We are all very familiar with and thankful for the sacrifices of front-line healthcare professionals in the COVID-19 pandemic. We’re less familiar, however, with the heroic efforts of public and private testing labs during the health crisis. These labs have performed millions of tests around the world, and cranked up a testing process from scratch in a short period. The process wasn’t perfect, and after the pandemic passes it will need to be improved. One key aspect of improvement will be greater use of digital and AI-assisted lab processes, which are just beginning to be developed.
COVID-19 is detected with molecular diagnostic lab processes, but a key function of labs is also anatomic pathology, which often involves the detailed examination of human cells and tissue. Pathology is finally beginning to move beyond the slide and the microscope for this task. Since the modern founder of pathology Rudolf Virchow encouraged his students to “think microscopically” in the 1850s, the focus of the field has been the systematic observation of cells in slides under a microscope. Most pathologists continue to work in microscope-equipped labs, rather than in front of computer screens. But “digital pathology,” in which pathologists examine digital images rather than slides, is on the rise, and artificial intelligence is likely to become an important component of it. No one, however, is predicting the imminent obsolescence of human pathologists.
In fact, they are needed now more than ever. The number of biopsies is increasing; a prostate cancer patient, for example, might have twelve biopsies collected in one needle. At the same time, the number of pathologists is falling—in part because of insurance reimbursement changes. An article in JAMA Network Open suggests the number of pathologists in the US fell by about 17.5% between 2007 and 2017; Canadian pathologists fell by more than 20%. Any technology that could improve their productivity and effectiveness would be good for the profession. According to one study by Quest Diagnostics, lab results account for 70% of decisions about inpatient care, but only 3-4% of budgets. Changes in pathology lab processes, then, could have significant impact on patient therapy and decision-making at relatively low cost.
Proscia, Digital Pathology, and AI
Proscia, a company founded in 2014 and based in Philadelphia, originally developed a digital pathology platform called Concentriq. It uses workflow-based software to automate, speed up, and provide analytics on the process of evaluating digital pathology images. It has also allowed—along with some temporary regulatory flexibility—pathologists to assess images from home during the coronavirus office closings. One of digital pathology’s challenges is that it lacks a common data standard for pathology images (unlike radiology, which has used the DICOM standard for several decades), but Concentriq works with a variety of proprietary standards for slide imaging and lab information systems. Proscia has established relationships with academic medical centers, hospitals, and commercial pathology labs.
As in radiology, once a medical field begins to rely on digital images, researchers and vendors are motivated to analyze them with AI. Deep learning, in particular, is an excellent technology for predicting whether an image is problematic from a disease standpoint. Proscia’s leadership team decided that they would focus first on identifying potentially cancerous cell images from dermatology biopsies. There are both efficiency and effectiveness reasons for that focus; a review of AI in pathology suggests that the algorithmic decisions in AI systems could reduce the relatively high rate of variability in pathologists’ judgments about whether a biopsy is cancerous.
Of course, for this strategy to work, Proscia needed access to images of biopsies that have been labeled by pathologists as malignant or benign, among other outcomes. There is no widely-available repository of such images, so Proscia’s approach was to partner with academic medical center pathology labs for access to their images.
Nathan Buchbinder, a co-founder and Chief Product Officer for Proscia, told me that the company’s Concentriq workflow offering and its first AI product, which it calls DermAI, are closely related. DermAI plugs into Concentriq and can perform tasks like preclassification of cases and workflow prioritization. What that means is—let me repeat—that no one is assuming that a pathologist won’t have to look at an image before declaring it cancerous. The process will allow an AI system to classify an image as “likely not to be cancerous,” or identify a potentially cancerous image by giving it a score, or find the most likely image among several for a pathologist to look at, or draw a circle around the suspicious part of the cell in an image.
The automation of those tasks could accelerate the work of pathologists, who might be able to look at fewer images or spend less time on them. Such efficiencies would allow them to spend more time communicating and translating results with a patient’s care team—an oncologist or surgeon, for example. There is a precedent for someone other than a full-fledged pathologist helping to make decisions on cases; residents in academic medicine often play that role. Buchbinder said that some Proscia customers compare their AI system to a “digital resident.”
