New AI foundation model can detect rare cancers, but needs digital support to grow

Virchow, developed by New York-based digital pathology company Paige, is one of the largest image-based AI foundational models for cancer detection. Born from 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 cancer detection and diagnosis that was previously not possible.

Rare cancers, which account for more than 50% of all cancers, are exceptionally difficult to diagnose. Given that 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 hallmarks with 94% accuracy is significant, according to the company, and has the potential to impact hundreds of thousands of lives.

Paige says this is a sign that AI is becoming so advanced that it can detect cancers it hasn’t even been trained to detect yet.

Healthcare IT News spoke with Dr. David Klimstra, chief medical officer and co-founder of Paige’s Virchow Model, about the basic model, the complexities that rare cancers present for pathologists and their patients, how the breadth and depth of the basic model’s data enables the model to identify rare cancers, and a patient case study.

Q. Can you talk about the Paige Virchow Foundation Model, at a high level, and how it represents advances in AI for cancer diagnostics?

A. Training pathology AI models has evolved dramatically over the past five years. To “teach” the AI ​​model what cancer looks like, it was initially necessary to manually annotate the tumor in each training image. This process was so slow and tedious that it could not generate the amount of data needed to train clinical-quality AI models.

The first major step forward was multiple instance learning, where image annotation does not need 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 amount of data the models not only learned what features distinguished cancer from benign, but the models could also better generalize over the enormous variation in appearance that a cancer can have.

Paige used manifold learning to train the prostate cancer detection model, as published in Naturopathy in 2019 and then cleared by the FDA for clinical use in 2021To date, this algorithm is the only FDA-cleared AI product for use in surgical pathology. The FDA approval was based on a clinical study that demonstrated both improvements in sensitivity and specificity in the diagnosis of prostate cancer in core needle samples when using the AI ​​as a second-read tool.

Paige has continued to create AI diagnostic tools using multiple instance learning; while this method is much more efficient than manual image annotation, it still requires huge amounts of data that reflect the specific diagnostic task. For very common cancers it can work, but to help diagnose rare cancers where thousands of images aren’t available, a new method was needed.

Paige therefore developed Virchow, a base model trained on truly massive amounts of data (up to 3,000,000 images with over 1 billion parameters) representing the full spectrum of neoplastic diseases, non-neoplastic conditions, and normal histology. This kind 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.

Because the model has already been exposed to so much pathology image data, it can learn specific tasks very efficiently. Evidence of the improvements that come from using Virchow is published in Nature Medicineand we believe that this model, along with future second and third generation baseline models that Paige is developing, will form the basis for training all pathological AI in the future.

Not only does it enable the detection of rare cancers and unusual variants of more common cancers, it also offers the opportunity to train to detect important ‘digital biomarkers’, such as genomic alterations in cancers, based on subtle morphological clues in the routinely prepared pathology images.

Q. Can you describe the complexity that rare cancers pose for pathologists, and other clinicians, and their patients, as detection and diagnosis options are limited or non-existent?

A. According to the National Cancer Institute, slightly more than a quarter of all cancers are considered rare, based on a limited number of cases (fewer than 40,000 per year in the U.S.). Furthermore, all of the more common cancers have a range of pathologic variants, some of which have very distinctive histologic features, genetic changes, and clinical behavior.

Rare cancers and variants present challenges for diagnosis and treatment, as an individual pathologist or oncologist’s experience with each of the hundreds of subtypes is likely to be limited. If a pathologist has little experience with a rare variant, he or she may not recognize it or know how important it is to distinguish it from other cancer variants.

Accurate diagnosis and classification of cancers is the role of the pathologist, but for some such cancers, subspecialty expertise may be required to accurately provide important diagnostic information. Outside of large, specialized centers, diagnostic experience with rare cancers may not be sufficient for pathologists to make the most critical and accurate diagnoses.

Q. How does the breadth and depth of data from this base model position it to identify small, complex, and/or rare cancers and cancer markers, and how might this change the rules of the game for cancer pathology?

