Breast cancer is the most commonly diagnosed cancer among all people worldwide

Unlocking AI’s full potential for beating cancer

Image credit: Science Photo Library

AI has been hailed as a silver bullet to beat breast cancer for decades. What’s stopping it?

For decades, experts have been exploring how artificial intelligence can help detect breast cancer. Yet cases continue to rise. Deaths continue to devastate, and all of AI’s promise is yet to materialise. We look at what is hindering progress in this space, and what needs to be done for AI’s true, life-saving potential to be realised.

Four women are diagnosed with breast cancer every minute across the globe, a third of whom die from the disease. As of 2020 – the last year for which we have full and verified figures – breast cancer is the most prevalent form of the disease, affecting one in eight women during their lifetime.

This ongoing prevalence has seen breast cancer become one of the most well researched, well funded and well understood cancers of our time. It’s also why, of all the diseases, conditions and applications where AI and machine learning promise real impact, breast cancer is the one with the most hype.

Yet, almost 50 years since researchers first began exploring the relationship between image processing, artificial intelligence and breast cancer diagnosis, relatively poor progress has been made. Certainly not in a way that has shifted the dial, and not in the revolutionary ways that the breakthroughs and respective headlines over the years would have us believe.

The reason? Little has changed in how we approach the technology, how we measure its success, and how willing healthcare systems are to use it. We’ve reached an impasse – one that’s potentially costing millions of lives, but it’s one that the next generation of AI pioneers are hoping to break.

To highlight how stunted progress has become, you only need look at the way the technology and the views surrounding it have remained static. In one of the earliest papers on the use of image processing for breast cancer diagnosis, from 1976, researchers praised the potential of the early form of AI, while highlighting several key challenges. The potential discussed in that paper – to use technology to make mammography screening more accurate, more efficient, and more impactful – is the same potential being touted today. The challenges listed – those surrounding how best to apply the technology, how it performs compared with humans and the reluctance of clinicians to adopt AI in medical settings – are the same challenges being faced today.

This is, in part, because the sector has been historically slow-moving. Mammography technologies that are today considered the gold standard of breast cancer screening took more than 50 years – between the 1910s and the 1960s – to reach widespread acceptance and adoption. Once mammography technologies became the norm, computer-assisted detection systems of the 1980s rose in popularity. These systems run rule-based algorithms and were used to offer second opinions on cases, or to assist radiologists where there was a lack of good-​quality evidence. They helped improve detection rates, but they were far from ubiquitous, and in the decades that followed, progress largely stalled once more.

More recently, however, the rise of modern AI and deep machine learning, coupled with an influx of capital into healthcare and the advent of Covid, has reinvigorated things. It’s difficult to determine which providers are trialling which AI systems, due to the closed nature of many of the trials. It is estimated, however, that the trends surrounding healthcare have accelerated digital adoption by as much as seven years.

Plus, while little has changed in terms of getting the latest imaging and diagnostic technology into the mainstream, very much has changed in other areas. Namely breast cancer survival rates. Even though the number of cases is rising – caused by a general growth in population, and a growth in the number of women reaching high-risk ages – the number of breast cancer deaths hasn’t risen at the same rate. This is thanks to the advances in screening that help radiologists catch cancers early.

BCAM cancer chart - Inline

Image credit: World Health Organisation

Across all age groups, the World Health Organisation found that when women attend screening programmes, breast cancer mortality is reduced by 20 per cent compared to groups of women who don’t attend. In high-​income countries where screening is more widely accessible, such as the US, Australia and the UK, the five-year survival of breast cancer is as high as 90 per cent. In low- and medium-income regions, where effective screening is not readily available, five-year survival rates plummet. In India, only 66 per cent of patients fight off the disease. In South Africa, this drops as low as 40 per cent.

Improving screening practices to catch cancers sooner is where early researchers saw the greatest benefit of software then, and it’s where AI and machine learning still offers the most potential now.

The underlying processes and inherent use of data in modern cancer screening programmes makes them a near-perfect match for the application of AI and machine learning. AI can be trained to read millions of images. It can identify patterns, find correlations, and spot anomalies in ways that aren’t possible for radiologists to do at scale. AI can take on repetitive admin tasks, while automating and speeding up such processes to lighten the load on clinicians.

