A Real PAIGE-Turner
Revolutionizing cancer diagnostics through computational power – exploring the ultimate ambitions of the Pathology Artificial Intelligence Guidance Engine
Luke Turner | | Longer Read
At a Glance
- Paige.AI is striving to shift pathology from a qualitative field to a quantitative one by digitizing vast number of slides and creating decision support systems
- Applications include clinical uses to save time for pathologists and an image retrieval technique to identify similar cases
- Collaboration with Memorial Sloan Kettering (MSK) Cancer Center facilitates the digitization of 30,000 slides per month, now being ramped up to 100,000 to develop effective algorithms
- Paige.AI has recently been granted Breakthrough Device designation by the FDA, boosting its quest for faster and more accurate diagnosis
“PAIGE helps pathologists to be more effective, researchers to be more quantitative, and patients to be more confident in their diagnosis,” reads the splash page for the Pathology Artificial Intelligence Guidance Engine (PAIGE). But what does “more effective” truly mean in a pathologist’s workflow? To find out, we have to dive into the origin of the venture – a computational pathology paper published in 2008.
Against the backdrop of a digital transformation that is rapidly gaining traction in pathology, Paige.AI’s mission is simple: to shift the discipline from a qualitative to a quantitative one. To improve the quality of image processing in pathology, Paige.AI, in partnership with the Memorial Sloan Kettering (MSK) Cancer Center, is digitizing vast numbers of slide images to develop decision support systems for pathologists.
“We have two baskets of applications. One is aimed at clinical pathology uses, where we take tasks that pathologists already carry out and help them to be faster, more robust, or more reproducible,” says Thomas Fuchs, founder of Paige.AI. “The other part relates to integrating pathology data with other important healthcare data to develop new diagnostic approaches and treatment paradigms, which involve processes that pathologists could not complete without computational pathology and PAIGE.
For example, if you notice a strange lesion or some unusual morphology in a patient’s sample, at the moment, you can only query your own brain or ask someone for a second opinion if they are available.” What PAIGE aspires to achieve is an image retrieval technique that marks the region in question and searches through an entire archive of cases to identify patients with similar tissue morphology. Based solely on an image (rather than text), the search would allow pathologists to learn which treatments those patients received and what their outcomes were, leading to better treatment selection for current and future patients.
Only last month, Paige.AI announced that it has been granted Breakthrough Device designation by the US Food and Drug Administration (FDA). As what appears to be the first-ever such designation for artificial intelligence (AI) in cancer diagnosis, this represents a significant milestone and underlines Paige.AI’s ambition to deliver effective diagnostic technologies for cancer patients that surpass those already available. It will enable timely access to Paige.AI’s technologies for patients and health care providers alike by accelerating the development, assessment, and review of breakthroughs. “We are honored to achieve this designation by the FDA, which demonstrates the groundbreaking nature of our technology,” says Fuchs. “We see it as the next step to producing leading clinical-grade AI in computational pathology, combining vast amounts of high-quality data with unique deep learning architectures to deliver better patient care.”
One of the ambitions of the project is to create powerful predictive technologies that span multiple cancer types – a goal MSK facilitates by providing a platform to prospectively scan a large number of slides. “We can easily multiplex from one organ type or task to another. We are, of course, focused on the most prevalent types of cancer – such as prostate, breast, and lung – but we’re also applying this to the more rare types, such as cholangiocarcinoma and pancreatic cancer. MSK enables this because it is a specialty hospital and we screen tens of thousands of patients to collect a significant amount of data.”
Another benefit of the PAIGE project is the sheer wealth of information that is being digitized – data on treatment, survival, and disease recurrence in patients. Fuchs believes that this will prove crucial to future research and education. “Because we have all the required licensing for the correlative information, we have been able to start building technologies that go beyond the image domain. You can search for patterns that enable you to predict where mutations might arise, and ask whether some mutations lead to different growth patterns by using this huge bank of data.”
