By Shane Walker
Radiologists attending the RSNA 2020 conference starting Sunday can expect to see a range of validated AI solutions—whether embedded, on premises, or in the cloud—being offered by companies eager to promote the technology, which can reshape medical imaging and transform how radiologists work.
Among companies with AI offerings that will have a presence at RSNA 2020 are Intel, with the Intel AI Builders ecosystem and OpenVINO developer’s toolkit; GE Healthcare
, with the Edison Open AI Platform and Edison Developer Program; Siemens Healthineers
, with the AI-Rad Companion product line and Digital Marketplace; Samsung Medison
, with the BiometryAssist and LaborAssist products; Philips Healthcare
, with the HealthSuite digital platform; Agfa HealthCare
, with the Enterprise Imaging solution; and Nvidia
, with Inception accelerator program and the Clara Imaging platform.
Radiologists, once wary that AI might altogether supplant their role in hospitals and treatment centers, are increasingly warming to AI because of the clinical and operational benefits possible with the proper use and implementation of the technology. Convening from Nov. 30 to Dec. 2 for the annual meeting of the Radiological Society of North America, an all-virtual event this year because of COVID-19, radiologists will see AI showcased in imaging demos, education sessions, and product exhibits. The international society of over 54,000 global members offers extensive resources on AI, and it issues an AI challenge each year to explore the ways that AI can benefit radiology and improve patient care.
A significant primary benefit in using AI is speed and capacity: Even the most skillful radiologists cannot come close to matching the sheer volume of data that can be analyzed by AI algorithms. Moreover, AI can find complex patterns in images, such as molecular markers in tumors not discernible to the human eye. AI-driven algorithms can also help improve diagnostic workflows by reducing time spent on routine or manual operations involving patient setup, screening, measurement, segmentation, and formatting.
But while there is enthusiasm among radiologists to incorporate AI into their practice, the knowledge to do so is often lacking, according to a study published in October in the journal Academic Radiology. The study found that almost 40% of the radiographers and radiologists participating in the survey were not familiar with AI, and that there appeared to be a mismatch between awareness of AI’s potential and expectations about its role. Given this discrepancy, it seems appropriate to provide some context for AI’s evolving acceptance in radiology—particularly with RSNA approaching.
Key concerns when evaluating AI in radiology diagnostics
As part of their work, radiologists may need to evaluate the merits of an AI solution or the soundness of the AI algorithms being deployed. To assess the efficacy of AI algorithms for use in the interpretation of medical images, a methodological framework in which important AI parameters are evaluated can be of immense value.
For instance, to understand an algorithm’s level of sensitivity and specificity in, say, differentiating lung cancer from benign diseases when examining lung nodules, a so-called receiver operating characteristic curve analysis can help determine the discrimination performance of the algorithm under consideration. Also of importance is calibration performance, which evaluates how similar predicted probability values will be to actual probabilities. The use of data collected with a minimum of spectrum bias is likewise of critical importance: Subjects forming the training data—both with and without disease—must be representative of the patients in whom the algorithm will be used, including severity or duration of the disease, presence and severity of comorbidities, and demographic characteristics. Otherwise, the results risk being biased and ultimately become clinically unusable.
Furthermore, unnatural disease prevalence within the data used to develop an algorithm could lead to problems. Determining the probability of lung cancer, using the example earlier, will be inaccurate in a population in which disease prevalence differs from that in the dataset used to develop the algorithm. Finally, beyond performance metrics, the best demonstration of clinical efficacy from a diagnostic AI tool is improved patient outcomes, which can be identified in clinical trials and observational outcome research.
features regulatory-approved diagnostic AI that analyzes medical images. The developer, product name, application, and image source for each is provided.
Various initiatives exist today for third-party providers to work with the AI development teams at some of the world’s biggest companies, producing AI-enabled solutions specific to healthcare that could impact the work of medical professionals, including radiologists.