Jonas Muff, CEO of deeptech firm Vara, adds that breast cancer screening “presents almost the ideal use case for AI”. Vara has built a platform that is used internationally by radiologists to help improve screening quality using AI and automation. “Not only can AI increase the efficiency and accuracy of mammography screening,” Muff continues, “but it is key to helping shift the dial from diagnostic to preventative care. It can help drive down costs, improve access to adequate care and it can help plug gaps where resources of expert knowledge are lacking.”

To put these lack of resources into perspective, in the US and Europe, there are around 10-12 radiologists per 100,000 habitants. In Mexico, there are just three experts for every 100,000 people.

When viewed in this way, AI promises a lot. Fewer missed cancers, fewer false positives, fewer lives potentially lost. There seems little reason not to adopt it.

However, the biggest sticking point in all the talk about AI being breast cancer’s silver bullet is that the focus is too narrow. Until now, the narrative has centred on how AI can either mitigate a shortage of experts or replace radiologists completely. This doesn’t do AI justice, according to Muff.

“Instead of using AI to find and diagnose cancer alone, it offers increased value when it’s more closely integrated into the entire clinical routine from start to finish, and when it’s complemented by automation,” Muff says. “AI’s power in screening programmes is also increased when it continuously monitors both the data being produced, as well as its own performance, against all the prospective data gathered.”

In this way, radiologists and the AI get real-time feedback on how they’re doing. Under current screening processes, this feedback can take months or even years to be produced and shared.

The narrow focus also doesn’t account for the fact that AI, on its own, has been repeatedly found to fall short of expectations. In a scoping review in September 2021, researchers from the University of Warwick Medical School concluded that AI, on its own, is simply not ready. At best, the team found existing AI models have “poor” methodological quality. At worst, three of the models failed to catch cancers that had been successfully identified by human radiologists, one of them missing as many as 10 per cent of identified cases.

The narrow-minded approach additionally doesn’t factor in cultural barriers. It doesn’t account for the vast economic and infrastructure gaps between countries. It doesn’t appreciate the regulatory and liability hurdles that lie ahead.

“It is a difficult, slow and expensive process, that needs a lot of investment and needs to be validated and regulated before it reaches the market,” adds Luis Enrique Hernández Gómez, CEO of Thermy, oncologist surgeon and Royal Academy of Engineering LIF Advance participant. According to Gómez, the process of going through scientific and clinical validation, establishing a viable product, all while adhering to the standards for industrialisation and regulation, can add at least two years to an already lengthy process.

Anything that provides diagnostics using AI, particularly using image processing, is problematic from an ownership point of view, too. “Firstly, there’s the issue of getting patentability,” says Peter Finnie, partner and patent attorney at Potter Clarkson. “Digital innovation is not fully recognised by the UK patent office, while the US has issues with granting patents for diagnostic methods. Then the regulatory framework isn’t there to allow dynamic AI solutions to be used. It’s never as simple as: ‘Here’s the software, switch it on and start making clinical decisions.’ That isn’t the way the medical profession or the regulators work.”

AI firms in this space are then met with liability considerations – if a machine is making the decision, who is at fault should the wrong decision be made? – as well as societal pushback.

“Patients want to be treated by people. Doctors are aware that a software fail in a clinical setting can lead to a much bigger problem than it might in another industry,” explains Harry Briggs, health tech investor and managing partner at OMERS Ventures. Not to mention the privacy risks around the collection of data and its analysis. “These risks make administrators and purchasers reluctant to invest in software until it is fully verified, approved and also being demonstrated in wide-scale use.”

According to AlbionVC partner and former surgeon Andrew Elder, who has been investing in health tech for two decades, hesitancy ultimately comes down to a “better the devil you know” mentality.

