Search

For traditional computer-aided diagnosis, the feature extraction of lung nodules only relies on artificial design, and the use of morphological features may lose nodular information, causing differences in classification results. This paper proposes a method of lung nodule feature extraction and classification as benign or malignant. First, the region of interest (ROI) of lung nodules was obtained from the original CT image using a threshold probability map. Next, the deep features of the lung nodules were extracted using the deep belief network (DBN). Finally, an extreme learning machine (ELM) was used as the classifier for benign and malignant classification. On the publicly available LIDC database, our method reaches a high accuracy of 95±0.3% in the diagnosis of lung nodules, and the area under the ROC curve is 0.932, which is superior to other feature extraction methods. Our method also avoids the complexity of artificial extraction and differences in feature selection, and it can provide a reference for clinical diagnosis. [ABSTRACT FROM AUTHOR]