CEO SUMMARY: Use of artifical intelligence (AI) to analyze digital pathology images and aid in diagnosis—or even in making the primary diagnosis—is much discussed. Experts in pathology regularly predict that use of AI in image analysis will transform the pathology profession. But that leaves one important question unanswered: When will AI be ready for prime time in the diagnosis of digital pathology images? In this exclusive interview, one expert explains how AI developers are tapping decades of lab test results to develop AI solutions for two common types of cancer.
FOR ALMOST TWO YEARS, anatomic pathologists have been bombarded by a seemingly-endless stream of press releases trumpeting some company’s new algorithm or image analysis solution that uses artificial intelligence (AI) to diagnose a whole-slide image (WSI).
However, these press releases leave two essential questions unanswered for pathologists interested in digital pathology. One: Is the AI in any vendor’s product robust and consistently accurate in the answer it produces? Two: Is the product truly ready for daily use in diagnosing cancers and other diseases?
Artificial intelligence is regularly touted as the most important technology poised to revolutionize anatomic pathology since pathologist Rudolf Virchow’s work with light microscopes in Germany 130 years ago. But surgical pathologists still wait for the first AI-based pathology product that, when used, transforms the basic diagnostic processes that pathologists use every day.
That day may not be far off, given the ongoing improvements to products that incorporate artificial intelligence, particularly in image analysis and diagnostics. For this reason, it is important for pathologists to understand the development curve for artificial intelligence.
One individual who is uniquely qualified to explain the technology development curve of artificial intelligence and its capabilities for use in anatomic pathology is Ajit Singh, PhD, a partner at Artiman Ventures in Palo Alto, Calif. He has a unique career trajectory involving imaging and informatics. In the 2000s, he was the CEO of Siemens Medical Solutions Image and Knowledge Management Group.
In 2008, Singh became CEO of BioImagene, one of the early entrants in the digital pathology marketplace. BioImagene was sold to Ventana Medical Systems, a division of Roche Diagnostics in 2010. Early in 2011, Singh joined Artiman Ventures. Over the past decade, he has been involved in diagnostic start-ups, several of which incorporate image analysis and artificial intelligence in their systems intended for use by pathologists.
Artificial Intelligence in 2018
In 2018, Singh was the closing speaker at the Executive War College and gave attendees a comprehensive presentation on artificial intelligence and its then-current state of development.
During his session, Singh pointed out that the capabilities of AI at that time made it effective for use in situations where there were not more than 15 to 20 variables in the problem to be solved. He gave the example of a patient who presents at the `physician’s office. Singh observed that the AI technology of 2018 was effective at managing the variables of:
- Who is the patient?
- Does identification presented to the physician match a real person?
- As of that date, does the patient have active health insurance?
- What is patient’s co-pay/deductibles with his/her coverage?
- How much of the yearly deductible/out-of-pocket has the patient met and how much of the patient-pay requirement does the physician need to collect?
To illustrate why the 2018 technology version of AI was not ready for use in anatomic pathology, Singh used the example of breast cancer. Because of the complexity of breast cancers, Singh observed that an AI solution would need a database of five billion breast cancer cases before the current technology of AI could reliably diagnose a breast cancer case with comparable accuracy to a trained pathologist.
Fast forward three years to today. What is different about artificial intelligence in 2021, compared to 2018? How has AI gained capabilities that make it ready to be a prime-time tool for pathologists in their daily work?
Singh has answers to these questions. In this recent interview with The Dark Report, Singh explained that multiple companies are bringing digital pathology analytical systems to market that utilize AI and demonstrate the ability to diagnose whole-slide images for at least two of the less complex types of cancer.
True of Prostate Cancer
“This is particularly true of prostate cancer, which has far fewer variables compared to breast cancer,” he said. “It is now possible to do a secondary read, and even a first read, in prostate cancer with an AI system alone.
“In cases where there may be uncertainty, a pathologist can review the images,” he continued. “Now, this is specifically for prostate cancer and I think this is a tremendous positive development for diagnostic pathways.
