning models.

“We can obtain NVIDIA’s AI-assisted annotation technology and integrate it into our image browser within a matter of days,” said Mark Michalski, Executive Director of the MGH & BWH Center for Clinical Data Science. “We currently need to annotate a large volume of images—sometimes around 1,000 or more per day—so any technology that helps automate this process can significantly reduce annotation time and costs. We are very excited to leverage AI-assisted workflows and collaborate with NVIDIA to address these critical medical imaging challenges.”

Regarding the application of Clara in China, Kimberly Powell stated: “The same software suite can be deployed either on-premises at hospitals or in the cloud. For the Chinese market, I believe that supporting such a hybrid operational environment offers significant advantages. We recognize that in some remote provinces or rural areas of China, where network connectivity is poor and cloud services are inaccessible, local deployment can be chosen. Conversely, hospitals in major cities, equipped with robust hardware infrastructure, can opt for cloud-based operation.”

This model has long been applied in the fields of gaming and autonomous driving, and its application in the medical sector will become increasingly complex. In addition to cloud-based solutions such as the Clara platform, some companies have chosen other means to deploy computing power for their projects.

To better optimize the performance of its algorithmic models, Hisi Heterogeneous Computing, which focuses on research in fields such as digestive endoscopy and ultrasound, has established a dedicated R&D platform for artificial intelligence technologies in medical imaging. Built with 64 NVIDIA Tesla V100 GPUs, the platform delivers outstanding computational power, reducing the training time for models that traditionally required 15 days to just 52 minutes. The company’s self-developed supercomputing parallel training software maintains 90% linear acceleration on a 1,024-GPU system.

Other innovative models are still underway. Chongqing-based startup Titanium Star Blockchain has set its sights on mining rigs abandoned in the wake of Bitcoin’s sharp price collapse. With specialized processing, these machines, originally designed for hashing power, can be daisy-chained to provide computational support for AI workloads. Compared with NVIDIA’s Clara Annotation Assistant SDK, this approach may seem crude and straightforward, but it remains an excellent example of upcycling waste hardware.

Overall, the healthcare industry will have the greatest demand for computing power worldwide. As researchers gain insights into the molecular, atomic, and even more microscopic realms, the provision of computational power may foster more innovative business models. NVIDIA must remain vigilant to maintain its leading position.