Recognizing Non-Small Cell Lung Cancer Subtypes by a Constraint-Based Causal Network from CT Images

Abstract

The primary goal of non-small cell lung cancer (NSCLC) recognition from CT images is to discover representative features, with each being responsible for NSCLC diagnosis. A key challenge in CT image feature selection is the fact that rich causal dependencies are often neglected among either radiomics or deep learning-based features. This leads us to present a constraint-based model to construct a causal network that explicitly discovers and leverages the inherent local causal variability of these deep and radiomics features under a global view. In particular, an identified network skeleton is generated to characterize a unique causal configuration of a particular NSCLC subtype as a variable number of nodes and links, and as a result, the resulting causal network satisfies the causal Markov property and all local cause-effect dependencies are globally consistent. Furthermore, a representative node selector is devised to select the most representative causal features from the causal network for NSCLC subtype recognition. Empirical evaluations on one benchmark dataset and one in-house dataset suggest our model significantly outperforms the state-of-the-art methods.