To learn more, I talked with two pathologist customers of Proscia who are exploring AI-assisted pathology. But it’s important to point out that customers are using AI-assisted digital pathology for research, not day-to-day clinical practice yet. Neither the FDA nor the College of American Pathologists have signed off on AI for image analysis-based diagnosis. It seems likely that AI will first be adopted “for research use only,” and then later specific use cases would be approved for clinical use.
Dr. Alex Baras at Johns Hopkins
Alex Baras, MD/PhD, is the director of informatics for the department of pathology and the director of precision medicine informatics at the Johns Hopkins comprehensive cancer center. He has fully implemented the Concentriq digital pathology platform as part of routine pathology research at Johns Hopkins. As head of informatics, he’s a strong advocate for greater use of technology in pathology, including AI. He says that pathology is a data-rich field, but it has not fully leveraged this data in conjunction with the recent advances in machine learning and artificial intelligence. He commented that radiology has had the opportunity to embrace digital, biomedical imaging data given that this field has adopted imaging standards such as DICOM and has attracted a wide variety of data scientists for many years. Dr. Baras is strongly advocating for a similar movement in pathology and has supported this at Johns Hopkins with the development of an ongoing collaboration between the departments of pathology and biomedical engineering. Because of the rich data that he believes exists in pathology, the only way to synthesize it all is via quantitative metrics that could be employed in algorithmic decision-making and clinical decision support tools.
Dr. Baras pointed out that pathologists continue to use conventional slides and microscopes for two reasons:
First, the user experience of looking at a slide on a microscope is still superior to anything on a computer—you can move much faster on microscope, the optics of a microscope are superb, moving slide around is easy, and there are preset zoom levels. Looking at images on a computer is like navigating around Google Maps with keyboard and mouse — you need move the mouse around in a clunky way and you find yourself having to continuously zoom in and out. Second, the physical slide still needs to be generated and then scanned in. This is analogous to a radiographic film first being developed and then being scanned in on a flatbed scanner, which is certainly not how the majority of radiographic digital images are generated today.
To change all this, Dr. Baras indicated that the software needs to improve. Machine learning / AI can support pathologists by completing the less complicated tasks that pathologists and their support staff have to perform today. These types of functionalities will require a variety of “quantitative metrics” derived from digital pathology imaging. To get there, the profession needs to start generating a large corpus of data. For these reasons, Dr Baras and his colleagues have embarked on a large-scale annotated digital pathology data-generating collaboration with Proscia.
Dr. Jason Lee at Thomas Jefferson University Hospital
Dr. Jason Lee is a dermatologist and dermatopathologist at Thomas Jefferson University Hospital in Philadelphia, and teaches at Jefferson’s Sydney Kimmel Medical College. Ai-assisted pathology, he believes, is coming whether pathologists like it or not. He’s trying to embrace the idea and be a part of its adoption.
Lee’s view of the primary value of DermAI is triage—separating cases into one of several buckets for differential treatment. He was involved in a study with Proscia to learn just how well AI can do that. The study employed a total of 13,500 biopsy slides from four different labs (including his own) and scanned using a variety of whole slide scanners. The image classification models were trained on five different outcome categories. Lee concluded that the system categorized certain types of cases well, and on others was not as robust. But he said that he is confident that, for example, it will soon be able to accurately classify melanocytic (skin cell) lesions as either malignant, benign, or needing further examination. In other words, the system can’t yet provide a diagnosis with confidence, but it can provide triage. Of course, even that function would require FDA approval before clinical usage.
Dr. Lee doesn’t feel that the models used in these deep learning analyses will be interpretable by humans. He’s not particularly worried about it, however, if the classifications are accurate; human diagnoses are also often poorly understood and explained. Like Dr. Baras, he is hoping that the machines will be more accurate. If they are, he believes, it won’t be long until they are widely used in the clinical practice of pathology.
