Artificial intelligence—more commonly referred to as AI—is a catch-all term that encompasses a multitude of different types of technologies and tools built to perform tasks that are typically associated with human minds. As technology has advanced over the last several decades, researchers have sought ways to apply these tools to the advancement of health care.
Within the lung cancer realm, different types of AI are being applied to advance and enhance the understanding of screening, modeling, and even treatment.
“AI is such a broad term, and in medicine we are just starting to get comfortable and understand the scope of how AI can be used to improve the health care system and outcomes for patients,” said Lecia Sequist, MD, MPH, Landry Family Professor of Medicine at Harvard Medical School and the program director of the Cancer Early Detection and Diagnostics Clinic at Mass General Cancer Center. “AI could help in so many different aspects, including screening, monitoring cancers, understanding if treatments are working, and improving patient experiences and the way hospitals work.”
In other words, the field is only viewing the tip of the iceberg when it comes to the potential applications of AI to the health care system. Lung Cancers Today recently spoke with several researchers attempting to harness these technologies in three unique ways.
Image Interpretation
Matthew Schabath, PhD, a senior member of the departments of cancer epidemiology and thoracic oncology and program leader of cancer epidemiology at Moffitt Cancer Center, describes AI technologies as one of the newer toolboxes that scientists can reach into for data analysis.
“In the case of AI, these new capabilities are especially remarkable for the analysis of medical images,” Dr. Schabath said. “Fundamental to the understanding of AI analysis of medical images is the understanding that these images are not pictures, but data.”
Dr. Schabath and colleagues at his institution and others have been working for years to harness AI to help with the early detection of lung cancer through image analysis. Recently, some of these efforts have come to fruition.
For example, late last year the United Kingdom’s University of Glasgow announced the launch of an AI-enhanced chest x-ray reporting solution that has begun clinical trials in the National Health Service.1 The initiative aims to improve early lung cancer detection and streamline the time from referral for chest x-ray to treatment.
The solution, qXR-LN (Qure.ai), distinguishes standard chest x-rays from those with abnormalities; the software can identify suspected pulmonary nodules ranging from 6 mm to 30 mm. The technology has also gained clearance from the US Food and Drug Administration in the United States.2
However, AI analysis of medical images can also be used to aid in the treatment of those already diagnosed with lung cancer. Dr. Schabath published a retrospective study looking at the use of AI to aid in biomarker-based treatment decisions in non-small cell lung cancer (NSCLC). The study used two residual-convolutional network models trained and validated with positron emission tomography/computed tomography (CT) images and clinical data from patients with NSCLC and tested with an external cohort. The deep learning scores generated by the models performed just as well as a traditional lab biomarker test in identifying patients with EGFR mutations and PD-L1 expression.3
“Traditionally, we take a portion of the lung cancer tissue and measure it for specific genes or proteins,” Dr. Schabath explained. “Now we are able to do that with AI. We don’t need the tissue.”
Expanding Screening
In her research, Dr. Sequist is coming at the issue of early diagnosis from a different angle. Instead of using image data to answer the yes or no question of whether someone has cancer, her research attempts to use AI and images to quantify future risk.
“Historically, in many people’s minds, smoking is the cause of lung cancer. That has proven to be overly simplistic,” Dr. Sequist said. “Not everyone who is going to get lung cancer will be from the high-risk population.”
Currently, the US Preventive Services Task Force (USPSTF) recommends lung cancer screening in adults aged 50 to 80 years who have a 20 pack-year smoking history and currently smoke, or who have quit within the past 15 years.4 However, it is estimated that as many as 20% of lung cancer cases are diagnosed in people who have never smoked.5
Dr. Sequist imagines more broad use of lung cancer screening that would incorporate the one-fifth of individuals who are not eligible for low-dose CT screening. For example, with cervical cancer screening, women receive a pap smear to test for the presence of human papillomavirus. Those who test positive undergo more frequent screening, and those who test negative have less frequent screening intervals.
“Everyone gets screening, and then you modulate based on what is found,” Dr. Sequist said. “With lung cancer screening we have decided to do it in a less sophisticated way, and that is not working.”
