The Hematology Laboratory of the Future

Recent and profound advances in artificial intelligence (AI), data science, and computer vision are having a profound impact on many industries, including medicine. The clinical hematology laboratory and hematopathology are well-positioned to take advantage of this technological revolution to improve routine patient care and accelerate innovation.

A session in this year’s Scientific Program, “AI, Data Science, Computer Vision and the Hematology Laboratory of the Future,” is designed to brief attendees on the development and application of computational approaches for the quantitative analysis of tissue and liquid biopsy samples. Three experts will present information on this growing area of innovation, including David L. Jaye, MD. Here, Dr. Jaye previews his presentation on the use of computer vision and AI in routine clinical care.

What topics will you be discussing in your presentation?

I’m looking forward to discussing the state of the art in employing computer vision and AI to the analysis of blood and bone marrow specimens. Since this is a young field with relatively modest impact in clinical practice at this point in time, I will also discuss my expectations for future growth and some possible challenges in developing and applying these new technologies for routine patient care. I will be speaking from my perspective as a practicing hematopathologist, rather than a computer scientist, and will include some early data from my research using these technologies to automate differential cell counts in bone marrow aspirate smears.

Yinyin Yuan, PhD, and Metin Gurcan, PhD, who are outstanding computer scientists with expertise in computation pathology, AI, and machine learning, will discuss the more technical aspects of the field.

What is computer vision and how is it used in the hematology lab?

Computer vision broadly employs AI algorithms that can interpret, synthesize, and make extrapolations from images of any type. These are complex computer programs that are thought to emulate human cognition in some ways. A concrete example would be algorithms that could automatically differentiate digitized images of one cell or tissue type from another, such as different types of blood cells, or cancer from benign cells and tissues. In fact, the U.S. Food and Drug Administration (FDA) just authorized marketing of AI-based software to help pathologists identify prostate cancer from digital images. In radiology, the technology is more advanced and is used to help interpret some types of patient radiographic scans. These algorithms also may be able to extract information from images in novel ways that humans typically do not, such as deriving better predictions of outcomes or response to therapy, but such technologies are in developmental stages.

In the hematology lab today, there are limited uses of computer vision and AI, for the most part. Approved devices are, I believe, mostly restricted to the morphologic analysis of blood smears. Based on trained and validated computer algorithms, these technologies automatically locate and pre-classify white cells, red cells, and platelets on digitized images of blood smears into a limited but substantive number of categories (e.g., neutrophils, blasts, sickled red cells, platelet clumps, etc.). Cell images are then presented to a human expert, such as a technologist, hematopathologist, or hematologist, for final classification. So, in this case, the use of these algorithms falls into what we call a decision support role, since the human is the ultimate arbiter. The accuracies, sensitivities, and specificities for classification depend on the cell and disease type, but even as a pre-classifier, these technologies overall perform quite well with common cytomorphologic variants.

Since blood smear review is perhaps one of the most common lab tests and has been performed manually by microscopy with glass slides for decades, these computer vision- and AI-based systems, though imperfect, partially automate the process, fit well into our workflows, and thus continue to make us more efficient in the lab, which is a terrific advance.

How has the “technological revolution” of AI and data science affected your work in the lab?

As a hematopathologist, the impact has certainly not been sizeable to this point. But soon, these technologies will likely have greater impact on the interpretation of complex genomic and genetics data that are generated in the molecular diagnostic laboratory.

In the future, we might envision a scenario where AI-based algorithms can help us integrate complex data in more efficient and novel ways with practical applications in patient care. Such complex and varied information might include clinical history, radiologic data, laboratory test results, and information from digitized pathology images. Such a model is quite a way off, but it is certainly worth discussing the potential pros and cons before we deploy AI in medical practice. We only have to look at the failure of the IBM Watson AI system for clinical use several years ago as a cautionary tale about overpromising and underperforming on the capabilities of AI in medicine.

What are the challenges to applying these new technologies in routine clinical care?

Gaps certainly exist between AI development and clinical implementation. For computer vision/AI in hematopathology, we don’t have a validated commercially available system and platforms outside of blood smear analysis.

More generally, I think there are other, less technical issues for consideration. For example, do we want the AI-powered devices, software, and platforms to have regulatory approval before using them for routine clinical care? Although it is feasible to develop in-house methods that employ computer vision and AI for use at a single or a few institutions, it is difficult to quantify how common this is across intuitions and practices currently. The barriers to implementation can be high.

There are also other mundane, but critical, things to account for, including making a sound business case to obtain an adequate budget for implementation, determining how the technologies will fit into workflows and existing IT infrastructure, and understanding how the work using these technologies will be reimbursed.

In the lab, we also need to determine best practices for test validation, quality control, and quality assurance. Regulatory frameworks for use of software as a medical device are evolving around these many complex issues. For example, in the future, would we ever want to deploy an adaptive AI system that could continue to “learn” on its own after deployment? I am encouraged that national agencies and organizations such as the FDA and the Centers for Medicare and Medicaid Services, as well as pathology organizations, are grappling with these topics. Eventually, these discussions will help reduce the barriers to clinical implementation of these promising technologies.