A. The Virchow Foundation model was exposed to virtually all types of rare cancers and unusual variants during training using millions of images from Memorial Sloan Kettering Cancer Center. This means that the computer model can potentially take on 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 in which Virchow-based diagnostic AI can be useful. Often, biopsies taken to diagnose a lesion or screen for cancer can contain very few cancer cells, compared to the other non-cancerous tissue elements in the sample. Even when these cancers have easily recognizable features of malignancy, finding them in a sea of ​​other cells – the proverbial “needle in a haystack” – can be very challenging and time-consuming for a pathologist.

In these types of tasks, AI models such as those based on Virchow are very suitable, because they can quickly assimilate all morphological findings in the image and draw the pathologist’s attention to very small suspicious areas for a final judgment.

Another application is to distinguish 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 the correct classification.

Finally, some pathologic diagnoses are highly subjective, requiring judgment to divide a continuum of histologic changes into discrete categories. Many studies have demonstrated significant diagnostic variation in these subjective determinations, such as grading the severity of a cancer precursor, but these diagnostic categories can have significantly different clinical management.

AI can make assigning cases to these subjective categories much more reproducible, as it removes the subjectivity of human interpretation. Although Although the development of effective AI models for subjective diagnoses is still in progress, there is potential to provide a solution to this difficult problem.

Q. Please share the patient case where the pan-cancer application was able to detect a rare and small metastatic foci of neuroblastoma in a pancreatic case and why this is important.

A. Cancer detection using AI models based on the Virchow basis model can help pathologists identify very small regions of rare cancers.

In a real-world example Paige encountered while validating the model to detect cancer in 17 different tissue types, a portion of resected, relatively normal-appearing pancreatic tissue was flagged by the model as suspicious for cancer.

Because these images were reviewed without any knowledge of the patient (age, sex, previous diagnoses, etc.), the most common cancer types one would expect to see on a pancreatic image would be ductal adenocarcinoma (the most common type of pancreatic cancer in adults) or a pancreatic neuroendocrine tumor.

In resection specimens, neither cancer type is usually very subtle, and examination of this tissue section quickly confirmed that neither was present. In fact, most of the pancreatic tissue appeared normal, except for a few clusters of small cells that initially appeared to represent benign inflammatory cells (lymphocytes).

However, based on the detection of these cells by the AI, the patient’s history was obtained. The patient was a child with a history of neuroblastoma, which usually arises in the adrenal gland. The neuroblastoma cells do indeed resemble lymphocytes, and once the patient’s history was known, the pathologist reviewer was able to verify that the AI ​​had indeed detected a cancer that is extremely rarely found in a pancreatic specimen.

This specific example provides concrete evidence of the capabilities we have encountered across a wide range of cancer types and variants: Virchow’s extensive training makes it possible to detect cancer types that are extremely rare in the practice of most pathologists.

Q. What should C-suite executives and other healthcare IT leaders at hospitals and health systems learn from all of this about the Paige Virchow Foundation Model?

A. The integration of AI into pathology practice has been slow, limited by the slow adoption of digital pathology platforms required to enable the use of AI tools. Digital adoption has been limited by the high cost of digitization, logistical challenges, and user reluctance after a century of practice using glass slides and microscopes.

As new AI tools are introduced, the increased efficiency and accuracy achieved with AI will increasingly justify the cost and effort of tackling digital pathology practices. Access to the Virchow foundation model means that the development of useful AI tools for pathology can be significantly accelerated.

Now, AI diagnostic tools can be developed more quickly, using smaller datasets, both by companies like Paige and, with access to Virchow, by academic departments interested in building their own AI. This means we can project a tipping point where the barriers to digital pathology adoption are outweighed by the benefits these technologies will bring to pathologists, treating clinicians, and their patients.

Follow Bill’s HIT reporting on LinkedIn: Bill Siwicki
Send him an email: bsiwicki@himss.org
Healthcare IT News is a publication of HIMSS Media.