To screen a broader population, Dr. Sequist and colleagues developed a deep learning model called Sybil, which requires one low-dose CT and no clinical data or radiologist review. Built with collaborators at the Massachusetts Institute of Technology, Sybil is not looking at scans the same way a radiologist would but is instead looking for pattern recognition within the raw data contained in the CT scan, Dr. Sequist explained.
Sybil was recently put to the test using three data sets: scans from more than 6000 participants in the National Lung Screening Trial (NLST), more than 8800 low-dose CT scans from Mass General, and more than 12,000 low-dose CT scans from a hospital in Taiwan. Sybil predicted cancer within 1 year, with an area under the curve (AUC)—where 1.00 is perfect—of 0.92, 0.86, and 0.94 for the three data sets, respectively. AI could predict lung cancer within 6 years, with AUCs of 0.75, 0.81, and 0.80, respectively.6
With the encouraging results of this trial, Dr. Sequist and colleagues are now trying to validate Sybil in larger and more diverse patient cohorts to figure out the best way to make recommendations to patients.
Risk Assessment
Finally, Thomas Callender, MBChB, PhD, a senior clinical research fellow at University College London, is applying AI to try to simplify lung cancer screening risk assessments.
Currently, there are approximately 7 million people in the United Kingdom who are potentially eligible for lung cancer screening, Dr. Callender explained.
“If you were to attempt to scale the existing approach, taking about 5 minutes to go through the 17 questions and enter them into the risk assessment, we would need somewhere between 350 to 650 dedicated staff sitting there all day, every day,” Dr. Callender said. That is just for lung cancer risk assessment.
“Our interest at the moment is using risk to be able to determine both eligibility and, in due course, screening intervals,” Dr. Callender said. He and his colleagues are attempting to do that using an AI method called ensemble machine learning.
Dr. Callender said this method takes advantage of the fact that any risk-prediction model will always have a certain amount of uncertainty within it. However, when lots of those models are combined, the errors within each model will start to cancel out.
“It is a bit like having five engines working at the same time,” Dr. Callender said. “You put in a small amount of information and get a score at the end, and underneath that information has been parsed through different models and combined together.”
Dr. Callender and colleagues used just three predictors—age, smoking duration, and pack-years—in their models. These predictors were selected because of their presence in the UK Biobank, the NLST, and the US Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial datasets.
Using these three variables, their models performed the same or better than leading comparator risk models. The model predicting lung cancer death had an AUC of 0.803, and the model predicting lung cancer incidence had an AUC of 0.787 in an external validation model.7
With additional external validation using a variety of other datasets, Dr. Callender said he could envision these three factors being used as part of an annual checkup or to prompt referral to a respiratory physician.
“People would have to basically answer three questions and that could prompt the discussion of whether or not they could benefit from lung cancer screening,” Dr. Callender said. “In the [United States], the uptake of lung cancer screening is only about 10% despite being recommended for about a decade. Anything that could simplify the screening process would be useful.”
Limitations, Expectations
These examples of the use of AI in lung cancer research only scratch the surface of the possible applications of this ever-expanding technology. Additional studies are looking at using AI to predict EGFR mutation subtypes in brain metastases from lung cancer, guide management of pulmonary nodules detected incidentally on CT scans, or predict prognosis after lung cancer surgery.8-10
As research continues, it will be important to keep in mind the possible limitations of AI. One ongoing limitation is generalizability, Dr. Schabath said.
“If you train a high-performing model and your population in 95% non-Hispanic White [patients] with high income, that model is likely only applicable in that population,” Dr. Schabath said. “If you don’t have a diverse population you are going to have disparities and inequities because the model won’t include data from diverse, underserved, or minority populations.”
Dr. Schabath also acknowledged that AI can be a bit of a “black box,” with a lot of inner workings, the nuances of which only highly trained individuals really understand.
“This technology could be misused if someone is not knowledgeable,” Dr. Schabath said, equating it to trying to hammer in a nail with a screwdriver. “You have to know how to use the right tools—natural language processing, machine learning, and so on—at the right time, otherwise the data you put out there will not be reproducible.”
