New AI platform model can detect rare cancers – but needs digital support to become widespread
Virchow, developed by New York-based digital pathology company Paige, is one of the largest image-based AI platform models for cancer detection. Born out of a collaboration between Paige and Microsoft Research, Virchow has become proficient at diagnosing small, complex, and rare cancers to provide pathologists with a level of insight into cancer detection and diagnosis that was previously impossible.
Rare cancers, which account for more than 50% of all cancers, are notoriously difficult to diagnose. Considering that more than 70% of cancers occurring in children and adolescents are rare cancers that were previously undetectable, the impact of the model’s ability to identify these cancers and their cancer characteristics with 94% accuracy is significant, according to the company, and could potentially impact the lives of hundreds of thousands of people.
According to Paige, this signals that AI is becoming so advanced that it can detect cancers that it hasn’t even been trained on.
Healthcare IT News sat down with Paige co-founder Dr. David Klimstra to discuss the underlying model, the complexities that rare cancers pose to pathologists and their patients, and how the underlying models leverage deep data to identify rare cancers and patient cases.
Q. Please talk about the Paige Virchow Foundation Model at a high level and how it represents advances in AI for cancer diagnosis.
ONE. Training pathology AI models has come a long way in the past five years. Initially, to “teach” an AI model what a cancer looks like, it was necessary to manually annotate the tumor in each training image. This process was so slow and tedious that it was impossible to generate the volume of data needed to train clinical-grade AI models.
The first big step is case learning, where there is no need to do image annotation and the only annotation given to the computer is whether a given image contains cancer or not. This allows tens of thousands of images to be used, and with that amount of data, models can not only learn the features that distinguish cancer from benign, but can also better generalize across the wide range of appearance variations that cancers can have.
Paige used multiple instance learning to train the prostate cancer detection model, as published in Nature Medicine in 2019 and then FDA approved for clinical use in 2021. To date, the algorithm remains the only AI product approved by the FDA for use in surgical pathology. The FDA approval was based on a clinical study that showed both sensitivity and specificity improved in diagnosing prostate cancer in core needle specimens when using AI as a second-reader.
Paige has continued to build diagnostic AI tools using multiple instance learning; however, while this approach is much more efficient than manually annotating images, it still requires a large amount of data that reflects the specific diagnostic task. It can work for very common cancers, but to help diagnose rare cancers without thousands of images, a new approach is needed.
Paige therefore developed Virchow, a foundational model trained on a truly massive amount of data (up to 3,000,000 images containing over 1 billion parameters) reflecting the entire spectrum of neoplastic, non-neoplastic, and normal histology. This type of self-supervised or generative AI can learn with fewer data labels, creating a virtual encyclopedia of pathological knowledge that can be applied to any downstream diagnostic task.
Since the model has been exposed to a lot of pathological imaging data, it can learn specific tasks very effectively. Proof of improvement is allowed for use Virchow has been published in Nature Medicine and we believe that this model, along with the second and third generation platform models that Paige is developing, will be the basis for training all pathology AI in the future.
Not only does it enable the detection of rare cancers and unusual variants of more common cancers, it also allows training to detect important “digital biomarkers,” such as genomic alterations in cancer, based on subtle morphological clues in conventionally prepared pathology images.
Q. Please describe the complexity that rare cancers pose to pathologists, other clinicians, and their patients because of limited or non-existent detection and diagnostic capabilities.
ONE. According to the National Cancer Institute, just over a quarter of all cancers are considered rare, based on the limited number of cases (less than 40,000 cases per year in the United States). In addition, all of the more common cancers have many pathologic variants, some of which have very distinctive histologic features, genetic alterations, and clinical behavior.
Rare cancers and variants present a challenge to diagnose and manage, as an individual pathologist or oncologist may have limited experience with each of the hundreds of subtypes. If a pathologist has little experience with a rare variant, they may not recognize it or they may not know the importance of distinguishing it from other cancer variants.
The role of the pathologist is to accurately diagnose and classify cancers, but for some cancers, specialist expertise may be required to provide accurate diagnostic information. Outside of major specialist centres, the experience of diagnosing rare cancers may not be sufficient for pathologists to make the most important and accurate diagnosis.