, the chipmaking giant works closely with clients on integrating their AI mechanisms with Intel’s AI systems into a seamless and unified workflow that radiologists can then use. Intel sees its main role throughout the product development continuum as one of making algorithms work faster and of optimizing the pipeline. “Faster AI algorithms can translate into speedier results and seamless clinical workflows, a critical factor in clinician adoption. By optimizing these compute-intensive applications with tools like OpenVINO, clinicians don’t have to wait for AI, the AI becomes an invisible aid,” says Dr. Anthony Reina, chief AI architect for health and life sciences at Intel, and a medical doctor with extensive experience in neurophysiology, telemedicine, and data science.
Intel’s AI technologies, such as its OpenVINO software tool kit, can help clients innovate high-performance AI solutions that are also easy to scale. The company currently has 35 AI builders focused on healthcare alone. Claire Celeste Carnes, strategic marketing director for health and life science at Intel, points out that Intel is enabling AI across the company’s product line—not just on its Xeon Scalable Processors where AI acceleration is built in—allowing R&D to scale more quickly.
K Elizabeth Hawk, MD, PhD, a clinical instructor at Stanford University School of Medicine, department of radiology, says that AI should work as an ally in healthcare, not as an adversary. “I firmly believe that technology like AI should really be designed to deepen the relationship between the provider and the patient—not pull them further apart or put up a divide between the two,” she remarked.
To be sure, the use of AI in healthcare has grown tremendously. A survey undertaken by Intel in July among US healthcare leaders found that nearly 80% of those surveyed deploy AI or expect to do so in their clinical workflow, up a massive 30% from the same time two years earlier in 2018. And while issues of trust and cost remain barriers to adoption of AI, trust is growing in the technology, which has also been found to yield meaningful time savings.
At GE Healthcare
, its Edison intelligence platform can help providers achieve greater efficiency, improve patient outcomes, and increase access to care. Embedded within existing workflows, Edison applications can integrate and assimilate data from disparate sources, and apply analytics or advanced algorithms to generate clinical, operational, and financial insights. Edison solutions can be securely deployed in various ways: via the cloud; the Edison HealthLink technology that offers clinicians access to analytics tools to interpret data; or directly onto smart devices.
Meanwhile, the Edison Developer Program consists of identifying, analyzing, building, and testing a developer’s product, culminating in distribution across the entire GE Healthcare ecosystem. GE is also running accelerator programs in India and China, where they assist startups—typically without the resources to scale, and unfamiliar with how to best design their software to integrate into clinician workflows—prepare for market entry.
In cases where the solution is not FDA approved, GE Healthcare will work with the startup all the way through to the approval process. Market-ready solutions typically come to GE from providers, with the solutions having already received regulatory approval but needing assistance with scale, integration, and relationship-building with a broader network of providers. Among Edison’s more than 50 intelligent applications is Edison Open AI Orchestrator, which simplifies AI selection, deployment, and use in imaging workflows at scale. The GE Edison Developer Program also helps to solve IT ecosystem headaches, while the Edison Healthlink and Centricity Open PACS AI Solution supplies providers with the benefits of unified billing and trusted security certification.
Karley Yoder, vice president and general manager of artificial intelligence at GE, says the company continues its rapid pace of releasing products that leverage upstream AI, in which AI creates faster, higher-quality data and is often deployed on or near devices. Examples of impactful upstream AI include MR AIR RECON DL, X-ray Critical Care Suite, and TrueFidelity CT.
In its paper AI in Healthcare: Keys to a Smarter Future
, the company says embedding AI into clinical workflows will produce profound results for clinicians and patients alike. “AI holds tremendous promise to expand access to quality healthcare by freeing up human attention to focus on higher-value problem solving while ensuring a uniformly high quality of performance,” the paper stated.
GE Healthcare and Intel have partnered on several projects to improve the efficiency of AI in medical diagnostics. GE Healthcare’s AirX™ tool, for instance, accelerates magnetic resonance imaging by using Intel AI technologies. Tapping Intel Xeon CPU-based platforms and Intel Distribution of OpenVINO Toolkit, GE Healthcare designed AirX to precisely identify and align MR scans for diagnostic neuroimaging. GE Healthcare estimates that the new tool may decrease MR set-up time by 40% to 60% while increasing accuracy and consistency.
