Views on artificial intelligence (AI) today tend to fall into two camps: There are those who lean fully into its hype, making claims that AI is more profound than fire, and there are the cynics, who see it as a biased black box that can never compare to the abilities of the human brain. The truth is somewhere in the middle. In healthcare, AI does hold great promise for helping physicians and impacting patients, perhaps on a greater scale than common clinical technologies. Yet to deliver on that promise, it must be consistently tested and improved, and used not as the final word in diagnosis, but simply as a tool to aid physicians. This will ensure that AI always remains safe, effective, and equitable for patients.
The Promise of AI
Despite the mixed views on the importance of AI, it has undoubtedly become an integral part of our daily lives. There is little that AI hasn’t touched, from banking, to Netflix, even customer service chatbots. Naturally, AI has also made its way into healthcare, and now, pathology is the next logical step. Introducing AI into pathology certainly offers many benefits. First, AI can run automatic quality control, ensuring that digitized slide images that will be assessed are scanned clearly, in focus, and free from debris. Once digital cases have been created, AI can also help pathologists quickly triage cases by indicating which slides are suspicious for cancer. Upon review, AI can then guide the pathologist to the area on the slide itself that seemed potentially cancerous, or it can offer an instant second opinion when used after a pathologist’s initial read. Finally, AI offers quality assurance of the sign out process to further enhance pathologist and patient confidence.
Importantly, all of this can be done on hematoxylin and eosin (H&E)-stained slides. As it stands today, H&E is ubiquitous thanks to its low cost and the reproducibility of the pre-analytical process. With AI, we can pull even more information out of H&E-stained slides, including quantifying and grading tumors, and identifying known molecular biomarkers or even novel biomarkers that go beyond what is discernable for the human eye today. This gives pathologists the ability to create more comprehensive reports at the time of diagnosis, reducing the need for additional stains, minimizing delays, and accurately guiding additional testing or treatment.
Challenges and Considerations for Clinical Application
As a result of its prominence in routine clinical use, H&E provides a vast dataset from which to train AI, addressing one of the key challenges that comes with building clinical grade tools. Pathology, as we know, is an empirical science that can’t be learned from a book. To become an excellent pathologist requires the review of tens of thousands of slides that represent the breadth and diversity of biology. Therefore, if AI is to be used to support a pathologist in their diagnosis, it also has to be exposed to at least as many slides as a human expert to be equally talented. So, any AI trained on only a small, curated dataset, as many early iterations were, will be unfit for the realities of clinical practice. In essence, it would be like training a self-driving car in an empty parking lot when it needs to perform on busy highways; It would be bound to fail.
Another challenge with building pathology AI is the complexity of the machine learning system required. Pre-built, off the shelf machine learning models are not an option, because whole-slide images are far too large. Manual annotation will not work either, as it is highly subjective and impossible to do at scale. It also limits AI to learning only what pathologists already know, thereby eliminating the potential of AI to provide discovery. Instead, multiple instance learning (MIL) provides a scalable, reliable approach.
MIL is done by training the system from both whole-slide images and their corresponding pathology reports to indicate to the algorithm which slides contain at least one instance of cancer, and which do not. The algorithm then compares these images repeatedly, finding the difference between those slides that contain cancer and those that do not. After many iterations of this process, the AI is able to learn what cancer looks like such that when new slides are analyzed, it can not only detect whether cancer is present, but where it is located on the slide.
The Paige Approach
This is the exact approach Paige took to training our AI. To train the system from a large and diverse dataset, Paige and Memorial Sloan Kettering digitized over 5 million whole slides representing patients from around the globe. Paige’s AI modules were thoughtfully trained on this data to ensure they can be trusted and work robustly across institutions, diverse patient data, and regardless of pre-analytical slide variations. In a foundational study conducted on whole-slide images from over 15,000 patients across the globe, we found that Paige AI was able to detect prostate cancer reliably at an unprecedented clinical-grade level.1
Of course, we did not stop there. Continuous testing is key to ensuring AI remains safe and effective for real-world use for patients. Our latest study tasked 16 pathologists with the review of 610 whole-slide images prepared at multiple institutions globally. They reviewed the slides once without assistance, and then again with assistance from Paige AI. When Paige AI was used, diagnostic errors reduced by 70%.2
Paige’s rigorous approach to building and testing our AI ultimately helped us earn breakthrough designation from the FDA, a great leap forward for bringing pathology AI to the fore. Today, our AI for prostate cancer, Paige Prostate Detect, is the only FDA-approved AI-powered digital pathology product. Our hope is that it can help set the stage for the whole field of pathology AI to grow and positively impact patient care.
The Future of AI in Pathology
At the same time, we are further tapping into the potential of AI to teach us something new. Where with clinical AI we can make the existing diagnostic process easier, when applied to computational biomarkers, AI can help us discover entirely new ways of understanding and treating cancer.
Using H&E-stained slides, which are lower cost, faster, and more reproducible than other commonly used stains for biomarker identification, AI has the potential to identify the presence of known biomarkers as well as discover new biomarkers that are unknown to pathologists today. The benefit of this approach is that it reduces the subjectivity that comes with other biomarker tests and can be done in minutes rather than days or weeks. This means oncologists can make treatment decisions faster, which also benefits the patient. It can also guide enrollment in clinical trials, again laying the foundation for better patient outcomes. And better patient outcomes is what it’s all about! The true power of AI is the impact it can have on cancer research and clinical practice to transform patient lives.
References
1Campanella, G., Hanna, M.G., Geneslaw, L. et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med 25, 1301–1309 (2019)
2Raciti, Patricia., Sue, Jillian., et al. “Clinical Validation of Artificial Intelligence Augmented Pathology Diagnosis Demonstrates Significant Gains in Diagnostic Accuracy in Prostate Cancer Detection.” Archives of Pathology & Laboratory Medicine (In Press).
