Benign-malignant pulmonary nodule classification in low-dose CT with convolutional features

PurposeLow-Dose Computed Tomography (LDCT) is the most common imaging modality for lung cancer diagnosis. The presence of nodules in scans does not necessarily portend cancer, as there an intricate relationship between nodule characteristics and cancer. Therefore, benign-malignant pulmonary classification at early detection a crucial step to improve diagnosis prolong patient survival. aim this study propose method predicting malignancy based on deep abstract features.MethodsTo efficiently capture both intra-nodule heterogeneities contextual information nodules, dual pathway model was developed integrate with attributes. proposed approach implemented supervised unsupervised learning schemes. A random forest added second component top networks generate results. discrimination power evaluated by calculating Area Under Receiver Operating Characteristic Curve (AUROC) metric.ResultsExperiments 1297 manually segmented show that integration context target features have great potential accurate prediction, resulting 0.936 terms AUROC, which outperformed performance Kaggle 2017 challenge winner.ConclusionEmpirical results demonstrate integrating images into unified network improves power, outperforming conventional single convolutional neural networks.