‘Accurate, robust’ AI detects lung nodules in chest x-rays
Convolutional neural networks (CNNs) can be trained to detect lung nodules on chest x-rays, according to a new study published in Artificial Intelligence in Medicine.
“In recent years, thanks to the large amount of available data and computational power of modern computers, CNNs have shown state-of-the-art performance in a number of computer vision applications,” wrote Xuechen Li, Shenzhen University in China, and colleagues. “Because CNNs can be trained end-to-end in a supervised framework to learn highly discriminative features, they are well suited to lung nodule detection in chest x-rays.”
The authors used data from the Japanese Society of Radiological Technology and hospitals in Guangzhou, China, and Shenzhen, China, for their research. All chest x-rays were preprocessed to improve visibility and enhanced. Multi-resolution CNNs were then trained using four different fusion methods: voting-fusion, committee-fusion, late-fusion and full-fusion.
The preferred performance metrics for this study were free-response receiver operating characteristic curve (FAUC) and refined competition performance metric (R-CPM). The full-fusion method, which used features extracted at multiple resolutions, was the most effective, achieving a FAUC of 0.982 and R-CPM of 0.987 on one of the study’s databases. The committee-fusion and late-fusion methods achieved a “slightly lower performance” than the full-fusion method in terms of both FAUC and R-CPM.
“The performance was better than the average radiologist,” the authors wrote. “Compared with previous studies, the proposed method achieved much higher sensitivity at much lower false positives per image. The proposed method was accurate, robust and has the potential to be used in real clinical practice.”
Of course, the team added, the work in this area is not complete. Li et al. already have plans for future research.
“Our future work will mainly try to employ a large chest x-ray database collected from different hospitals and machines for network training to improve the detection performance and robustness of the network,” the authors wrote. “The architecture of the CNN should also be optimized for a larger-scale database.”
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.