MethodsOne hundred patients with cervical cancer who underwent preoperative MRI between June 2020 and March 2022 were retrospectively enrolled. Training (n = 70) and testing cohorts (n = 30) were assigned by stratified random sampling. The clinical and pathological features, including age, histological subtypes, tumor grades, and node status, were compared between the two cohorts by t-test or chi-square test. Radiomics features were extracted from each volume of interest (VOI) on T2-weighted images (T2WI) and apparent diffusion coefficient (ADC) maps. The data balance of the training cohort was resampled by synthesizing minority oversampling techniques. Subsequently, the adiomics signatures were constructed by the least absolute shrinkage and selection operator algorithm and minimum-redundancy maximum-relevance with 10-fold cross-validation. Logistic regression was applied to predict the cervical cancer stages (low [I-IIa]) and (high [IIb–IV] FIGO stages). The receiver operating characteristic curve (area under the curve [AUC]) and decision curve analysis were used to assess the performance of the radiomics model.
ResultsThe characteristics of age, histological subtypes, tumor grades, and node status were not significantly different between the low [I-IIa] and high [IIb–IV] FIGO stages (p > 0.05 for both the training and test cohorts). Three models based on T2WI, ADC maps, and the combined were developed based on six radiomics features from T2WI and three radiomics features from ADC maps, with AUCs of 0.855 (95% confidence interval [CI], 0.777–0.934) and 0.823 (95% CI, 0.727–0.919), 0.861 (95% CI, 0.785–0.936) and 0.81 (95% CI, 0.701–0.918), 0.934 (95% CI, 0.884–0.984) and 0.902 (95% CI, 0.832–0.972) in the training and test cohorts.