Can Deep Learning Enhance Pulmonary Nodule Detection on Chest X-Rays?

A deep learning algorithm that emphasizes bone suppression in assessing chest X-rays for pulmonary nodules demonstrated significantly enhanced sensitivity over radiologist assessment and a convolutional neural network algorithm using original chest radiographs, according to newly published research.

For the study, published in JAMA Network Open, researchers compared sensitivity rates and false-positive markings per image (FPPI) for a deep learning bone-suppressed (DLBS) model, a convolutional neural network (CNN) algorithm and radiologist assessment of pulmonary nodules on chest X-rays. According to the study, the DLBS model was trained with data from 998 patients (mean age of 54.2) and researchers assessed the model in two external data sets comprised of 246 patients (mean age of 55.3) and 205 patients (mean age of 51.8).

In the external data sets, the DLBS model had 91.5 percent and 92.4 percent sensitivity rates in comparison to 79.8 percent and 80.4 percent for the CNN algorithm. Researchers also noted a slightly reduced FPPI for the DLBS model with the first external data set (.07 versus .09 with the CNN model) and a 7 percent reduction in the second external data set (.09 versus .16 for the CNN algorithm).

“We assumed that our DLBS algorithm could generate lung parenchymal images while subtracting the overlying bony structures from chest radiograph images and therefore efficiently detect lung nodules from lung parenchymal images as the overlying bony structures had already been subtracted,” wrote Jin Hur, M.D., Ph.D., who is affiliated with the Department of Radiology and the Research Institute of Radiological Science and Center for Clinical Image Data Science at the Severance Hospital and Yonsei University College of Medicine in Seoul, Korea, and colleagues.

“The main finding was that our bone-suppressed model (the DLBS model) could more accurately detect pulmonary nodules on chest radiographs compared with the original model (the CNN algorithm). In addition, radiologists experienced improved nodule detection performance when assisted by the DLBS model.”

(Editor’s note: For related content, see “Deep Learning Model May Predict Lung Cancer Risk from a Single CT Scan” and “Deep Learning Model Predicts 10-Year Cardiovascular Disease Risk from Chest X-Rays.”)

Using the second external data set, the researchers also compared the DLBS model assessment versus the assessment of three thoracic radiologists with more than five years of experience. Hur and colleagues noted a 14.6 percent higher sensitivity rate for the DLBS model (92.1 percent) in comparison to the mean sensitivity rate of the radiologists (77.5 percent). The combination of the DLBS model with radiologist assessment resulted in 12 percent, 15.3 percent, and 14.2 percent individual increases in sensitivity rates in comparison to the sensitivity rates for the individual radiologists, according to the study. The study authors also pointed out that the thoracic radiologists had a reduced FPPI rate when utilizing the DLBS model (7.1 percent) in comparison to not using the model (15.1 percent).

In regard to study limitations, the authors said selection bias was a possibility due to validation of the deep learning model with retrospective data sets. They also noted that interstitial lung disease, pleural effusion and pneumonia were not considered in the study. Hur and colleagues maintained that a prospective multicenter study is needed to determine the viability of the deep learning model for application in clinical practice.