AI software can come to the aid of overwhelmed radiologists by aiding in the crucial tasks of detection, quantification, and future risk prediction of lung cancer on CT exams -- both in low-dose screening exams and in nonscreening chest CT exams.
Radiologists inundated with lung nodule findings may find that AI can come to their aid in the crucial tasks of detection, quantification, and future risk prediction of lung cancer on CT exams -- both in low-dose screening exams and in nonscreening chest CT exams.
Among common incidental findings in medical imaging, pulmonary nodules appear on about one-third of chest CT scans leaving radiologists and primary care physicians in sometimes difficult positions of advising and managing follow-up lung imaging, procedures, and evaluations. Compounding the pressure, updated lung cancer screening recommendations for people 50 years old and older, depending on their cigarette smoking history, mean demand for workup of pulmonary nodule findings is likely to increase amid concerns about fewer expert radiologists in the field.
AI software can come to the aid of overwhelmed radiologists by aiding in the crucial tasks of detection, quantification, and future risk prediction of lung cancer on CT exams -- both in low-dose screening exams and in nonscreening chest CT exams.
Radiologists inundated with lung nodule findings may find that AI can come to their aid in the crucial tasks of detection, quantification, and future risk prediction of lung cancer on CT exams -- both in low-dose screening exams and in nonscreening chest CT exams.
Among common incidental findings in medical imaging, pulmonary nodules appear on about one-third of chest CT scans leaving radiologists and primary care physicians in sometimes difficult positions of advising and managing follow-up lung imaging, procedures, and evaluations. Compounding the pressure, updated lung cancer screening recommendations for people 50 years old and older, depending on their cigarette smoking history, mean demand for workup of pulmonary nodule findings is likely to increase amid concerns about fewer expert radiologists in the field.
"CT utilization continues to grow exponentially, meaning more volume, which places a demand on radiologists to read more and to do so as efficiently and as accurately as possible," explained Jared Christensen, MD, MBA, professor and vice chair of radiology at Duke University Medical Center (DUMC) in Durham, NC, in an interview with AuntMinnie.com. "Couple that with advances in technology, we can see more than ever before. Fifteen years ago, a typical chest CT had 50 images; now it has 500 images and the overall image quality has improved, meaning we can reliably detect nodules as small as 2 mm in size."
Jared Christensen, MD, MBA.
Christensen leads the Duke Health Lung Cancer Screening (LCS) Program. He's also chair of the American College of Radiology (ACR) Lung-RADS Committee, which develops national guidelines for the reporting, classification, and management of screen-detected nodules.
"We are diagnosing more nodules than ever before. Unfortunately, we do not have a reliable way to determine which of these nodules represent lung cancers, so it creates a large burden for imaging follow-up as growth then becomes the primary marker for a suspicious nodule," Christensen said.
Quick scan for lung cancer
With AI working behind the scenes in some way at many healthcare sites, deep learning-based computer-aided detection (CAD) software is being tested for detecting and assessing nodules detected on nonscreening chest CT exams. At ECR 2024 in Vienna, a scientific presentation highlights the possibilities for the real-world impact of AI-supported CAD for routine thoracic CT, for example, reinforcing the opportunity for incidental lung nodules at CT to catch early-stage lung cancer.
AI is also focused on malignancy risk scoring, nodule growth prediction, and lung nodule dynamics to support lung cancer screening programs. Regardless, AI-assisted software models focused on pulmonary nodules alone could relieve the demands of the large volume of lung CT findings, particularly with an increase in low-dose computed CT scans.
However, the variety of currently available products may also be a challenge for institutions seeking to broadly integrate AI for lung cancer screening.
"Most of the work in AI for LCS has focused on detection and characterization, with a more recent focus on prediction," Christensen told AuntMinnie.com. "If I want to fully integrate AI, I may have to have three or four different products to have a comprehensive clinical program.
"For example, one company may only detect and characterize solid nodules, but not atypical pulmonary cysts," Christensen explained. "I may need another product for nodule risk prediction or one that automatically assigns a Lung-RADS classification. Other products offer solutions to integrate AI tools into PACS and dictation software. And that’s only on the imaging side of the equation. I may then need another AI tool for tracking and managing patients in the screening program to ensure appropriate follow-up and adherence. Because of a precision-based approach, it is difficult to find an 'all-in-one' AI solution to LCS."
Can AI accurately determine which nodules are cancer and, of those, which are more likely to be aggressive and at high risk for developing metastatic disease? "Several nodule risk calculators are available," Christensen added, including the PanCan (or Brock), Veterans Affairs model, and others. "These use a combination of imaging and clinical data to derive a probability of a nodule being cancer. These calculators can be helpful for making treatment decisions but are not perfect," Christensen continued.
"AI-based predictive tools focus almost exclusively on radiomics -- extracting quantitative data from imaging that is not detectable to radiologists," according to Christensen. "At a high level, this includes defining imaging biomarkers or patterns that incorporate morphologic characteristics and voxel-based features to predict risk of malignancy, histology, therapeutic response, and survival. Combining AI predictive tools with clinical data is particularly robust in early validation models."
AI is diving into territory that radiologists on their own cannot really accurately assess, Christensen said.
A major predictive leap
Meanwhile, a team at Massachusetts Institute of Technology in Cambridge, MA, and Mass General Cancer Center has developed an individualized risk model for future lung cancer prediction. This model does not require clinical information, image preprocessing, or annotation. It is freely available and, if approved by the U.S. Food and Drug Administration, would make a major predictive leap in forecasting lung cancer for individuals.
Sybil, as the AI model is called, looks for predictive information captured by a single low-dose chest CT scan and synthesizes information beyond visible nodules, according to the original report published in January 2023 in the Journal of Clinical Oncology. The program was designed using a 3D convolutional neural network architecture to produce a set of six scores representing calibrated probabilities of a future lung cancer diagnosis extending one to six years following the low-dose CT scan.
The Sybil AI-based software analyzes the thinnest axial CT slices, treating each scan as a unique data point. Sybil was trained using low-dose chest CT scans from National Lung Screening Trial (NLST) tissue-confirmed lung cancers. Developers suggested that the software run in the background at a radiology reading station as soon as low-dose CT images are available since there is no need to provide demographic or other clinical data, or requirements for a radiologist to annotate areas of interest.
Florian Fintelmann, MD.
Florian Fintelmann, MD, an associate professor of radiology at Harvard Medical School, and radiologist in the division of thoracic imaging and intervention at Massachusetts General Hospital, told AuntMinnie.com that Sybil is alive and well one year after debuting.
"Sybil takes the long view," Fintelmann said, going beyond nodule detection and characterization and emphasizing that Sybil is not computationally demanding. Sybil is also open access so anyone in the world can explore the tool.
"But before we can make recommendations on how best to use the information it generates, we would need to understand how different levels of risk fit into the current management paradigms," Fintelmann said, adding, "current lung cancer screening eligibility revolves around smoking history and age but since many patients we see diagnosed with lung cancer today would not have been eligible for screening, we need a better way to identify those at risk for lung cancer. Sybil might allow us to evaluate future lung cancer risk in people who never smoked."