The Critical Role of AI in Cancer Diagnosis

Cancer has been famously called “the emperor of all maladies, the king of all terrors”—and very aptly so. It is responsible for tens of millions of deaths every year, is costly to treat, and can be difficult or impossible to cure. Early detection is essential to treating cancer effectively—but all too often, cancer diagnoses come much too late.  Siddartha Mukherjee, who coined the term “emperor of all maladies” in his book of the same name, has called the history of cancer a military one. “Here, too, there are victories and losses, campaigns upon campaigns, … survival and resilience.” In the fight against cancer, we search unendingly for our next great weapon in the oncological arms race.

Artificial intelligence, with its ability to analyse vast quantities and diverse forms of data and unlock powerful new insights, promises to be such a weapon. Already today, it is enabling researchers to detect and treat cancer more rapidly and effectively. 

In this article, I explore the critical role of AI in cancer diagnosis, both as it exists today and in its potential to revolutionise the detection and treatment of cancer in the future. 

What’s ten thousand pictures worth?

One of the ways AI is already being used in the fight against cancer is by detecting anomalies in images of tissue samples. As cancer develops, the uncontrollable growth of abnormal cells results in characteristic abnormalities in the surrounding tissue. Medical imaging enables oncologists to see these abnormalities and detect the cancer. 

Image detection also happens to be an important domain of artificial intelligent research. By training AI algorithms on tens of thousands of medical images, researchers have been able to mimic the diagnostic expertise of oncologists with an extremely high accuracy. Researchers at Tulane University, for example, exposed their algorithm to over 13,000 medical images of colorectal cancer from more than 8,000 patients. Following this training, the algorithm was able to detect colorectal cancer in medical images with an impressive 98% accuracy. That’s more accurate than even many of the most skilled oncologists.  Not only do such algorithms improve diagnostic accuracy, they free up oncologists from the time-consuming and repetitive task of reviewing medical images, enabling them to focus on patients and their treatment. 

In addition to spotting cancer earlier and more accurately, they can also be used to reduce the need for invasive biopsies and tissue samples. Researchers at the Department of Radiology at NYU Langone Health and its Laura and Isaac Perlmutter Cancer Center, for example, were able to use AI to reduce tissue sampling by 27%, while still improving accuracy by an astounding 37%. Of course, there are hurdles to overcome before this kind of technology becomes widely available. Nonetheless, the analysis of huge datasets of medical images by sophisticated algorithms promises earlier, more accurate, less invasive detection and diagnosis. 

Analysing risk factors—and finding new ones

Accurately predicting the emergency of cancer is the holy grail of cancer research. Unfortunately, it’s no easy task. There are dozens, perhaps hundreds of factors that may improve or worsen one’s chances of developing cancer. Speaking broadly, these include genetic factors and lifestyle factors, such as diet, exercise regime, and smoking. 

On a more granular level, we can speak of “prognostic markers” — biological factors associated with greater risks of cancer. Increased tissue density in mammograms, for example, is often associated with an elevated risk of cancer. Likewise, the presence or activation of certain genes is known to be associated with certain cancers. For example, the genes TP53, WT1, Ki67, Topo-II, BRCA1 and BRCA2 are all prognostic markers of ovarian cancer.  Important inroads have been made in analysing these risk factors using artificial intelligence, with very promising results. For example, researchers at the Universities of Iowa and Wisconsin have successfully used artificial neural networks (ANN) to improve prediction of recurrence and divide patients into good (5+ years) and bad prognosis.  Perhaps the most exciting aspect of analysing prognostic markers and risk factors with AI is that machine learning algorithms are able to pick up on novel markers that may have gone hitherto unnoticed.

Indeed, given the opportunity to develop their own predictive models, machine learning algorithms can “stumble upon” markers human researchers weren’t looking for.  In 2021, researchers turned the power of machine learning on mammogram imagery from more than 6,300 women. The algorithm was able to predict the later emergence of interval cancers better than models using traditional clinical risk factors, including tissue density. “The results showed that the extra signal we’re getting with AI provides a better risk estimate for screening-detected cancer,” says Dr. Shepherd, one of the authors. “It helped us accomplish our goal of classifying women into low risk or high risk of screening-detected breast cancer.”

Understanding these new signals will provide oncologists with better tools for assessing risk and making earlier diagnoses. 

Better subclassification, greater personalisation

There are hundreds of types of cancer, and thousands upon thousands of ways it can present and progress. The unique interaction of genes, environment and maladie make continue to complicate cancer diagnosis and treatment. The ability to “zero in” on specific subclassifications of cancer, predict their severity and evolution, and a patient’s response to treatment would be a precious one indeed. 

Fortunately, AI is bringing us closer to achieving this. Several cutting-edge cancer projects, spearheaded by Microsoft, show great promise. 

● InnerEye uses machine learning and natural language processing to help radiologists and oncologists unlock a more detailed understanding of how tumours progress. 

● BC Cancer and Microsoft Canada’s “Single Cell Genomics” will help medical professionals view the genomes of individual cancer cells, allowing for more targeted and specific combinations of treatments and therapies. It will also help oncologists predict how individual cancer cells are likely to respond to chemotherapy. 

● Microsoft’s Bio Model Analyzer is a cloud-base tool that helps model how cells interact, communicate and connect with each other. Its uses include detecting cancer earlier and better understanding how to treat cancer by modelling which therapies will be most effective and predicting resistance to them. 

● Microsoft and AstraZeneca are collaborating on “Project Hanover”, which uses the above tool to sort through fragmented information related to drug interactions and resistance in patients with leukaemia and find the most relevant pieces of data for experts to then use in creating treatment plans.

Tools like this will change the way we understand cancer, right down to individual cancer cells and genes. This kind of granularity may well herald a new age of hyper-specific cancer diagnosis and treatment planning.