Virchow, developed by New York based digital pathology company Paige, is one of the largest image-based AI foundation models for detecting cancer. Born out of a partnership between Paige and Microsoft Research, Virchow has mastered the complexities of diagnosing small, complex and rare cancers to give pathologists a level of insight into detecting and diagnosing cancer that previously was not possible.
Rare cancers, which represent more than 50% of all cancers, have been exceptionally hard to diagnose. Considering more than 70% of cancers occurring in children and adolescents are previously undetectable rare cancers, the impact of the model’s ability to identify these cancers and cancer features with 94% accuracy, according to the company, is significant and has the potential to affect hundreds of thousands of people’s lives.
It signals that AI is becoming so advanced it can detect cancers it hasn’t even been trained on, according to Paige.
Healthcare IT News sat down with Dr. David Klimstra, chief medical officer and cofounder of Paige’s Virchow model, to discuss the foundation model, the complexity that rare cancers present for pathologists and their patients, how the foundation models’ breadth and depth of data position it to be able to identify rare cancers, and a patient case.
Q. Please talk about the Paige Virchow Foundation Model, at a high level, and how it represents progress in AI for cancer diagnosis.
A. The training of pathology AI models has evolved tremendously over the past five years. Initially, in order to “teach” the AI model what cancer looks like, it was necessary to manually annotate the tumor in each training image. This process was so slow and tiresome that the volume of data required to train clinical-grade AI models could not be generated.
The first major step forward was multiple instance learning, in which image annotation does not have to be performed, and the only annotation provided to the computer is whether or not a given image contains cancer. This allowed tens of thousands of images to be used, and with that volume of data, not only did the models learn what features distinguish cancer from benign, but the models could better generalize across the vast array of variation in appearance a cancer can have.
Multiple instance learning was used by Paige to train the prostate cancer detection model, as published in Nature Medicine in 2019 and subsequently cleared by the FDA for clinical use in 2021. To date, this algorithm remains the only FDA-cleared AI product for use in surgical pathology. The FDA clearance was based on a clinical study that showed both sensitivity and specificity improvements in the diagnosis of prostatic carcinoma in core needle specimens when using the AI as a second-read tool.
Paige has continued to create diagnostic AI tools using multiple instance learning; however, even though this method is much more efficient than manual image annotation, it still requires huge amounts of data reflective of the specific diagnostic task. For very common cancer types, it can work, but to help diagnose rare cancers for which thousands of images are not available, a new method was needed.
Hence, Paige has developed Virchow, which is a foundation model trained on truly massive amounts of data (up to 3,000,000 images containing more than 1 billion parameters) reflecting the full spectrum of neoplastic diseases, non-neoplastic conditions and normal histology. This sort of self-supervised or generative AI can learn with even fewer data labels, creating a virtual encyclopedia of pathology knowledge that can be applied to any downstream diagnostic task.
Since the model has already been exposed to so much pathology image data, it can learn specific tasks very efficiently. Evidence of the improvements allowed using Virchow has been published in Nature Medicine, and we believe this model, along with future second and third generation foundation models Paige is developing, will be the basis for training all pathology AI in the future.
Not only does it allow detection of rare cancers and unusual variants of more common cancers, but it also allows training to detect important “digital biomarkers,” such as genomic alterations in cancers, based on subtle morphologic clues within the routinely prepared pathology images.
Q. Please describe the complexity that rare cancers present for pathologists, among other clinicians, and their patients as detection and diagnosis capabilities have been limited or nonexistent.
A. According to the National Cancer Institute, just over one quarter of all cancers are considered rare, based on limited numbers of cases (less than 40,000 per year in the U.S.). Additionally, all more common cancers have a range of pathological variants, some of which have very distinctive histologic features, genetic alterations and clinical behavior.
Rare cancers and variants represent a diagnostic and management challenge, since the experience of an individual pathologist or oncologist with each of hundreds of subtypes is likely to be limited. If a pathologist has little experience with a rare variant, they may not recognize it, or they may not know the importance of its distinction from other variants of cancer.
Accurate diagnosis and classification of cancers is the role of the pathologist, but for some such cancers, it may require subspecialty expertise to precisely provide the important diagnostic information. Outside of large, specialized centers, diagnostic experience with rare cancers may not be adequate for pathologists to provide the most critical and accurate diagnoses.