At Siemens Healthineers
, AI is prevalent across the German maker’s medical product line, including its computed tomography (CT) equipment with 3D camera for positioning and dose reduction tools. The company’s Digital Marketplace features 44 products from 10 partners as well as 28 Siemens Healthineers items, including 4 from its AI-Rad Companion product line, introduced in 2019, of AI-powered, cloud-based imaging software. The Al-Rad Companion Chest CT, for instance, detects lung nodules, and following automatic segmentation of the lung lesions, calculates the volume and maximum diameter of the lesions and the tumor burden. Siemens Healthineers develops its own AI at the Siemens Corporate Technology North American Headquarters located in Princeton, New Jersey, in collaboration with other teams from Germany and Bangalore, India, utilizing use-case collections representative of world markets.
Valentin Ziebandt, head of diagnostic imaging and digital health at Siemens Healthineers, notes that the company has processes in place to identify new AI vendors. For instance, if healthcare providers or customers inform Siemens Healthineers of a tool they already are using in clinical practice, Siemens Healthineers will investigate implementing the tool. However, if customers approach Siemens Healthineers with an application but no solution, the company will search for a solution or will develop one for the customer. A third process involves Siemens Healthineers surveying the AI landscape on its own and then identifying potential new solutions to investigate.
Other important players with AI healthcare-related offerings expected at RSNA this year are Samsung Medison, Philips Healthcare, Agfa HealthCare, and Nvidia.
, the medical imaging division of Samsung Electronics, uses an Intel Core i3 processor, the Intel Distribution of OpenVINO toolkit, and the OpenCV toolkit for the company’s BiometryAssist product that automates fetal measurements, Samsung Medison has also developed LaborAssist, which automatically estimates the fetal angle of progression during labor for a comprehensive understanding of a patient’s birthing progress, without the need for invasive exams.
Meanwhile, Philips Healthcare
provides its HealthSuite Insights and Insights Marketplace to support the adoption of analytics and AI in key healthcare domains. For its part, Agfa HealthCare
integrates an ecosystem of embedded algorithms into its Enterprise Imaging platform.
, the new giant in healthcare, will be highlighting the Clara Imaging platform, an application framework and partner ecosystem that brings together AI and smart sensors to improve patient care in healthcare facilities.
According to Nvidia, a goal of Clara Imaging is to increase AI adoption in healthcare by enabling tools that make data annotation, training, and deployment seamless for medical imaging applications. For instance, the Clara Train application framework facilitates the development of medical imaging applications with APIs that can add AI-assisted annotation to any medical viewer and a set of domain-specialized pre-trained models.
Nvidia recently announced it would acquire venerable UK chip designer Arm from Japan’s SoftBank Group for $40 billion. By combining Nvidia’s substantive knowhow in AI with Arm’s extensive ecosystem, Nvidia hopes to bring the power of AI and high-performance computing to practically every smart or IoT-connected entity in the world. The focus today on healthcare by Nvidia is understandable. “The world is confronting COVID-19, one of the greatest challenges in human history,” said Nvidia CEO Jensen Huang in remarks during the company’s annual meeting of stockholders.
To that end, Nvidia’s newest AI supercomputer, Cambridge-1, and Nvidia-accelerated computing are being deployed in the scientific community to help sequence and image the novel coronavirus causing COVID-19, to search for a vaccine or treatment, and to build robots to disinfect hospitals.
About the author: Shane Walker is a healthcare industry analyst and subject matter expert with a wide range of system level and component level technology coverage including investigative, diagnostic, and therapeutic equipment and software for clinical, ambulatory, and laboratory environments. Prior to founding Village Intelligence Corp. (www.villageintelligence.com), Mr. Walker served as a lead consultant for healthcare projects at IHS Markit and continues to serve on the board of the American Aerospace Technical Academy where he has provided research on the NDT industry. He can be reached at email@example.com.