Many complicated diseases make up cancer. The American Society of Clinical Oncology’s Cancer.Net provides “individualized guides for more than 120 types of cancer and related hereditary syndromes.” Selecting the right cancer treatment for an individual patient gets even more complex when considering the various diagnostic tests, the wide range of medicines and treatments, and different treatment regimens. With so much information to consider and analyze, artificial intelligence (AI) might be, well, just what the doctor ordered.
Instead of relying on the human brain—or even teams of experts—oncologists can now turn to high-powered computation. Although AI might enhance oncology in many ways, here we will examine using these computational tools to make better decisions about an individual patient’s treatment. This includes the an oncologist’s tools such as digital pathology, molecular testing using a patient’s DNA sequence, and more, to select an appropriate medication.
When asked about the main pathological challenges in oncology that artificial intelligence might address, Manuel Salto-Tellez, chair of molecular pathology at UK-based Queen’s University Belfast, mentioned four: analysis of tissues ahead of next-generation sequencing (NGS), objective scoring of biomarkers associated with therapeutic intervention, more accurate determination of phenotypic features related to prognosis and/or therapeutic intervention, and data processing/AI as a surrogate of molecular status. He called these “four main challenges in the practice of surgical pathology that digital pathology—together with artificial intelligence—has the opportunity of solving.”
Some tools already use AI to improve oncology. Others remain in development, but promising. Nonetheless, some experts think that AI is not ready for oncology applications. Everyone interviewed here, though, believes that AI is changing oncology now or soon will.
Advances Allowing AI
Although AI is not new—emerging as a field of thought in the 1950s—scientists can now do unique things with it. Hans Hofstraat, head of global research for precision digital solutions at Netherlands-based Royal Philips first applied AI in marine environmental research about 30 years ago, but today he sees many features that expand how it can be used. “The big difference now is in ubiquitous digitization, computational power and access to huge datasets,” he said. “Digital pathology datasets, for instance, are quite humongous, and you need high-performance computing to use AI with these apps.”
But big datasets aren’t enough. “You need really good, well-annotated datasets for the analysis,” Hofstraat said. “And now, we make all of this data available and create an interface so that it can be used with AI.” This information can all be combined to create insights and information that, Hofstraat said, “can be made available to healthcare professionals and patients, so decisions can be made.”
Plus, today’s scientists can easily gain access to extremely powerful computing. Even a modern laptop outruns the computing resources available to scientists at the birth of AI. Moreover, today’s tools for crunching big data are far easier to use.
Many of the steps in diagnosing cancer have stayed the same for many decades. A pathologist examines a tissue sample under a microscope to detect cancer or identify variants of cancer. AI can change that. At Philadelphia-based Proscia, Chief Product Officer Nathan Buchbinder envisions AI influencing pathological applications for cancer in two major ways. “First, through pattern recognition, AI can automate many of the routine tasks performed today, driving tremendous improvements in quality and efficiency.” Second, Buchbinder sees AI augmenting the role of the pathologist. “For the past 150 years, pathologists have diagnosed cancer by interpreting patterns in tissue using a microscope,” he said. “With AI, we can augment these diagnoses with additional insight into treatment options or outcomes.”
To push pathology in that direction, Proscia created its Concentriq software, which can be used through a web browser or installed on-site. It uses AI—deep learning based on convolutional neural networks and some in-house methods—to analyze a digitized slide. “This works with enormous whole-slide images of tissues that can be shared for analysis and collaboration,” Buchbinder explained. “We are creating a series of disease-specific AI modules.” For example, Proscia’s DermAI screens and classifies skin-cancer biopsies before or after a pathologist’s review to help reduce errors and improve laboratory confidence and efficiency. “There are more than 300 possible diagnoses for skin cancer,” Buchbinder said. “It’s not clear cut, and there’s a range of what a cancer might look like for a specific diagnosis, so having this extra layer of AI review is helpful.”
Other companies also aim at improving digital pathology with AI. According to Hofstraat, Royal Philips applies AI to oncology “on many levels.” When asked what areas of oncology healthcare benefit the most from AI, Hofstraat said, “AI plays a big role in precise treatment selection and execution.”