“Why and how did AI find its first success with prostate cancer?” asked Singh. “There are two reasons and they are familiar to all surgical pathologists. One, as noted earlier, the number of variables is less.
“Why and how did AI find its first success with prostate cancer?” asked Singh. “There are two reasons and they are familiar to all surgical pathologists.”
“Second is the pool of data about prostate cancer testing and outcomes that spans at least 35 years,” he noted. “Let me explain. The first FDA clearance for the PSA test was in 1986, and since then men have had PSA exams, even though this test is highly inaccurate for diagnosis. The inaccuracy of these PSA tests are recorded in medical records, along with the results of prostate needle biopsies and prostatectomies.
“Today, there are some 35 years of data that include PSA test results, prostate biopsies, prostatectomies, and diagnosis codes sitting in various electronic patient records and non-electronic patient records,” stated Singh. “In recent years, it became possible to digitize and assemble this data. From all that data comes an unexpected and exciting development for the use of artificial intelligence in prostate cancer diagnosis.
“Researchers and AI developers went back and looked at the prostate cases and began to identify the variables common to the cases that could be associated with the PSA scores,” he explained. “There were instances where the PSA test would say positive, but the patient’s biopsy showed no cancer. There were also cases where the PSA test showed negative, but the doctor observed hematuria or other symptoms and decided to do a prostate biopsy and discovered that the patient actually had a partial carcinoma.
“This was a gold mine of useful diagnostic data for the AI developers,” Singh said. “For large numbers of patients, they could look at the PSA scores, see the diagnoses, then look at the slides made from the biopsies to see what features and characteristics of the tissue could be associated with the PSA test variables.
“Working retrospectively with this data, the AI developers were able to identify the tissue structures consistent with the different variables,” added Singh. “Was it a positive PSA test and a negative biopsy? Was it a negative PSA and a positive biopsy? Researchers could now identify the tissue characteristics consistent with each type of diagnostic outcome and program the AI to accurately recognize and classify these different elements in a prostate cancer.”
State of Development
According to Singh, AI algorithms have reached the state of development where they also can be used in the diagnosis of skin cancers.
“AI is happening quickly in dermatopathology because melanomas have been tested and diagnosed in similar ways to prostate cancer,” noted Singh. “There are decades of patient cases where an initial indication caused the dermatologist to take multiple skin biopsies from the patient,” noted Singh. “However, when these biopsies were read by dermatopathologists, many of them were negative for cancer.
“Consequently, there are huge numbers of cases where researchers can see the initial symptoms that caused the physician to do skin biopsies, along with the final diagnoses. They then compare the slides made from the biopsies to see characteristics of the tissue associated with the negative diagnoses and positive diagnoses.
“Melanoma is a much less complex type of cancer than, say, breast cancer, so the decades of diagnoses and slides provided an immense amount of relevant data that AI developers could use to build their image analysis algorithms. Like with prostate cancer, this is an exciting development,” he emphasized.
Coming Next in Diagnostics?
What may come next with AI and cancer diagnostics? “Pathologists might want to watch the development of AI for use in diagnosing the types of cancers where diagnostic tools often trigger unnecessary biopsies,” predicted Singh. “The frequency of lung cancer and colon cancer would make each a good candidate for an accurate AI-powered diagnostic tool.
“Awareness of smoking as a cause of lung cancer gives that disease high visibility,” he continued. “Colon cancer is an interesting opportunity for AI because there is very low compliance on colonoscopy and there are frequent overcalls on colon cancer. For example, if polyps are found, it creates a concern for the physician and the patient. That concern leads to unnecessary biopsies.”
Singh also noted that cervical cancer—especially the type caused by the human papillomavirus (HPV)—is another type of cancer ripe for development of an AI diagnostic tool. “Like with prostate cancer and melanomas, there exists decades of HPV test data, Pap smear results, biopsy results, and patient outcomes,” observed Singh. “This huge volume of data is what allows researchers to develop and tune AI to acceptable performance for use diagnosing digital pathology images in support of clinical care.”