Q. What breadth and depth of data does this platform have to be able to identify small, complex, and/or rare cancers and cancer features, and how could this change the game in oncology?
ONE. The Virchow platform model has been exposed to virtually every rare cancer type and uncommon variant during training with millions of images from Memorial Sloan Kettering Cancer Center. This means the computer model can essentially take on the role of an entire team of semi-specialized experts and not rely on the personal experience of a single pathologist to recognize unusual diagnostic findings.
There are a number of ways that Virchow-based diagnostic AI could be useful. Many times, biopsies taken to diagnose lesions or to screen for cancer may contain very few cancer cells, compared to the other non-cancerous tissue components in the sample. Even if these cancers have recognizable features of malignancy, finding them in a sea of other cells—the proverbial “needle in a haystack”—can be difficult and time-consuming for a pathologist.
In this type of task, AI models such as Virchow-based models prove to be very effective as they can quickly assimilate all the morphological findings in the image and draw the pathologist’s attention to very small suspicious regions to make a final judgment.
Another application is to distinguish rare variants, which a pathologist may have limited experience with. An AI model can draw on extensive prior exposure to even rare cancers to help the pathologist make the right classification.
Finally, some pathologic diagnoses are highly subjective, requiring judgment to separate a spectrum of histologic changes into distinct categories. Many studies have shown significant diagnostic variation in these subjective determinations, such as grading the severity of precursor neoplasms, but these types of diagnoses may have significantly different clinical management.
AI can make classifying cases into these subjective categories much more reproducible because the subjectivity of human interpretation is removed. Although Although the development of effective AI models for subjective diagnosis is still a work in progress, it has the potential to provide a solution to this difficult problem.
Q. Let’s talk about a patient case where systemic oncology was able to detect small and rare metastases of neuroblastoma in a pancreatic case and why this is important.
ONE. Cancer detection using an AI model based on the Virchow platform model can help pathologists identify very small regions of rare cancers.
In one real-world example, discovered during Paige’s validation of the model to detect cancer on 17 different tissue types, a section of resected pancreatic tissue that looked relatively normal was flagged by the model as suspicious for cancer.
Since these images are evaluated without knowing patient information (age, gender, previous diagnoses, etc.), the most common cancers one might encounter in pancreatic imaging are ductal adenocarcinoma (the most common type of pancreatic cancer in adults) or pancreatic neuroendocrine tumors.
In resection specimens, none of these cancers are typically very subtle, and examination of the tissue quickly confirms that none are present. In fact, most of the pancreatic tissue appears normal, with the exception of a few small clusters of cells that initially appear to represent benign inflammatory cells (lymphocytes).
However, based on the AI’s detection of these cells, the patient’s medical history was gathered. The patient was a child with a history of neuroblastoma, most commonly found in the adrenal gland. Neuroblastoma cells actually resemble lymphocytes, and knowing the patient’s medical history, the pathologist was able to verify that the AI had indeed detected an extremely rare type of cancer in the pancreatic sample.
This particular example provides concrete evidence of the ability we found across a wide range of cancers and variants – that Virchow’s extensive training allows for the detection of extremely rare cancers within the practice of most pathologists.
Q. What should senior executives and other health IT leaders at hospitals and health systems take away from all of this about the Paige Virchow Foundation Model?
ONE. The introduction of AI into pathology practice has been slow, constrained by the slow adoption of digital pathology platforms needed to enable the use of AI tools. Digital adoption is constrained by the high cost of digitization, logistical challenges, and user reluctance after a century of practice using microscopes and microscopy.
As new AI tools are introduced, the increased efficiency and accuracy achieved using AI will increasingly justify the cost and effort of implementing digital pathology. Adopting the Virchow platform model means that the process of developing useful AI tools for pathology can be significantly accelerated.
Today, AI diagnostic tools can be developed faster, using smaller datasets, both from companies like Paige and from Virchow, and from academic departments interested in building their own AI. This means we can anticipate an inflection point where the barriers to adopting digital pathology are offset by the benefits these technologies will bring to pathologists, treating clinicians, and their patients.
Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: [email protected]
Healthcare IT News is a publication of HIMSS Media.
