Robust Identification of Subtypes in Non-Small Cell Lung Cancer Using Radiomics

Abstract

Identifying the histological phenotype of non-small cell lung cancer (NSCLC) is of crucial importance to its treatment and prognosis. The radiomics-based prediction model has the potential to non-invasively extract the tumor phenotype characteristics. However, the existing research ignores the stability of extracted features, which restricts the performance and robustness of the constructed model. While most of themethods in the literature use classification accuracy to solve theproblemofradiomics featuresstability, in this paper we propose the use ofSOM (Self-organizing Mapping) and K-means to evaluate the stability of different feature subsets. The subset with good clustering performance is selected as the optimal feature subset.When the optimal feature subset is used for modeling, compared with other feature subsets, the higher AUC(Area Under Curve) and lower SD(Standard Deviation) on the three classifiers show that the feature subset had excellent classification performance and good stability, and can distinguish NSCLC subtypes more accurately and robustly.