For example, Hofstraat and his colleagues are combining a large number of pathology-related features into a single visual diagnostic. “We show the information to a radiologist or pathologist and enable them to grasp all of the information there in one easier-to-understand heat map,” he said. “For instance, by analyzing an entire tissue slide—how many tumors cells, where immune cells are, and so on—it provides a more comprehensive and objective view than an individual pathologist can grasp within reasonable time.”
As Hofstraat noted, “To apply AI, you need high-quality data,” and that depends on the method of collecting it. “For digital pathology, we developed a scanner that is now FDA-approved,” Hofstraat said, “and that is the vehicle that started our path to AI-based applications.”
Various paths can be taken in applying AI to pathology. As another example, Dave Billiter, CEO and co-founder of Deep Lens in Columbus, OH, said that his company’s AI-driven pathology platform, VIPER “supports improved efficiency in clinical and research processes, and connects qualified patients to clinical trials at the time of diagnosis.” He adds, “This accelerates trial enrollment and expands access to therapeutic alternatives.” The VIPER platform leverages advanced machine learning techniques, such as next-generation convolutional neural networks, which Billiter said, “helps rapidly classify and identify difficult tumor sub-types and stages in real-time with very high accuracy.” Plus, the company has been validating this system for over a decade—working with pathologists at more than 85 major institutions in nine countries.
Expanding Data Collection
To make those decisions based on as much information as possible, even more resources must be analyzed. Then, all of that information must be combined in an actionable way. There’s more than one way to go about that.
At Swiss company SOPHiA GENETICS, scientists take a decentralized approach to analyze genomic and radiomic data of healthcare institutions from around the world. The company’s general manager in North America, Kevin Puylaert, says it analyzes about 15,000 genomic profiles each month for a total to date of 400,000. “Experts from almost 1,000 healthcare institutions use SOPHiA daily,” he said. “SOPHiA detects and characterizes genomic alterations, provides experts with an accurate report, and then the expert will go through each case and do their own interpretations.” Puylaert added, “AI in our system learns how those cases are being classified, learning how the experts are working with that data.”
The vision is that this platform’s AI will learn how to pre-classify future cases. Plus, everyone using the system benefits from the cases analyzed by others, because each case teaches the AI to better assess future cases.
The scientists at SOPHiA GENETICS are building their system to address many kinds of cancer from taking into account a variety of aspects. “The most crucial point of the collaborative approach of many experts is that it’s not a single expert opinion,” Puylaert said. “It’s not one way of looking at evidence, where everyone thinks the same thing, and this lets the field evolve.”
Other experts are also looking for ways to learn more by turning AI loose on information from more than one patient. At New York-based, Prognos, the company collects clinical laboratory test results from patients with a range of conditions. Currently, says Chief Data Scientist Adam Petranovich, the company’s data lake includes results from more than 25 billion tests. “Some are oncology-related tests and some are not,” Petranovich said. “We try to look at them all holistically.”
As an example, Petranovich created this scenario: Someone had a biopsy, a pathologist sees cancerous cells in it and other tests confirm the cancer, but the person has surely had other tests and no one really knows what pieces of information might help in predicting the future of that cancer or what treatment might work the very best. “We throw all the data and test results into our AI algorithm, and see what we can learn,” he said. “From this, we hope for increased accuracy in cancer predictions.”
In the future, such AI-driven analysis from Prognos will even include social factors, like where a person lives, information about the environment, the person’s job and more. “All of those things have the potential to inform an algorithm, just as much as a test result or image,” Petranovich stated.
These ideas make sense in general, but can Prognos prove that its algorithm really works? “We use many methods for model validation,” said Petranovich. “First is back testing.” That is, Petranovich and his colleagues have a data set that goes back 10 years, and they use these patients as a training set for the AI. The scientists select a random group—say, 10% of the people—who are left out of the training and see how the algorithm does at describing where their health is heading. Petranovich explained that they are working on this now, and they should know in a couple months how their system performs with its oncology models.
If the results are promising, great; if not, Petranovich and his colleagues will adjust the model and try again. “We can fail fast,” he said. “Computing power is relatively cheap, and we can essentially spin up thousands of computers, run an algorithm overnight and make big predictions,” he says. “We can rapidly prototype and try new things.”
