Patients
This study was approved by the ethics committee of our hospital, and the requirement for informed consent was waived. A total of 347 EOC patients who were diagnosed with EOC between April 2017 and June 2022 were analysised retrospectively via our hospital’s PACS system. We included 232 EOC patients by the inclusion criteria: histopathologically confirmed EOC, and receiving CT enhancement examination a week before surgery. Of the 232 EOC patients, we excluded 123 EOC patients by the following exclusion criteria: patients with preoperative chemoradiotherapy (n = 75), those clinical data such as CA125 are incomplete (n = 23), and patients with other malignant tumor or combined other malignant tumor (n = 16), with poor CT quality (n = 9). Finally, a total of 109 EOC patients were included in the study. The flowchart of patient selection was shown in Fig. 1. The clinical and pathological characteristics, contain preoperative CA125, age, and CA199 levels, FIGO stage, pathological type, and menopausal status, which were acquired from our patients’ medical records. Patients were divided into training and test cohort in a ratio of approximately 7:3.
Ct Images Acquisition And Preprocessing
CT imaging was obtained through the following CT scanners: Brilliance 16 (Philips Medical Systems, the Netherlands) or SOMATOM definition flash (Siemens Healthcare, Munich, Germany). The parameters for CT protocol were as follows: rotation time, 0.4 s; slice collimation, 64×0.625 mm; matrix, 512×512; tube current, 250 mAs; pitch, 1.21 and tube voltage, 120 kVp. After a plain scan, Left anterior elbow vein administration of the iodinated contrast media (370 mg I/mL) at a flow rate of 4.0 mL/s using a high-pressure syringe (Ulrich, German), and the enhanced arterial phase and venous phase images were acquired with a delay of 25-30s and 60-70s.
The CT images were exported in DICOM format, and folders are established and uniformly named. Before image segmentation, preprocessing was required. All CT images were preprocessedin Python 3.9, and the images were resampled to 1.0× 1.0× 1.0 mm³ by linear interpolation, and then all CT images were normalized to abdominal window. The heterogeneity of CT scan parameters was eliminated by performing resampling and intensity normalization.
Two observers (L.Y.P and W.X.W, with 3 and 8 years of gynecologic diagnosis experience, respectively), who were blinded to the histological results, independently analyzed in consensus the preprocessed CT images of all patients to assess metastasis and measure tumor size. When two observers appeared disagreement, consult another experienced radiologist (G.L.G, with 35 years of experience in gynecological diagnosis) to get the final opinion. The largest cross-section of the tumor focus was selected for measurement, and the mean value of the two measurements was calculated.
Tumor Segmentation And Feature Extraction
After the two radiologists above performed CT evaluation images to identify the tumor area, one of them (L.Y.P) was arranged to segment the tumors. The volume of interest (VOI) was manually drew layer by layer independently via 3D slicer software (version 4.13.0; www.slicer.org), and avoid blood vessels and other abdominal organs when sketching.
A total of 851 radiomics features were extracted automatically from the VOI using the 3D slicer built-in radiomics feature extraction plug-in pyradiomics (version3.0.1). The extracted features include first-order histogram features, 3D shape based features, gray-level run length matrix (GLRLM) features, gray-level dependency matrix(GLDM) features, gray-level co-occurrence matrix (GLCM) features, neighborhood gray-tone difference matrix(NGTDM) features, gray-level size zone matrix(GLSZM) features, and wavelet features based on Fourier transform. Then, Z-Score transform was used to normalize all eigenvalues.
To evaluate the reproducibility and accuracy of the extracted feature values, 20 patients were randomly selected by the same radiologist (L.Y.P) and another radiologist (W.X.W) for tumor segmentation a month later. Subsequently, intra-observer consistency analysis was conducted on the VOI drawn by the same observer, and inter-observer consistency analysis was conducted on the VOI drawn by two observers.
Feature Selection And Model Construction
The observers were evaluated by intra-class correlation coefficient (ICC). The consistency of radiomics features was extracted and the features with high consistency (ICC > 0.75) were screened. First, based on the heuristic scoring criteria, a multivariable ranking algorithm (minimum redundancy maximum relevance [mRMR]) was used to recognize the most important features, and only the top-ranked features with maximum correlation and minimum redundancy were retained. Variance threshold method, univariate selection method and the least absolute shrinkage and selection operator (LASSO) technique were used to reduce the dimension of radiomics features and select metastatic and non-metastatic related features with nonzero coefficients from the 851 radiomics features in the training set. Then, The radiomics score (Rad-score) for each patient was acquired by counting the linear combination of these features, which weighted by their corresponding nonzero coefficients.
Multivariate logistic regression method was used to find out the independent risk factors of metastasis of EOC, and a clinical model was established. Then, we combined the radscore with clinical characteristics, and to develop a combined model. Finally, we construct a visual radiomics nomogram based on the combined model. The workflow is shown in Fig. 2.
Model Evaluation
The predictive performance of the models was evaluated via the area under the curves (AUCs), sensitivity, specificity and accuracy, which was calculated by the receiver operating characteristic (ROC) curves. The net benefits of each model at different threshold probabilities was calculated by decision curve analysis (DCA), which to assess the clinical utility of the radiomics nomogram. The goodness of fit of the nomogram was assessed via the Hosmer-Lemeshow test. Calibration curve was a visual tool to assess the agreement between predictions and observations of the predicted values.
Statistical Analysis
SPSS 26.0 (IBM), Python (version 3.9) and R software (version 4.2.1 ) were used for statistical analysis. Qualitative data were compared by using the chi-square test or Fisher’s exact test. Quantitative data were tested for normal distribution by using Kolmogorov-Smimov test. Quantitative data conforming to the normal distribution were expressed as mean ± SD, and the t-test was used for comparison between groups; and the quantitative data of non-normal distribution were expressed as medians with interquartile ranges, and Mann Whitney U-test was used for comparison. A P-value < 0.05 was considered to be significant. The package version are presented in the Additional A1.