“Pathologists might want to watch the development of AI for use in diagnosing the types of cancers where diagnostic tools often trigger unnecessary biopsies.”
Because of the growing numbers of companies entering the anatomic pathology space with image analysis algorithms, machine learning products, and artificial intelligence tools, The Dark Report was interested to learn which companies or academic centers Singh would single out as worth watching.
“There many companies coming into the pathology AI market,” noted Singh. “Of these, I think three are farthest along with their AI offerings for analysis of pathology images. They are Ibex Medical Analytics (Tel Aviv, Israel), Paige.AI (New York, N.Y.), and PathAI (Boston).”
Ongoing Development of AI
Pathologists and clinical lab administrators following the development of artificial intelligence capabilities and how they are used in different aspects of healthcare—including surgical pathology—need to remember that the AI’s enabling technologies are being improved at a steady pace. We may still not have the self-driving car that was promised just a few years ago, but, as Singh points out, where there are fewer variables, artificial intelligence can already be used to great success and that is happening already in certain sectors of healthcare.
Contact Ajit Singh, PhD, at email@example.com.
How an Early Image Analysis Solution Created the TC-PC Model That Changed Anatomic Pathology
ONE CHALLENGE FOR ALL CLINICAL LABORATORY ADMINISTRATORS and pathologists is to know the definition of artificial intelligence (AI) and use that definition to correctly assess if any lab test system, product, software, or image analysis algorithm truly uses AI.
Pathologists with long memories recall ChromaVision Medical Systems, Inc., founded in 1993. At the time, Chromavisions’ flagship product—the ACIS System—was an image analysis tool that allowed pathologists to “detect, count, and classify cells of clinical interest based on recognition of cellular objects of particular color, size, and shape.” Its FDA clearance was as a staining device for cytokeratin 18.
Era of TC-PC Arrangements
Although this was not true artificial intelligence used to diagnose a digital pathology image, the ChromaVision system did transform anatomic pathology in a fundamental way. The ACIS was quickly adapted for use in measuring estrogen receptors in breast cancer. When used in this manner, the ACIS system became an essential tool in the earliest versions of the TC-PC (technical component-professional component) business model.
The TC-PC model was simple in concept and execution. Pathologists at community hospitals sent their breast cancer biopsies to a centralized laboratory. The referral lab processed the tissue to produce the glass slides and the images used by the ChromaVision system and billed for the TC. The lab then transmitted the digital images back to the referring pathologists, who then used the ChromaVision system to diagnose the case, thus allowing them to bill for the PC.
US Labs, a pathology company based in Irvine, Calif., was the fastest to jump on this TC-PC model in the early 2000s. It became the biggest buyer and user of ChromaVision ACIS systems.
Meanwhile, executives at nearby Chromavision Medical Systems watched the growth and profits at US Labs. They decided to restructure their company. They renamed it Clarient, Inc., and converted their instrument manufacturing company into a pathology laboratory organized around the TC-PC model. (See TDRs, January 3, 2005, and August 30, 2004.)
The TC-PC model was so attractive that LabCorp acquired US Labs in 2005. GE Healthcare described Clarient as a “molecular diagnostics and imaging firm” when it purchased the company for $580 million in 2010. GE divested Clarient to Neogenomics for $275 million in 2015. (See TDR, Oct. 26, 2015.)
The saga of ChromaVision and its pioneering system, which could do basic analyses from a digital image of a pathology slide, might be considered one of the earliest applications of a computer algorithm being used with a digital image of a pathology slide in support of clinical care.
Early Image Analysis
Of course, this happened with technology that dates back to 1993. Because the current generation of image analysis algorithms and artificial intelligence systems are much more robust and capable, the pathology profession may be poised for widespread adoption of AI for use in digital image analysis.
There is another lesson that pathologists and pathology practice administrators should take from ChromaVision’s role in expanding the use of the TC-PC business model. That lesson is that there is fast adoption anytime pathologists recognize something new can increase the revenue they generate from the cases they read.