AI program results in early detection of lung nodules

Use of an artificial intelligence program detected lung nodules up to 1 year before their detection by radiologists, researchers reported at the virtual European Respiratory Society International Congress.

Hervé Delingette, PhD, researcher with the Epione project team of the Inria Centre at Côte d’Azur University in Nice, France, and colleagues assessed a deep-learning AI program trained on a public database of 888 low-dose CT scans for lung nodule detection in which radiologists characterized lung nodules based on radiological criteria only. Researchers evaluated this AI program in 1,179 participants of the National Lung Screening Trial who had biopsy-confirmed lung nodule malignancy with two CT scans 1 year apart.

One hundred seventy-seven patients had diagnosed cancer in the past year. These patients had the AI program applied on the image at diagnosis a year before the diagnosis was made.

From 1,179 patients, the AI program detected 75% of annotated nodules (68% of 2,352 benign nodules). This led to an average of 12 extra candidate nodules per scan.

Researchers observed the AI program detected 172 of 177 malignant lung nodules (97% sensitivity). The five lung nodules missed by the AI program were located closer to the center of the chest, which is a more difficult area to distinguish malignant areas, according to the researchers.

When testing the AI program against the CT scans from 1 year prior to lung cancer diagnosis, the program successfully detected 152 of 157 malignant lung nodules.

The researchers noted that the AI system resulted in an average of 12 false-positive detections per scan.

The AI program requires further work to improve distinguishing between abnormal but benign lung tissue and possibly cancerous tissue before being used in clinical settings to avoid unnecessary biopsies, according to a press release issued by ERS.

“We found that the system still provides a large number of false positives per scan, similar to overcomplicated diagnosis software. However, it showed good performance to detect biopsy-confirmed malignant lesions, even 1 year prior to the detections by radiologists,” Delingette said. “This is encouraging for the future use of AI systems to assist radiologists in lung cancer screening.”

Reference:

• Artificial intelligence could help diagnose lung cancer a year earlier. https://www.ersnet.org/news-and-features/news/artificial-intelligence-diagnose-lung-cancer-a-year-earlier/. Published Sept. 8, 2021. Accessed Sept. 27, 2021.