Deep learning-based prediction models can help healthcare providers diagnose small pulmonary nodules, according to a new study published in Academic Radiology.
“While it is important to apply deep learning-based models quickly and easily in the clinical field, there have been few articles focusing on their usefulness published to date,” wrote Kum J. Chae, MD, of South Korea’s Biomedical Research Institute of Chonbuk National University Hospital, and colleagues. “Thus, the purpose of the present study was to develop a fast and simple deep learning−based model for the classification of small pulmonary nodules on CT images and to preliminarily evaluate the model's performance and usefulness for human reviewers.”
Chae’s team gathered data from patients who underwent nonenhanced CT scans from January 2015 to December 2017 at a single hospital. Nodules smaller than 5 mm and larger than 20 mm were excluded from the study, a decision made “to evaluate the effects of the deep learning model on difficult-to-diagnose nodules requiring a second opinion.”
With a final dataset that included 208 nodules—114 benign and 94 malignant—the researchers assigned 30 benign and 30 malignant nodules to the test set. Another 84 benign and 64 malignant nodules were used for the training and validation set. All CT scans were reviewed by two subspecialist radiologists.
The authors developed a custom deep learning model, CT-lungNET, for this study. Its performance was then compared to AlexNET, a trusted convolutional neural network (CNN) architecture used to diagnose pulmonary nodules using transfer learning.
Overall, CT-lungNET achieved an area under the receiver operating curve (AUROC) of 0.85 when classifying small pulmonary nodules. This was higher than the AUROC recorded by AlexNET (0.82). CT-lungNET also had a faster processing speed per one image slice—0.90 seconds compared to 8.79 seconds—than AlexNET.
The model did not make a significant impact on radiologist interpretations when used as a second reviewer—but it did help nonradiologists, resulting in a mean AUC improvement of 0.13.
“Our preliminary results demonstrated that CT-lungNET showed more accurate and faster classification of small pulmonary nodules on nonenhanced chest CT compared to AlexNET,” the authors wrote. “Whereas the performance of radiologists was not increased significantly after reviewing the results of CT-lungNET, nonradiologist diagnostic accuracy was enhanced.”
The researchers observed that deep learning models such as CT-lungNET can increase the confidence in a radiologist’s conclusion, even if it doesn’t necessarily improve their reading performance.