Machine Learning for Histologic Subtype Classification of Non-Small Cell Lung Cancer: A Retrospective Multicenter Radiomics Study

Abstract

Background: Histologic phenotype identification of Non-Small Cell Lung Cancer (NSCLC) is essential for treatment planning and prognostic prediction. The prediction model based on radiomics analysis has the potential to quantify tumor phenotypic characteristics non-invasively. However, most existing studies focus on relatively small datasets, which limits the performance and potential clinical applicability of their constructed models. Methods: To fully explore the impact of different datasets on radiomics studies related to the classification of histological subtypes of NSCLC, we retrospectively collected three datasets from multi-centers and then performed extensive analysis. Each of the three datasets was used as the training dataset separately to build a model and was validated on the remaining two datasets. A model was then developed by merging all the datasets into a large dataset, which was randomly split into a training dataset and a testing dataset. For each model, a total of 788 radiomic features were extracted from the segmented tumor volumes. Then three widely used features selection methods, including minimum Redundancy Maximum Relevance Feature Selection (mRMR), Sequential Forward Selection (SFS), and Least Absolute Shrinkage and Selection Operator (LASSO) were used to select the most important features. Finally, three classification methods, including Logistics Regression (LR), Support Vector Machines (SVM), and Random Forest (RF) were independently evaluated on the selected features to investigate the prediction ability of the radiomics models. Results: When using a single dataset for modeling, the results on the testing set were poor, with AUC values ranging from 0.54 to 0.64. When the merged dataset was used for modeling, the average AUC value in the testing set was 0.78, showing relatively good predictive performance. Conclusions: Models based on radiomics analysis have the potential to classify NSCLC subtypes, but their generalization capabilities should be carefully considered.