He recognizes that the biggest hurdle in terms of developing the project is implementing PAIGE within the clinic. “This is something at which big corporations usually fail, and I think that the crucial element to achieving clinical relevance is truly understanding pathology and its detail,” he says. The typical pathology department is a network of well-oiled machinery, in which the loss of a few seconds over a certain step can culminate in large disruptions. And this is where the collaboration with MSK strengthens PAIGE’s practical potential. “I believe that the AI revolution will emanate from centers like ours, where you have not only the data and machine learning knowledge and experience, but also an abundance of domain experts who can test and optimize the technology in the clinic.”
To provide this seamless transition into the clinical workflow, PAIGE’s slide viewer software is vendor-agnostic, so it is not tied to the products of a specific manufacturer. The advantage of this approach is that hospitals don’t need to replace long-established systems to use PAIGE, and the technology can be installed in laboratories with a variety of different workflows. Fuchs thinks that this will be instrumental to clinical adoption. “It’s important to work with the pathologists and institutions whose methods are already ingrained, rather than rushing in and trying to replace everything, which is the strategy that big companies often take. This is only possible if you have that larger set of key opinion leaders who form an integral part of the whole initiative.”
The changing face of pathology
Not much has changed in pathology over the last 150 years, so Fuchs can forgive those who question the large-scale changes associated with the switch to using AI in the lab. He believes that there is absolutely no danger of pathologists’ being replaced by the technology. Rather, it will allow them to reduce the amount of time they spend on repetitive tasks and increase the time they can devote to crucial aspects of the job. “Nobody wants to spend their time counting nuclei. When there are hundreds of slides that need meticulous examination, a machine could easily complete the task, leaving pathologists to think about the statistics instead of creating them.”
Another benefit of embracing PAIGE in the laboratory relates to the current shortage of pathologists in many locations. With demand continually increasing as more and more cases require expert diagnosis, the use of computational methods offers an attractive opportunity for workloads to be managed more effectively. “Pathology will look completely different in 10 years. It will be much more diverse and will include algorithms that aid with not only the imaging side, but also the genomics,” says Fuchs.
The widespread adoption that results from these changes will have huge potential benefits for remote areas and those that suffer from a lack of pathologists. “Imagine a small hospital somewhere in the Midwest that has a patient with a strange or rare type of cancer. They won’t be able to send every case to MSK, but imagine if they could use PAIGE – which has been trained by the best pathologists at MSK – to analyze these slides.”
The project has ambitious targets and strives to have international impact. MSK has strong ties with Nigeria, and the people behind PAIGE are in discussions with Indian pathologists and Chinese cancer centers to optimize their machine learning algorithms for the global stage. The concept of uploading an image from a small rural village and getting a meaningful result back quickly is an enticing prospect.
PAIGE’s slide viewer was rolled out institution-wide at MSK in 2017 and is the single entry point for pathologists and cancer researchers there. But how close are we to experiencing PAIGE in clinics around the world? “We are very close with our slide viewer and with our initial disease modules, as evidenced by our recent Breakthrough designation by the FDA. We expect to roll out beta versions of the slide viewer and initial disease modules with partner hospitals and commercial labs later this year, and aim to start selling them in 2020,” said Fuchs.
MSK currently scans up to 30,000 pathology slides each month; however, they are ramping up this input to an ultimate goal of 100,000 slides per month. Fuchs’ present focus is on digitizing the 25 million slides in the MSK archive to create the single largest digital dataset in pathology, but he recognizes that progress will be slow and steady. In the future, when PAIGE works with more cancer centers, the digitization rate of pathology slides will increase.
One thing is certain – Paige.AI has launched an ambitious project that in no way hides its aspirations to change the face of cancer diagnostics for pathologists and patients alike.
Thomas Fuchs is Founder and Chief Scientific Officer at PAIGE.AI, and Associate Faculty Member at Memorial Sloan Kettering Cancer Center.
Subscribe to The Pathologist Newsletters
- TJ Fuchs et al., “Computational pathology analysis of tissue microarrays predicts survival of renal clear cell carcinoma patients”, Med Image Comput Comput Assist Interv, 11, 1–8 (2008). PMID: 18982583.
- TJ Fuchs et al., “Computational pathology: challenges and promises for tissue analysis”, Comput Med Imaging Graph, 35, 515–530 (2011). PMID: 21481567.