By Eftyhia Helis
What if a computer was better than a trained clinician for making an accurate disease diagnosis? And, what if it could detect a health issue before it has even fully developed? This may be the case in the field of lung cancer diagnosis.
According to a November 18, 2020 article by Elizabeth Svoboda, published in Nature, computers that were trained to recognize and interpret images from medical scans were as good and, in some cases, better than doctors for detecting lung cancer from patient medical imaging information. The detection was reported to be accurate even when the abnormal patterns on the scans were in such early stages that the human eye could have easily missed them.
Lung cancer affects millions of people worldwide. According to the Canadian Cancer Society, lung cancer is the most commonly diagnosed cancer in Canada and is the leading cause of cancer death for both men and women. However, due to non-specific symptoms in the early stages of the disease, in many cases a diagnosis for lung cancer comes too late in the progression of the illness, thus making it difficult to successfully provide life-saving treatments. Early screening, especially for at-risk individuals (e.g., heavy smokers, people with a family history of lung cancer, and people who are exposed to certain chemicals in the workplace), and an accurate diagnosis are critical for reducing the risk of dying from lung cancer.
Low-dose computed tomography (also called low-dose CT scan) is the recommended screening procedure for lung cancer, especially for adults who have no symptoms but are at high risk for developing the disease. The screening aims to identify any abnormal growth on the lung and assess whether it is benign (non-cancerous) or malignant (cancerous). This evaluation is usually done by a radiologist who will assess whether cancer has developed. An accurate assessment is crucial for deciding on next steps for the treatment plan of the patient, such as biopsy or surgery, but it is also important for avoiding any unnecessary procedures. In the case of lung cancer, accurate assessment can be challenging, leading to a high number of wrong diagnoses. Artificial intelligence (AI), which is based on the concept of developing and training computer systems to perform tasks that would require human intelligence, is being considered as a tool to support accurate diagnoses.
Canadian decision-makers are increasingly interested in the use of AI in the field of medical imaging. According to a recent pan-Canadian survey conducted by CADTH, AI is being used in at least 40 imaging departments across Canada (for both clinical and research purposes), with most of that use being with CT. In these imaging departments, AI is used not only for reading and interpreting images but also for improving image quality (image reconstruction) and for lowering radiation dose to manage radiation safety from exposure to CT.
But what does the evidence say for using AI in lung cancer diagnosis? CADTH reviewed available evidence from 7 studies that reported on the accuracy of AI for diagnosing lung malignancies compared with diagnosis based on human observation (i.e., diagnosis made by radiologists or other clinicians). In the identified studies, AI algorithms were used to read CT scans to classify lung nodules as either benign or malignant.
While the studies reported somewhat mixed results, there is evidence that AI models might be a promising support for improving accurate diagnosis in the field of lung cancer. Of the 7 studies CADTH reviewed, 4 studies reported that AI models were more accurate at classifying lung nodules when compared with radiologists (2 of these studies confirmed the difference in accuracy with a statistical test as well), while 3 studies found that AI models were either comparable or less accurate than human observers. In most studies that reported improved diagnostic accuracy by the AI algorithm, the authors noted that when compared with human observers, the AI models were more likely to correctly identify malignancies in individuals with cancerous nodules and were better at ruling out those nodules that were not cancerous.
The mixed findings reported in the reviewed studies may reflect the variability in the AI models that were evaluated in the studies. And, despite these promising trends, it may be premature to draw conclusions about using AI for lung nodule classification in clinical practice. More studies of high methodological quality and more real-world testing would help us understand how AI could be used for lung cancer diagnosis.
In addition, there are several aspects of AI use — including privacy and safety of patient data, how AI fits in the current clinical routine, and what happens as these AI systems evolve rapidly — that need to be considered before the technology is fully adopted in clinical practice. The hope is that AI will be a valuable assistant to doctors and support improved treatment and survival outcomes among diagnosed patients.
You can find the CADTH review described in this article at cadth.ca/artificial-intelligence-classification-lung-nodules-review-clinical-utility-diagnostic-accuracy-cost. If you’re interested in learning more about CADTH’s survey of medical imaging capacity in Canada — the Canadian Medical Imaging Inventory — visit cadth.ca/cmii, and for the latest evidence on oncology, visit cadth.ca/oncology. If you’d like to learn more about CADTH, visit cadth.ca, follow us on Twitter @CADTH_ACMTS, or speak to a Liaison Officer in your region: cadth.ca/Liaison-Officers.
Eftyhia Helis is a knowledge mobilization officer at CADTH.