Identifying pathological subtypes of non-small-cell lung cancer by using the radiomic features of

backgroundRadiomics provides promising opportunities in cancer diagnosis, endowing medical imaging with an increasingly important role in analyzing tumor phenotypes. Positron emission computed tomography (PET) imaging can detect functional changes before they become morphologically evident on computed tomography (CT) imaging. The aim of this study was to explore the feasibility of using quantitative PET radiomic and clinical features to identify subtypes of non-small-cell lung cancer (NSCLC).

methodsIn this study, one hundred patients who had been diagnosed with histologically confirmed NSCLC were collected retrospectively, including 61 patients with adenocarcinoma (ADC) and 39 patients with squamous cell carcinoma (SqCC). Then, the gross tumor volume (GTV) was delineated on PET images. A total of 107 features were extracted, which included 60 texture features and 47 metabolic features. The least absolute shrinkage and selection operator (LASSO) was used to select the optimal feature set, which was considered to be the best predictable features. Meanwhile, we analyzed the differences of selected features between two tumor types. Classification models were built by multivariable logistic regression analysis with three settings, namely: (I) radiomic features; (II) clinical features (smoking, age, sex, tumor size, T stage and N stage); and (III) radiomic features combined with clinical features. Finally, the area under the receiver operating characteristic (ROC) curve (AUC) was used to evaluate the performance of the classification models.

resultsFive out of 107 features were selected as the optimal feature set, which included four texture features and one metabolic feature. Significant differences were observed from these five features between ADC and SqCC subtypes (P<0.05). The radiomics features combined with clinical features model showed higher classification performance (AUC =0.781, sensitivity =1.000, specificity =0.700, accuracy =0.885) than the radiomic model (AUC =0.700, sensitivity =0.938, specificity =0.600, accuracy =0.808) and clinical features model (AUC =0.728, sensitivity =0.625, specificity =1.000, accuracy =0.769) in the validation datasets.

conclusionsThe