HKU Data Repository

File(s) not publicly available

The role of texture analysis of pre-treatment 18F-FDG PET/CT in the evaluation of cervical cancer

posted on 2020-08-06, 06:39 authored by Kit Chi Chan
This zipped folder contains a Microsoft Excel document recording the PET report details of the anonymized patients including the age, dates of the examinations, FIGO stage, histology, lymph node involvements, treatments received, progression/recurrence status etc. There are sub-folders containing the “Data” and “Results” arranged according to the thesis chapters (i.e. Ch.2, Ch.3 and Ch.4). The “Data” folders in each “Chapter” folder have the data (i.e. texture feature values and conventional PET indices) used for statistical analyses. The “Results” folders contain the results of Man-Whitney U Test and AUC from MedCalc.

Abstract of the Thesis

Intratumor heterogeneity has prognostic values in cervical cancer. Such heterogeneity can be depicted by 18F-FDG PET/CT imaging and then quantitatively characterized by texture features. In comparison to the conventional indices derived from PET, this thesis aimed to evaluate the discriminative performance and predictive ability of the texture features in staging, lymph node metastasis detection and disease progression or recurrence.

Materials and methods
Newly diagnosed patients with cervical cancer, who underwent pre-treatment whole body 18F-FDG PET/CT imaging were retrospectively recruited. The patients were categorized based on their FIGO stages, nodal involvements and disease progression or recurrence status. Thirty-five texture features together with the SUVmax, metabolic tumor volume (MTV) and total lesion glycolysis (TLG) of the primary cervical tumors were extracted (LIFEx software). Spearman’s Rank Correlation was used to determine the correlations between the texture features with SUVmax, MTV and TLG. Correlations among the features were also evaluated. Mann-Whitney U tests determined the differences in feature values while logistic regression was conducted to predict different status assessed by the area under the receiver-operating-characteristic curve (AUC). The analyses were performed on the following comparisons: (a) early staged and advanced stage (Chapter 2); (b) PET-negative lymph node and PET-positive pelvic lymph node only (Chapter 3); (c) PET-negative lymph node and PET-positive pelvic and para-aortic (PA) lymph node (Chapter 3); (d) patients with and without disease progression or recurrence (Chapter 4).

SUVmax, entropy derived from both the histogram and gray-level co-occurrence matrix (GLCM), together with a number of texture features that described regional heterogeneity remained significant in the analyses across the four areas of research. Twenty-one texture features were identified to be highly correlated with the conventional PET indices. Correlation was also found among the remaining fourteen features. The logistic regression models were comprised of selected uncorrelated features. The AUCs of the models in the prediction of different areas under study were (a) 0.747; (b) 0.677; (c) 0.777 and (d) 0.917. Correlation calculated based on the (GLCM) was a significant predictor in the models to predict staging (p=0.02) and PA lymph node metastasis (p=0.01). Zone percentage (p=0.03) and run percentage (p=0.04) were significant factors in the prediction of disease progression or recurrence. In comparison with AUCs of the conventional indices, there were no significant differences in the prediction of staging and lymph node metastasis. As for the prediction in disease progression or recurrence, statistical significant differences in the AUCs were observed between the model and MTV (p=0.001); and TLG (p=0.01). Yet, there was no significant difference with SUVmax (p=0.10).

In conclusion, when evaluated based on the texture features, the primary tumors possessed more heterogeneity in patients with advanced stage, nodal involvement and disease progression or recurrence. The significant predictors in the models might provide complementary information in the prediction. However, the models with the selected features did not perform better in the prediction when compared with SUVmax.


Usage metrics

    Research Postgraduates


    Ref. manager