The role of pre-treatment 18F-FDG PET texture features in the prediction of nodal involvement and tumor characterization in cervical cancer

Abstract

Purpose/Objective(s): Radiomics is gaining attention in oncology given the increased availability of information from radiological images. Quantitative analysis of texture features has been evolving. However, in order to determine the added value of the texture features, their performance should be compared with conventional indices, such as Standardized Uptake Value (SUV), Metabolic Tumor Volume (MTV), and Total Lesion Glycolysis (TLG). This study aimed to analyze the association between the texture features and nodal involvement, and the correlations between the selected texture features and conventional metabolic indices. Materials/Methods: 18F-FDG PET scans were analyzed retrospectively, 31 texture features and 5 histogram indices were extracted from the baseline images of 85 cervical cancer patients. Volumes of interest were first manually contoured, then segmented by fixed threshold set to 40% of the maximum SUV in the lesion. Texture features were computed with absolute resampling method. The patients were dichotomized by nodal involvement and further subdivided into pelvic and para-aortic involvement. The area under the curve (AUC) of receiver operating characteristics was used to evaluate the discrimination performance of the features and AUC>0.7 was considered moderate-good. Correlation between texture features was assessed by Spearman’s Rank Correlation. Results: GLRLM_RLNU, GLZLM_GLNU, GLZLM_ZLNU, and NGLDM_Coarseness were identified to have AUC > 0.7 (p<0.01) in the node positive group and in subgroup with pelvic node positive. Among which, GLRLM_RLNU and NGLDM_Coarseness were highly correlated with MTV and TLG (mean absolute correlation rs Z 0.92, p<0.01) and resulted in similar prediction performances. GLZLM_ZLNU was highly correlated with SUVmax (rs Z 0.907, p<0.01). GLZLM_GLNU was highly correlated with TLG (rs Z 0.913, p<0.01). Conclusion: The texture features that had the potential to predict nodal involvement were highly correlated with MTV, TLG, and SUVmax. The correlation should be accounted for when evaluating the results from texture analysis.