“When introducing any form of AI into clinical settings, the product is at an immediate disadvantage because it’s not being compared like-for-like. Existing breast cancer diagnostic technologies and systems aren’t perfect. They work, but they’re far from faultless. By contrast, any new product coming in to disrupt this status quo has to prove it’s near perfect. It must prove it’s free from bias, that it works with complete transparency and that it’s accurate. It must fit within medicine’s ‘do no harm’ ethos. All the while, the harm already being done by the status quo, and the harm that could be done from not implementing the technology, is failing to be factored into the decision-making process.”

In last year’s Warwick study, the researchers suggested that to navigate these many hurdles, AI should be viewed as complementary to radiologists, rather than contrasting. When used alongside current processes, liability concerns can be lowered, regulatory issues can be reduced, fears can be allayed, and progress can continue at pace.

This is the route Vara is taking with its decision referral pathway. This pathway is a screening process in which an algorithm only makes a decision on a case when it is confident about its prediction. Any cases for which it has ‘doubts’, it leaves to human radiologists to read, without any AI support or interference.

Vara is currently used in 30 per cent of clinical settings in Germany and has been trained and evaluated on more than seven million images taken from the German screening programme. “From our work with clinicians across Germany,” says Muff, “AI has proved to be most effective when it is optimised to pre-screen ‘normal’ mammograms and automatically prepare the relevant reports. It’s these scans, as well as the pages of admin that sit alongside them, that take up the bulk of the radiologists’ time,” he explains.  

Start-up Thermy takes a similar, complementary approach to Vara but instead combines AI with thermographic images. These images are taken on its own range of thermal scanners before being classified using AI. Breast cancer can be detected via thermographic imaging because when cancer cells grow and multiply, the blood flow and metabolism rise, which raises the temperature of the skin.
Instead of looking to replace mastography, Thermy believes its technology can be a complementary tool to help catch breast cancers early, especially in women under 40.

Moving away from diagnostics, a technology developed out of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) is using AI to personalise screening pathways. Called Tempo, the technology uses an AI-based risk model to determine how often a patient should be screened.

Depending on a country’s screening practices, older women who are typically at greater risk of breast cancer are encouraged to be screened every one to two years. This blanket approach can cause interval cancers to be missed, or not caught early enough.

Tempo’s personalised approach can increase the chances of catching cancers early among certain patients, while reducing the frequency of mammograms among others. In trials, Tempo has improved early detection rates while reducing the number of mammograms by 25 per cent, thus saving time and sparing resources.

One significant benefit to taking holistic, rather than narrow, approaches is that such technologies can be designed to work as well in low-income as high-income countries.

In addition to facing the same liability, regulatory and perception hurdles as high-income countries, lower-income regions have to battle against economic, technological and infrastructure hurdles. Screening is expensive to launch and expensive to run. It requires specific skills and infrastructure. It is often held back by other more pressing medical concerns and priorities in healthcare systems that are notoriously fragmented.

Effective, holistic and complementary AI systems can help reduce such costs. They can support the smaller number of trained radiologists and help health systems better manage their scarce resources.

Vara is one of a small group of platforms that are already in use in clinical settings. Earlier this year, the company launched its first screening units in lower-income countries – Mexico initially and then Greece – in partnership with healthcare providers in those regions.
“Current AI solutions promise great cost and efficiency reductions but largely rely on systems to already be in place,” Muff adds. “These solutions are missing an opportunity to improve access; to get the cost-, time- and life-saving technologies to all women. Not just those living in the richest countries. At Vara we believe increasing access to effective breast cancer screening should be the starting point, not an afterthought.”

Thermy’s Gómez agrees: “Cancer is one of humanity’s unsolved problems. This shortage of radiologists, and the limited availability of imaging equipment, has led to the need for tools that make the radiologist’s job more efficient, accurate and quick, and that allows them to tend more patients. It’s why AI is so attractive for tackling this problem.”

While the idea of AI sweeping in to save the day when it comes to breast cancer screening sounds hugely exciting, the reality is more boring – and rightly so. AI’s real promise lies in picking up the slack for radiologists in the routine heavy lifting and the everyday admin. This is the most realistic, most feasible and most likely way in which the technology will ever make it beyond the lab and into clinics and finally get ahead of the bureaucracy. In the name of saving lives.

Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.

Recent articles