An Actionable Algorithm
One of the latest buzz phrases, Big Data, gets applied to seemingly everything. “There’s this notion in business that big data is a panacea, but it’s in fact a problem in biology, especially in clinical medicine,” said Mark Kiel, chief science officer at Genomenon, Ann Arbor, MI. “There is so much data available for each patient, and connecting information in the medical literature to the patient is required to make the most accurate diagnosis and find the right treatment.”
Someone could create a system that makes all of that information accessible to a pathologist or oncologist, but what then? “It’s not about access, but making the information actionable,” said Kiel. “It’s about correlating published information with patient information.”
So, Genomenon developed the Mastermind Genomic Search Engine, which Kiel described as a database that indexes and understands complex genomic information from the medical literature. For example, the software uses Genomenon’s patented Genomic Language Processing (GLP)—facets of machine learning and AI developed by scientists at Genomenon—that “understands what authors of articles intended in describing research studies, and allows our users to connect this knowledge to their patient data,” Kiel explained. Mastermind can answer questions such as “how a drug can be utilized for a patient with a particular constellation of mutations determined through next-generation sequencing,” Kiel said. To keep up with the constant flow of NGS research, Mastermind is updated weekly with newly published literature.
In thinking about how oncologists can use Mastermind, Kiel said, “Clinicians need the evidence, but they don’t have time to search through Google results and a pile of scientific articles to find that evidence.” Mastermind organizes and annotates the information, and provides oncologists with the ability to find the right evidence to quickly make diagnostic and therapeutic decisions for their patients.
Intelligence is Not All Artificial
From N-of-One, which is owned by German parent company QIAGEN, Chief Scientific Officer Sheryl Elkin said the company provides clinical interpretation and decision support for molecular testing. She added, “We help oncologists and their patients interpret their data and the results of tumor testing, then link them to therapies or clinical trials with the supporting evidence.”
To do that, N-of-One does not use AI. Instead, Elkin and her colleagues—a group of Ph.D.-level scientists supported by a team of medical and research oncologists—rely on a large knowledgebase built from the analysis of more than 150,000 patient cases. “We have decided against machine learning, so far, because of black-box results and lack of reproducibility,” Elkin said. “There’s a lot of human judgment involved in our analysis, and the team undergoes rigorous training.”
But N-of-One is far from technology-free. “A sophisticated technology platform automates the workflow as much as possible,” Elkin explained, “as well as streamlining the scientists’ process by identifying literature and pulling in data from various sources, and then presenting it all to scientists. There is multivariant analysis inside the N-of-One system and lots of technology automation, but not AI per se.”
N-of-One and QIAGEN Clinical Insights (QCI) are in the process of integrating their solutions. Here, Elkin sees value in AI and expects more in the future. So far, she said, “a large and robust training cohort that is reliable for driving a machine learning–based clinical decision-support system isn’t there yet.” That said, many different AI-algorithms are already in use within QCI—for example, to identify literature for curation, network computations, knowledgebases and ontologies, and knowledge-driven inference algorithms. QCI also uses a large team of M.D. and Ph.D. scientists to curate the literature and then uses computational methods to classify variants. By merging these platforms, Elkin said, “we will have a combined solution that pairs a computational approach with a human analytical approach, allowing clients to submit files with large numbers of variants and ending up with a small set that is interpreted and summarized for a physician.”
More Tech to Come
The key is all about getting better information to pathologists and oncologists to improve the path for cancer patients. Scientists hope to continually enhance that process, and that takes increasingly advanced technology.
“When AI goes beyond what a human can do, the oncologist can correlate a patient’s biopsy to the likelihood of a response to therapy and more personalized approaches to medicine,” Buchbinder predicted. “We don’t have that yet, but there’s evidence of that being a really key area.”
Getting there requires the adoption of sophisticated technology in oncology. Some experts think there’s considerable room to improve that. “The healthcare industry is fairly behind compared to other industries in the types of technology and algorithms used,” said Petranovich. “Oncology technology, especially, is not up to what’s used in other industries.”
As shown here, though, many good, potentially beneficial pathways and applications are under development. Some of these show great promise in digging deeper into oncology data and pulling out predictions that cannot be gained by one mind alone. Instead, AI techniques explore evidence from around the world and at unprecedented depth. Now, we must wait and see if these tools can turn cancer into a manageable or, better still, curable disease.