Q. How does this foundation model’s breadth and depth of data position it to be able to identify small, complex and/or rare cancers and cancer features, and how might this change the game for cancer pathology?
A. The Virchow foundation model has been exposed to essentially all types of rare cancers and uncommon variants during its training with millions of images from Memorial Sloan Kettering Cancer Center. This means the computer model can essentially assume the role of an entire team of subspecialized experts and does not have to rely on the personal experience of an individual pathologist to recognize unusual diagnostic findings.
There are several ways diagnostic AI based on Virchow can be helpful. Many times, biopsies taken to diagnose a lesion, or to screen for cancer, may contain very limited cancer cells, compared with the other non-cancerous tissue elements in the sample. Even when these cancers have readily recognizable features of malignancy, finding them within a sea of other cells – the proverbial “needle in the haystack” – can be very challenging and time-consuming for a pathologist.
In this type of task, AI models like those based on Virchow excel, as they can rapidly assimilate all of the morphologic findings within the image and draw the pathologist’s attention to very small suspicious regions for final adjudication.
Another application is in the distinction of rare variants, with which an individual pathologist may have limited experience. An AI model can draw on the vast prior exposure to even rare cancer types to help the pathologist with proper classification.
Finally, some pathology diagnoses are highly subjective, requiring judgment to separate a continuum of histologic alterations into discrete categories. Many studies have shown significant diagnostic variation in these subjective determinations, such as grading the severity of a cancer precursor, yet these diagnostic categories can have significantly different clinical management.
AI can make assigning cases to these subjective categories much more reproducible since the subjectivity of human interpretation is removed. Although developing effective AI models for subjective diagnoses remains a work in progress, there is potential to provide a solution for this vexing problem.
Q. Talk about the patient case where the pan cancer application was able to detect a rare and tiny metastatic foci of neuroblastoma within a pancreatic case and why this is significant.
A. Cancer detection using AI models based on the Virchow foundation model can help pathologists identify very small regions of rare cancers.
In one real-world example, encountered during Paige’s validation of the model to detect cancer across 17 different tissue types, a section of resected relatively normal-appearing pancreatic tissue was flagged by the model as suspicious for cancer.
Since these images were assessed without any knowledge of the patient (age, gender, prior diagnoses, etc.), the most common cancer types one would expect to encounter in a pancreas image would have been ductal adenocarcinoma (the most common type of pancreas cancer in adults) or a pancreatic neuroendocrine tumor.
In resection specimens, neither of these cancer types is usually very subtle, and examination of this tissue section quickly confirmed neither was present. In fact, most of the pancreatic tissue appeared normal, with the exception of a few clusters of small cells that initially appeared to represent benign inflammatory cells (lymphocytes).
However, based upon the detection of these cells by the AI, the prior history of the patient was obtained. The patient was a child with a history of neuroblastoma, which most commonly arises in the adrenal gland. The cells of neuroblastoma indeed resemble lymphocytes, and once the patient’s history was known, the pathologist reviewer could verify that the AI had indeed detected a cancer that is extremely rarely encountered in a pancreas specimen.
This specific example provides concrete evidence of the capabilities we have encountered across a broad array of cancer types and variants – that the extensive training of Virchow enables detection of cancer types that are extremely uncommonly seen in the practice of most pathologists.
Q. What do C-suite executives and other health IT leaders at hospitals and health systems need to take away from all of this about the Paige Virchow Foundation Model.
A. The incorporation of AI into the practice of pathology has been slow, limited by the slow adoption of digital pathology platforms that are needed to allow the use of AI tools. Digital adoption has been limited by high costs of digitization, logistical challenges, and user reluctance after a century of practice using glass slides and microscopes.
As new AI tools are introduced, the greater efficiency and accuracy achieved using AI will increasingly justify the costs and effort to take on digital pathology practice. Having access to the Virchow foundation model means the development of useful AI tools for pathology can be significantly accelerated.
Now, AI diagnostic aids can be developed more quickly, using smaller datasets, both by companies like Paige and also, with access to Virchow, by academic departments interested in building their own AI. This means we can project an inflection point where the impediments to adopting digital pathology are outweighed by the benefits these technologies will offer pathologists, treating clinicians, and their patients.
Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.