Lung cancer is the deadliest cancer in Canada—and the world. Every year, lung cancer kills more than 20,000 Canadians. Of those who receive a diagnosis, about 75% are expected to die within five years. Yet, there’s no national standard in Canada for lung cancer screening and, of all the cancers, lung cancer receives some of the lowest funding for research.
Late-stage lung cancer is much harder to treat, so early detection is critical for boosting survival rates. Yet, half of lung cancer diagnoses are made at Stage 4, and about 20% are made at Stage 3, according to the Canadian Cancer Registry database at Statistics Canada. The five-year survival rate for someone with Stage 4 lung cancer is only 1%. On the other hand, someone with Stage 1A lung cancer has a five-year survival rate of 49%.
Aside from low funding and a stigma attached to lung cancer—because it’s often, but not exclusively, related to smoking—early symptoms such as fatigue and a persistent cough often go unnoticed. And non-smokers may assume they’re not at risk, despite an increasing number of non-smokers being diagnosed with lung cancer. Indeed, the Centers for Disease Control and Prevention (CDC) reports that 10% to 20% of lung cancers happen in people who’ve never smoked or smoked only a few cigarettes in their lifetime.
That’s why lung cancer screening is so critical—and Canada has been breaking new ground in this area.
Current approaches to lung cancer screening
The Pan Can Lung Cancer Risk Prediction Model, developed in Canada, helps to determine whether a person should get a computed tomography (CT) scan to detect early-stage lung cancer. Current approaches to screening are based on age and smoking history, but the Pan Can model takes this a step further, looking at additional variables such as sex, family history, educational level, body mass index and chronic obstructive pulmonary disease.
Research suggests that lung cancer screening can improve outcomes, if caught at an early stage. Results of the NELSON trial—a large, randomized, population-based study in Belgium and the Netherlands—demonstrated the value of low-dose CT screening in people at high risk for developing lung cancer. Overall, the study found that CT scanning decreased mortality by up to 26% in high-risk men and 61% in high-risk women over a 10-year period.
In the U.K., Wythenshawe Hospital—part of Manchester University NHS Foundation Trust—has launched Lung Health Checks to boost lung cancer survival rates, providing quick, accessible screening via low-dose CT scanners in mobile units. As part of a pilot programme in 2016 between the Wythenshawe Hospital, Macmillan Cancer Improvement Partnership and the Manchester Clinical Commissioning, they were able to quadruple diagnosis rates for early-stage lung cancer. And eight of 10 cancers detected were at an early enough stage to allow for curative treatment in 90% of patients.
On a larger scale, the European Commission has adopted a Beating Cancer Plan aimed at improving cancer outcomes in the European Union through prevention, early detection, access to treatment and improving quality of life. In 2020, 2.7 million people in the EU were diagnosed with cancer—and cases are set to increase by 24% by 2035. While there’s much focus on breast cancer and cervical cancer, lung cancer remains the deadliest cancer in Europe.
However, each member nation manages its own affairs, so it’s challenging to harmonize a program across Europe. In some countries, smoking is part of the culture, so lung cancer has a lower level of concern among citizens—and not all governments have the money to support such a program. So the EC faces significant obstacles in rolling out its plan consistently across Europe.
The U.S., on the other hand, has a national program specifically for lung cancer screening, and the U.S. Preventive Services Task Force (USPSTF) recommends annual lung cancer screening using low-dose CT scans for high-risk individuals. Most insurance plans, as well as Medicare, help to pay for these tests in the U.S.
In comparison, there is no national program in Canada for lung cancer screening—even though the Canadian Task Force on Preventive Health Care is recommending lung cancer screening with three annual low-dose CT scans for high-risk individuals. Canadians considered high risk are those aged 55-74 who smoke, quit less than 15 years ago or have a history of smoking.
Challenges with the Lung-RADS screening model
Clinicians base lung cancer screening on a set of criteria: nodule size, nodule density or the appearance of a new nodule. But this isn’t a foolproof method.
According to the Lung-RADS screening model—the current best practice—there are four main categories for lung cancer screening. Categories 1 (negative) and 2 (probably benign) require the patient to come back in 12 months. Category 3 (suspicious) requires the patient to come back after six months. And Category 4 (very suspicious) requires the patient to come back in one to three months for a follow-up CT and tissue sampling.
The criteria for lung cancer screening was developed with consensus and careful thinking, but there’s no hard data that says you have to smoke a certain number of cigarettes after which you should be screened. A tumour board—which consists of the radiologist, oncologist, radiation oncologist and surgeon—determines the best path forward for each patient.
But there is no real data to show that 12 months is the right follow-up interval, causing a lot of anxiety in patients. But artificial intelligence can fix this huge Achilles heel in screening and those time frames can be completely finetuned with AI, saving the patient a lot of anxiety and saving the healthcare system a ton of money.
Training deep learning algorithms to detect cancer
While radiologists already use diagnostic tools, AI systems are based on deep learning, which use real-world data to determine what is and isn’t a tumour. Data is based on thousands upon thousands of CT scans (in patients with and without cancer), so the machines can ‘learn’ to recognize a cancerous nodule. And the more they ‘learn,’ the more accurate they become over time. This could be an important tool in helping radiologists verify their findings or to spot something so small it’s undetectable by the human eye.