Finally, an important consideration for AI is keeping its advances in perspective, Dr. Sequist said. Not every tool will pan out to be as useful as promised.
“Just like medicines, not everything works. Sometimes medicines are initially hailed as groundbreaking, and it turns out later they have bad side effects and have to be pulled from the market,” Dr. Sequist said. “Some AI tools will be adopted and some discarded, but the net effect is that it will help open up knowledge, speed processes, and give more time for the human interactions that are the root of what health care is all about.”
Leah Lawrence is a health writer and editor based in Delaware.
References
Brogan J. Digital health collaboration to trial AI-powered x-rays for lung cancer detection. PharmaTimes online. December 18, 2023. Accessed February 2, 2024. https://pharmatimes.com/news/digital_health_collaboration_to_trial_ai-powered_x-rays_for_lung_cancer_detection_1505087/
Qure.ai’s AI-driven chest x-ray solution receives FDA clearance for enhanced lung nodule detection. Qure.ai. January 7, 2024. Accessed January 30, 2024. https://www.qure.ai/news_press_coverages/qure-ai-s-ai-driven-chest-x-ray-solution-receives-fda-clearance-for-enhanced-lung-nodule-detection
Mu W, Jiang L, Zhang J, et al. Abstract PR-03: Radiomics and AI-based treatment decision support for non-small cell lung cancer. Clin Cancer Res. 2021:27(5_Supplement):PR-03. doi:10.1158/1557-3265.ADI21-PR-03
Lung cancer: screening. US Preventive Services Task Force. March 9, 2021. Accessed January 30, 2024. https://www.uspreventiveservicestaskforce.org/uspstf/recommendation/lung-cancer-screening
Lung cancer among people who never smoked. Centers for Disease Control and Prevention. Accessed January 30, 2024. https://www.cdc.gov/cancer/lung/nonsmokers/index.htm
Mikhael PG, Wohlwend J, Yala A, et al. Sybil: a validated deep learning model to predict future lung cancer risk from a single low-dose chest computed tomography. J Clin Oncol. 2023;41(12):2191-2200. doi:10.1200/JCO.22.01345
Callender T, Imrie F, Cebere B, et al. Assessing eligibility for lung cancer screening using parsimonious ensemble machine learning models: a development and validation study. PLoS Med. 2023. doi:10.1371/journal.pmed.1004287
Li Y, Lv X, Chen C, et al. A deep learning model integrating multisequence MRI to predict EGFR mutation subtype in brain metastases from non-small cell lung cancer. Eur Radiol Exp. 2024;8(1):2. doi:10.1186/s41747-023-00396-z
O’Dowd E, Berovic M, Callister M, et al. Determining the impact of an artificial intelligence tool on the management of pulmonary nodules detected incidentally on CT (DOLCE) study protocol: a prospective, non-interventional multicentre UK study. BMJ Open. 2024. doi:10.1136/bmjopen-2023-077747
Kinoshita F, Takenaka T, Yamashita T, et al. Development of artificial intelligence prognostic model for surgically resected non-small cell lung cancer. Sci Rep. 2023. doi:10.1038/s41598-023-42964-8
Types of Artificial Intelligence
Machine learning (ML): identifies and analyzes patterns; machines can improve with experience
Deep learning (DL): a type of ML with multilayer neural networks that enables machines to learn and make decisions on their own
Natural language processing (NLP): a process that enables computers to extract data from human language and make decisions based on that information
Computer vision (CV): using a series of images or videos to gain information and understanding
Source: Kaul V, Enslin S, Gross SA. History of artificial intelligence in medicine. Gastrointest Endosc. 2020;92(4):807-812. doi:10.1016/j.gie.2020.06.040
PQ
“Traditionally, we take a portion of the lung cancer tissue and measure it for specific genes or proteins. Now we are able to do that with AI. We don’t need the tissue.”
—Matthew Schabath, PhD
© 2025 Mashup Media, LLC, a Formedics Property. All Rights Reserved.