Semantic Characteristic Prediction of Pulmonary Nodules Using the Causal Discovery Based on Streaming Features Algorithm

Abstract

Early diagnosis significantly improves the survival rate of lung carcinoma patients. Thus, establishing a relationship between computational features and semantic features has become key in the study of lung cancer. However, it is difficult to choose relevant computational features to predict the semantic features of pulmonary nodules. In this paper, we exploit the causal discovery based on streaming features (CD_SF) algorithm to apply computational features derived from the description of shape, size and intensity, as well as textural image features, to describe 7 semantic lung nodule features. By treating each semantic feature as an individual learning task, CD_SF predicts semantic ratings by first selecting computational features and then building a causal structure network using 10-fold cross-validation evaluation schemes on 2,636 nodules, each with a diameter greater than 3 millimeters. The experimental results suggest that the predicted semantic ratings from CD_SF are closer to the radiologists’ ratings than the existing ensemble learning (EL) algorithm without feature selection. In addition, to improve the speed of the feature selection process, we analyzed the time complexity and improved it by using causal discovery with symmetrical uncertainty based on streaming features (CD_SU_SF).