Waiting for the results of a biopsy can be an harrowing experience. The anxiety it creates for the patient, as they wait for the results on a suspect growth or damaged tissue, can even lead to sickness, through a phenomenon known as biopsy stress.
The challenges, however, of obtaining a fast diagnosis and turnaround are immense, not in the least because a sample can require the expert vision of a number of highly trained doctors to successfully diagnose. The more doctors reviewing a case, the more accurate the diagnosis, but the practicality of getting five or ten – or, better yet, a hundred – pathologists to review the samples of every single patient at a hospital, and form a single consensus-based diagnosis, are of course nil.
Artificial intelligence may provide an answer to the problem, or at least provide tools to find one, says Hamid Tizhoosh, director of KIMIA Labs at the University of Waterloo’s Artificial Intelligence Institute.
Tizhoosh envisions a novel approach that is “pathologist-centric,” and has partnered up with several organizations, including St. Jacobs-based business Huron Digital Pathology as well as the Grand River Hospital, to bring the idea to life. The project has also received the backing of the province through a $3.1-million grant from the Ontario Research Fund.
“If you look at the mainstream AI right now, the absolute majority of people – companies and research units and universities – are actually investing in AI making decisions,” explains Tizhoosh, who is the lead researcher on the project.
Organizations are investing millions into developing AI that can look at tissue samples for a specific illness, and then spit out a binary answer: ‘yes, this sample is showing signs of malignant cancer’, for example, or ‘no, it is not.’
“And as long as you can come with several thousand if not millions of images [of biopsies], then you can train an AI to learn the difference. Which is fantastic, but the only problem is a piece of smart software telling the pathologist ‘yes’ and ‘no’ – ‘yes cancer’, ‘no cancer’ – it doesn’t really help,” says Tizhoosh.
This is because when a pathologist puts a sample under a microscope or, as is the increasingly the case nowadays, pulls it up on a screen, they are not simply looking for a ‘yes’ or ‘no’ answer. Rather, they have to bring their extensive training to the fore and come up with a detailed explanation why they believe their diagnosis to be the correct one. They have to parse over the cellular structure of the sample for clues, little irregularities and abnormalities, and then present their arguments in a detailed report that other doctors and experts can evaluate.
“In that sense, the majority of works that are being done world-wide, and several companies are getting started on doing this, are training artificial intelligence software to tell us ‘yes’ or ‘no’, is actually of not much use,” says Tizhoosh.
Instead, what Tizhoosh and co. are seeking to do is create an image search program that can take a look at a sample, and then search through a catalogue of millions of other samples produced over the years, to come up similar cases. In seconds, the program would be able to procure previous samples that were extremely similar, with detailed reports on how those cases were diagnosed, as well as how they were treated and the eventual outcome for the patient.
The image search would essentially allow the pathologist to virtually “collaborate” with his or her colleagues instantly, by showing how they had diagnosed similar-looking samples. Moreover, most hospitals will have an extensive number of biopsy samples to draw on, he adds.
“If you just look inside one hospital, you already have a huge amount of information from past patients, the past 10-20 years, which have been evidently diagnosed. And we know what was the outcome… So this is a huge amount of medical wisdom that is already there,” says Tizhoosh.
The advantage over an image search, as opposed to the binary ‘yes-no’ system, is that the AI keeps the trained pathologist at the centre of the whole process.
“We don’t want to replace them as the mainstream AI claims can do, which we find ridiculous and preposterous. We want to assist the pathologist … by giving him access to the collective wisdom of his colleagues in the same hospital.”
The technology is still in an early phase of development and needs extensive testing and verification, which is where Tizhoosh’s healthcare partners come in. Three hospitals are working with the group: Grand River Hospital (GRH), Southlake Regional Health Centre in Newmarket, and University of Pittsburgh Medical Center in the U.S.
“The role of the hospital is kind of a two-step role. One part of the role is to provide that clinical expertise that will be needed to validate what the University of Waterloo is trying to validate,” said Carla Girlametto, manager of research and clinical trials at the GRH. The second part is to provide Tizhoosh and his team with the digital biopsies they need to test their program.
On the hardware and commercial end of the project is Huron Digital Pathology in St. Jacobs, which develops the medical scanners that allow pathologists to digitize biopsy samples to extremely fine resolutions.
“We’re hoping we can develop some very specific solutions within a year, year-and-a-half timeframe,” explained Patrick Myles, CEO of Huron Digital Pathology, of the short-term prospects for the project. The software would be able to work with data from any scanner, but potentially could be integrated into Huron’s own hardware.
“Our goal is to move into something that can benefit Grand River initially and the Waterloo Region, but also to be able to commercialize it broadly so that it can help other hospitals in Ontario and Canada and hopefully worldwide,” added Myles.