“The limits of human vision also make it easy for radiologists to overlook tiny malignant lesions. Up to 35% of lung nodules are missed at the initial screening, for example. Using AI systems can help on both counts by shifting some of the burden from busy specialists and detecting lung spots invisible to the naked eye,” according to an article in Nature.
While lung cancer screening can help to save lives, AI can make that screening much more effective. A deep learning algorithm can be trained to read X-rays, CT’s, MRIs and other medical scans, recognize patterns and interpret images for more granular, accurate diagnoses. In fact, there’s very good evidence that when you apply deep learning on the whole lung using larger data sets, it could point out areas of abnormality that might otherwise be missed.
Computer-aided detection (CADe) can detect a pulmonary nodule—a spot on the lung—that might be cancerous, while computer-aided diagnosis (CADx) differentiates whether that nodule is benign or malignant. The more diverse the data set, the better the deep learning algorithm can detect nodules and predict their risk.
Researchers from a Terry Fox Research Institute-led study published a paper in the Lancet Journal that demonstrated how AI could be used in lung cancer screening. By training a deep learning algorithm using anonymized data from more than 25,000 patients, they could accurately estimate the three-year risk for lung cancer and related mortality. This was achieved through better timing of diagnostic tests.
“DeepLR recognises patterns in both temporal and spatial changes and synergy among changes in nodule and non-nodule features. DeepLR scores could be used to accurately guide clinical management after the next scheduled repeat screening CT scan,” according to the Lancet Journal.
Detecting cancer with AI-based liquid biopsies
Better screening can save lives, while lessening the burden on the healthcare system. But it would be challenging, if not impossible, for the Canadian government to provide lung cancer screening to every Canadian. In many remote, rural or Northern areas, the nearest CT scanner may be hundreds of kilometers away. And rather than systematic Canadian coverage, there are only isolated pockets of funding, such as Cancer Care Ontario or MUHC, CHUM and IUCPQ in Quebec.
AI-based liquid biopsies could be a useful tool in determining who is at risk for lung cancer—and who needs to be sent for further screening. A liquid biopsy is a low-cost blood test that can detect the DNA fragments from cancer cells that circulate in the bloodstream. While it’s useful for different types of cancer, it has great potential for diagnosing lung cancer—and makes the screening process much easier.
Typically, examining genetic mutations and markers on lung cancer cells—in order to determine the best treatment plan—requires a biopsy (removing a piece of tissue) and sending it to pathology. Liquid biopsies, on the other hand, involve a blood test, which is particularly useful for hard-to-reach tumors or where tumor tissue is scarce. When accessing tissue through the chest, for example, there’s a higher chance of complications such as pneumothorax or bleeding.
Here, too, AI could boost the efficacy of liquid biopsies. Researchers at the Johns Hopkins Kimmel Cancer Center have developed an AI-based blood testing technology called DELFI (DNA evaluation of fragments for early interception), which spots unique patterns in the fragmentation of DNA shed from cancer cells circulating in the bloodstream. In a sample of nearly 800 people, researchers were able to detect more than 90% of lung cancers. This approach would benefit not only high-risk individuals, but the general population—including non-smokers.
Taking a collaborative approach to AI
But there’s also a critical shortage of comprehensive and multimodal data sets to enable development of more comprehensive models that can improve and personalize the approach at the different stages of the patient pathway. For example, determining who could benefit from a liquid biopsy after a CT scan to adjust wait times, speed up diagnosis and improve overall efficiency. AI can be trained to screen for lung cancer, but it needs data, and there isn’t a national open data lake in Canada. As well, most healthcare institutions don’t have the infrastructure to develop their own AI-based solutions or train deep learning algorithms.
Much of this healthcare data lives in silos, so broader infrastructure is required to connect these silos, both internally and externally—because, even if innovations are made, those innovations can’t be translated into real-life settings in the broader community. There are also privacy concerns around patient and institutional healthcare data.
Imagia is the first and only shared ecosystem that empowers clinicians with AI-ready data sets and research tools—and drives to scale with industry partners in a collaborative, multi-institutional setting. To ensure strict data governance and institutional data privacy, AI models are trained on distributed datasets while data remains encrypted. This allows for AI training on data from multiple institutions, hospitals and clinics, without sharing the patient data.
While a lung cancer diagnosis was once seen as a death sentence, innovations in AI-based lung cancer screening and liquid biopsies hold great promise for catching lung cancer at an earlier stage and improving survival rates—all while helping the healthcare system control costs and provide the best level of patient care possible.
Blog Nov 22, 2021
It seems self-evident that healthcare research could gain tremendous value from using artificial intelligence to derive insights and solutio...
Blog Nov 18, 2021
Lung cancer is the deadliest cancer in Canada—and the world. Every year, lung cancer kills more than 20,000 Canadians. Of those who receiv...
Publications Oct 15, 2021
A technical 20-page report for next-generation data governance models. Authors: Francis Dutil, Alexandre See, Lisa Di Jorio and Florent Chan...
Please complete the form in order to direct your request to the appropriate department, and we will reach out as soon as possible.