With Training, Startups Like Paige.AI Could Soon Diagnose Cancer

Xconomy New York — 

It turns out interpreting biopsies isn’t entirely unlike mapping the rocky terrain of Mars. Thomas Fuchs, a former research technologist at NASA, is taking some of what he learned in training algorithms to navigate the Red Planet for the Mars Rover project and applying it to create cancer-detecting algorithms.

Fuchs (pictured) thinks his startup, Paige.AI, can leverage machine learning in pathology to increase accuracy, save pathologists’ time, and deliver better patient outcomes. Pathology, which involves reading slides from biopsies to evaluate tissue health, is a crucial step in the treatment of cancer patients. It influences everything to follow—from the treatments or therapy ordered, to the additional tests doctors decide to run.

New York-based Paige.AI is first focusing on training algorithms for prostate and breast cancer diagnosis. The company announced a $25 million Series A funding round led by Breyer Capital in February.

Using digitized slides, Paige.AI employs deep learning techniques to teach algorithms how to differentiate between slides of healthy tissue and those with abnormalities. By using algorithms to pull out only the slides most likely to be cancerous, Fuchs said, pathologists will be able to spend more of their time actually reasoning and analyzing the results, rather than sorting slides. This has the potential to cut the amount of time pathologists spend looking at slides in half, according to Fuchs.

To train machine learning algorithms to identify what type of cancer or cell abnormality is present, having access to quantity as well as quality of slides is crucial.

Fuchs, Paige.AI’s founder and chief scientific officer, is also the director of computational pathology in the Warren Alpert Center for Digital and Computational Pathology at Memorial Sloan Kettering Cancer Center. With approximately 30 percent of cases at the cancer center coming from outside, Fuchs said, the institution is already a popular resource for second opinions in pathology.

A recently inked agreement between Paige.AI and New York-based Memorial Sloan Kettering will give the startup exclusive access to the cancer center’s intellectual property in computational pathology, as well as its library of 25 million pathology slides for the next eight years. Fuchs said access to both Memorial Sloan Kettering’s library of slides as well as its world-class pathologists gives Paige.AI a leg up on the competition.

Fuchs ultimately hopes that by digitizing slides and perfecting his company’s algorithms, pathologists will be able to use the pattern recognition capabilities of Paige.AI to conduct something of an image search. Cross-referencing cancerous slides from a fresh biopsy with Paige.AI’s database of slides, the algorithms will make use of past cases where a patient presented with similar cell morphology.

Because raw images from the Memorial Sloan Kettering library include accompanying annotations and even genome sequencing tests, the image search could show the accompanying diagnosis, treatment, and even outcomes of patients who presented with similar cell morphology in the past—that’s the idea, at least.

Fuchs sees this potential to correlate images with sequencing data as Paige.AI’s real value for hospitals.

The genome sequencing data accompanying an image that matches a slide taken from a fresh biopsy could be used to predict how the tissue might mutate, meaning a hospital could potentially only need to do one expensive test in order to infer the results of the second.

But training the company’s deep neural networks takes a combination of several approaches. One of the methods involves fellows in pathology annotating hundreds of slides of a cancer—say, breast cancer—and passing those to senior pathologists for review before feeding the annotated slides to Paige.AI. Another approach incorporates the startup’s software into the workflow of practicing pathologists. At Memorial Sloan Kettering, Paige.AI’s accompanying digital slide viewer—something of a Google Maps for viewing cells—is already part of the clinic workflow.

“We have Microsoft Surface stations where they can, with a pen, just paint or correct on [the slides] so the AI can learn continuously from the pathologist,” Fuchs said.

A final approach to training the neural networks involves the incorporation of patient data. By training the neural networks on clinical data, from patient treatment to patient outcome, Fuchs said, Paige.AI will be able to help calculate survival rates.

Anant Madabhushi, a biomedical engineering professor at Case Western Reserve University, said he sees the ability to view patient outcomes as the real value of Paige.AI.

“I think right now digital pathology is hard, deep learning is hard. Everybody thinks that if they get pathology images they can start doing deep learning,” Madabhushi said. “I would argue that the most valuable data is the clinical trial data sets, with outcome information.”

At Case Western Reserve, Madabhushi is part of a team using deep learning to analyze biopsy tissue to determine whether or not chemotherapy would be beneficial for a cancer patient. By accelerating diagnosis and improving accuracy, the test has the potential to reduce misdiagnosis and save patients from unnecessary and grueling chemotherapy treatments, all at a lower cost than tests available today.

For artificial intelligence in pathology, Madabhushi thinks finding that value proposition is crucial for success.

Paige.AI is by no means the only company applying artificial intelligence to make pathology less qualitative and more quantitative.

Boston-area startup PathAI, one of Paige.AI’s most obvious competitors, raised $11 million in November in a round led by General Catalyst Partners. Initially, the startup is partnering with Philips to create a tool to help pathologists diagnose metastatic breast cancer. PathAI’s product has yet to be incorporated into the workflow of pathologists the way Paige.AI’s has at Memorial Sloan Kettering. But Andrew Beck, PathAI’s co-founder and CEO, has published extensively in the fields of cancer biology, cancer pathology, and biomedical informatics, lending credibility to the startup’s expertise in the domain. Google and IBM have their own initiatives in pathology, too.

While people and big organizations are clearly investing in artificial intelligence applied to medicine, Madabhushi is skeptical of the economic viability of many of these endeavors.

“A lot of groups have not really thought through, at least in my mind, what is the value proposition… how are you going to make your money?” Madabhushi said.

In the case of pathology, digitizing slides is time consuming and expensive. And what does AI need in order to be accurate? Not only well-annotated data, but a lot of it.

At the James Cancer Hospital at Ohio State University, biotechnology company Inspirata announced on April 18 that it has just digitized its 500,000th glass histopathology slide, up from 100,000 slides last August. Madabhushi, who also serves as a member of Inspirata’s scientific advisory board, said that’s the rate of progress, even with multiple slide scanners at full operation.

Fuchs said his team is digitizing 40,000 slides a month. While Memorial Sloan Kettering has been digitizing slides for the last five years, at the current rate, it would take some 52 years to digitize the center’s 25 million slides. In a press release, Paige.AI acknowledges it is focusing on digitizing “millions of additional archived slides in the next few years,” but not all 25 million. With the new funding in February, Fuchs said, ramping up Paige.AI’s rate of digitization is a primary objective.

Eventually, Paige.AI’s technology could process digital slides submitted from hospitals with limited access to pathologists. That resource could be life-saving for individuals with rare tumors that might not be immediately identifiable to many pathologists.

While Paige.AI is primarily working with Memorial Sloan Kettering, the startup anticipates eventually working with other cancer centers. Madabhushi said collaboration will be important if Paige.AI is to train algorithms with universal application to the way slides are prepared across different hospitals.

Paige.AI is currently a team of five, not including Memorial Sloan Kettering staff collaborating with the startup. Fuchs said a portion of Paige.AI’s recent funding will go toward expanding and hiring machine-learning engineers.

Fuchs is optimistic about attracting talent.

“What speaks for us is it’s a good cause,” Fuchs said. “You can spend your time stratifying teenagers on Facebook, or you can heal cancer patients. I mean, it makes a difference what you spend your life on.”

A 2017 graduate of the University of Wisconsin–Madison, Cadence Bambenek is a freelance science and technology writer and fact-checker. She previously interned at the Wisconsin State Journal, Business Insider, and Psychology Today. Follow @cadencebambenek

Trending on Xconomy