An Artificial Intelligence Representation of Human Knowledge for Lung Nodule Classification

Low-dose computed tomography (LdCT), a recommended screening method for detection of early lung cancer, has high false positive (FP) rate, and lung nodule biopsy is a follow-up option to eliminate the FPs. It is challenging and meaningful to differentiate the nodule pathology by the LdCT screening data to avoid the costly interventional biopsy procedure. In this paper, we propose an artificial intelligence (AI) model to represent the human knowledge about lesion properties to differentiate the malignant nodules from benign ones. Three lesion properties in terms of heterogeneity, elasticity and growth are quantitatively represented by the proposed AI model. An augmented feature selection strategy was developed to integrate all lesion properties for the lesion classification. Experimental results show that the proposed AI model can achieve an AUC (area under the curve of receiver operating characteristics) score of 0.78 in the cases where physicians cannot determine the lesion type with AUC score around 0.5.