A newly published study shows that a deep learning algorithm that emphasizes bone suppression in assessing chest radiographs for pulmonary nodules outperforms convolutional neural network algorithms using radiologist assessments and original chest radiographs. was also demonstrated to significantly improve sensitivity.
For this study, JAMA network open, investigators investigated sensitivity rates and per-image false-positive markings (FPPI) for deep learning bone suppression (DLBS) models, convolutional neural network (CNN) algorithms, and radiologist assessment of pulmonary nodules on chest x-rays. compared. According to this study, the DLBS model was trained on data from 998 patients (mean age 54.2 years), and researchers consisted of 246 patients (mean age 55.3 years) and 205 patients (mean age 51.8 years). We evaluated the model on two external datasets. .
On the external data set, the sensitivity rates of the DLBS model were 91.5% and 92.4% compared to 79.8% and 80.4% for the CNN algorithm. The researchers also noted a slight decrease in FPPI for the DLBS model using the first external data set (.07 vs. .09 for the CNN model) and a 7% decrease for the second external data set. (.09 vs. .16 for the CNN algorithm). .
“Our DLBS algorithm subtracts the overlying bony structures from the chest X-ray image to generate the lung parenchyma image, and efficiently detects lung nodules from the lung parenchyma image because the overlying bony structure is already subtracted. We assumed we could,” writes Jin Hur, MD, Ph. .D., affiliated with Severance Hospital, Department of Radiology, Institute of Radiological Sciences, Center for Clinical Imaging Data Science, Yonsei University College of Medicine, Seoul, South Korea, and colleagues.
“The main finding was that the bone suppression model (DLBS model) was able to detect lung nodules on chest radiographs more accurately compared to the original model (CNN algorithm). We experienced improved performance in nodule detection with the help of the model.”
(Editor’s Note: For related content, seeDeep learning models may predict lung cancer risk from a single CT scan” When “Deep learning model predicts 10-year cardiovascular disease risk from chest x-rays”)
Using a second external data set, researchers also compared the DLBS model ratings with those of three thoracic radiologists with more than 5 years of experience. Hur and her colleagues noted her 14.6% higher sensitivity rate for the DLBS model (92.1%) compared to the mean sensitivity rate for radiologists (77.5%). According to this study, combining the DLBS model and radiologist assessment resulted in 12%, 15.3%, and 14.2% increases in sensitivity rates compared to those of individual radiologists. The study authors also noted that the thoracic radiologist noted that her FPPI rate decreased when using the DLBS model (7.1%) compared to not using the model (15.1%).
Regarding study limitations, the authors state that validation of deep learning models using retrospective data sets can introduce selection bias. They also noted that interstitial lung disease, pleural effusion, and pneumonia were not considered in this study. He argued that a prospective multicentre study is needed to make a determination.
