Development and clinical application of deep learning model for lung nodules screening on CT images
Comprehensive Overview
This study, published in Scientific Reports (2020), investigated whether a computer program trained to recognize patterns in medical images, known as deep learning, could reliably detect small growths in the lungs, called pulmonary nodules, on low-dose CT scans, the type of chest scan increasingly used for lung cancer screening. Radiologists reviewing large volumes of these scans face a demanding, repetitive workload, and small nodules are sometimes missed or judged inconsistently between readers. Researchers in China built and trained a deep learning algorithm using CT scans from three hospitals, then compared its ability to detect and classify nodules against a group of nine radiologists reviewing a separate set of scans, with results checked against an independent public dataset of CT scans from a different country as an additional test. After confirming the tool's accuracy, the researchers used it to analyze CT scans from more than 64,000 adults to estimate how common these lung nodules actually are in the general Chinese population, information that had been largely lacking. Across both detection and classification tasks, the deep learning algorithm matched or exceeded the accuracy of the radiologists, while completing each scan in a matter of seconds compared to roughly two minutes for a human reviewer.
Outcomes and Implications
The most notable finding was that the deep learning algorithm not only matched but in several respects outperformed radiologists at identifying suspicious lung nodules, achieving higher overall accuracy while working dramatically faster, completing each scan in about five seconds compared with roughly two minutes for a radiologist. The algorithm was somewhat less likely than radiologists to flag a true nodule, but substantially better at avoiding false alarms, meaning fewer unnecessary follow-up concerns for patients without real abnormalities. When researchers used the validated tool to screen scans from more than 64,000 adults, they found that nearly 60 percent had at least one lung nodule, a notably higher rate than most previous studies from other countries have reported. Nodules became more common with increasing age, peaked around ages 60 to 64, and were consistently more frequent in men than women across nearly every age group. The vast majority of nodules detected were small and solid, most commonly located in the upper right portion of the lungs. This matters because most detected nodules are not cancerous, and a highly sensitive but overwhelmed screening system risks either missing dangerous nodules or burying radiologists in low-value work; an accurate, fast automated tool could help ensure nodules are caught consistently while easing radiologists' workload. The authors conclude that deep learning technology is a feasible, efficient assistant for large-scale lung cancer screening, and that the high nodule prevalence found here highlights the need for clear, consistent follow-up guidance to manage this screening burden.