This study was approved by the Institutional Review Committee of the First Affiliated Hospital of Anhui Medical University (approval number PJ2023–07-11). The requirement for informed consent was waived because of the retrospective study design and use of known data. By searching the electronic medical records, CC patients who had cervical lesions found by transvaginal sonography (TVS) in the US Department of the First Affiliated Hospital of Anhui Medical University (Hospital 1) from June 2018 to September 2023 and CC patients of Anhui Provincial Maternity and Child Health Hospital (Hospital 2) and Nanchong Central Hospital (Hospital 3) from January 2018 to October 2023 were analyzed retrospectively.

The inclusion criteria for this study were as follows: 1) pathologically confirmed CC patients who underwent radical hysterectomy, 2) cervical biopsy-confirmed CC patients, and 3) TVS was performed less than 2 weeks before the hysterectomy or biopsy. The exclusion criteria were as follows: 1) chemotherapy or radiotherapy before US examination, 2) patients with a history of other malignancies, 3) the US images were lost or of poor quality, and 4) incomplete clinical information. Ultimately, 209 CC patients were enrolled from Hospital 1, and patients were randomly divided into the training and internal validation sets at a ratio of 7:3. Fifty-two CC patients were enrolled from Hospital 2 and Hospital 3 as the external validation set. The following baseline data were collected: age, SCCa, FIGO stage, histological type, and L by TVS. The study flowchart is shown in Figure 1 .

TVS was performed using a US probe with a frequency of 5–9 MHz. The patients’ bladder must be empty within 10 minutes before the examination; the lithotomy position was taken; a little coupling agent was placed on the top of the probe, a condom was placed, and a little coupling agent was placed on the top again; the probe was slowly inserted into the vagina until the vaginal vault was reached. The gynecological conditions in the machine were selected, and the appropriate depth was adjusted according to the size of the cervical lesion. First, the longitudinal section of the cervix was scanned, and then the transverse section was scanned. Careful and comprehensive observation of the cervical lesion was conducted, 6–10 clear and complete cervical lesion US images were collected for each patient, images were stored in a picture archiving and communication system (PACS), and one US image of the maximum section of the lesion was selected finally.

Cervical US image segmentation was completed by reader 1 (with 7 years of experience in gynecological US examination) and reader 2 (with 15 years of experience in gynecological US examination) using the 3D slicer 5.3.0 software. Regions of interest (ROIs) were manually selected and segmented ( Figure 2 ). First, reader 1 and reader 2 delineated ROIs on the cervical US images of 50 randomly selected patients. Two weeks later, reader 1 again delineated the ROIs of these 50 patients. Intra- and inter-class correlation analyses were performed separately to test the consistency of extracted features between Reader 1 and Reader 2. The intra- and inter-class correlation coefficient (ICC) values were greater than 0.75, indicating a good consistency of the extracted features. The segmentation of the remaining cervical US images and the extraction of radiomics features were performed by reader 1 alone.

US-radiomics features were extracted using the Pyradiomics 3.0.1 software, which can extract a large number of features from the US images using many engineering algorithms ( Figure 2 ). The images were standardized before feature extraction. First, features with ICC ≥ 0.75 were included in subsequent analyses. Second, batch effects were removed between Hospital 1 and the external validation set using Combat ( Figure 3 ) (owing to very few cases originating from Hospital 3, the cases from Hospital 2 and Hospital 3 were consolidated as the external validation set). Then, univariate analysis was used for these features, with p < 0.05 features classified as significant correlation variables between stage I and non-stage I CC and included in the subsequent analyses. Again, Spearman’s correlation analysis was used to assess the correlation and redundancy of the features. If Spearman’s correlation coefficient ≥0.9, the variable was considered redundant and was excluded. Finally, the least absolute shrinkage and selection operator (LASSO) algorithm was applied for further feature selection ( Figure 2 ).

First, univariate logistic regression analysis was used to analyze the correlation between clinical parameters and the CC stage. The variables with significant correlations (p < 0. 05) were included in the multivariate logistic regression analysis, and then independent predictors significantly associated with the CC stage were identified.

The selected features were trained using six ML algorithms, including linear regression (LR), decision tree (DT), support vector machine (SVM), random forest (RF), K-nearest neighbor (KNN), and eXtreme Gradient Boosting (Xgboost). Fivefold cross-validation (CV) was performed in the training data set to obtain the optimal parameter configuration. The hyperparameters of the six ML algorithms were adjusted using the grid search method and 10-fold CV. The patients in the testing data set were used to evaluate the performance of the model ( Figure 2 ).

Statistical analyses were performed in the R 4.2.2 analysis platform. Quantitative data with a normal distribution were expressed as mean ± standard deviation, while quantitative data with a non-normal distribution were expressed as the median (interquartile range). Categorical data were expressed as numbers and percentages. For univariate analysis, the chi-square test, independent sample t-test, or the Mann–Whitney U test was used. Correlations were calculated using Spearman’s test. The DeLong test was used to compare the performance of the six ML algorithms and three models. The SHAP algorithm was run using the “Xgboost” and “SHAP” Python packages. For all tests, p < 0.05 was considered statically significant.

According to the inclusion and exclusion criteria, in this study, 209 CC patients were enrolled from Hospital 1 and divided randomly into the training set (n = 146) and the internal validation set (n = 63). Fifty-two CC patients were enrolled from Hospital 2 and Hospital 3 as the external validation set. The detailed characteristics of these patients in the training, internal validation, and external validation sets are presented in Table 1 . Patients with stage I CC in the training, internal validation, and external validation sets were 51.4%, 50.8%, and 46.2%, respectively.

A total of 944 radiomics features were extracted from US images. There were 920 features with ICC ≥ 0.75, then 131 related features were obtained through univariate analysis and removed by Spearman’s correlation analysis, and 38 imaging radiomics features were left. Finally, the following 12 features were identified as the most significant by the LASSO algorithm: wavelet.HHH_glszm_ZonePercentage, wavelet.HHH_firstorder_Mean, square_glszm_SmallAreaLowGrayLevelEmphasis, wavelet.HLL_glszm_LargeAreaHighGrayLevelEmphasis, original_glszm_ZoneEntropy, square_glcm_Correlation, wavelet.HLH_glrlm_GrayLevelVariance, wavelet.LLH_glrlm_RunLengthNonUniformityNormalized, wavelet.HLH_glcm_SumSquares, wavelet.LLH_ngtdm_Strength, wavelet.HHL_glszm_GrayLevelNonUniformity, and wavelet.LLL_ngtdm_Busyness. These were all included in the US-radiomics models.

In the univariate analysis, L (p < 0.001) and SCCa (p = 0.016) showed statistically significant differences between the C0 and C1 groups. The multivariate regression analysis showed that both L and SCCa were independent predictors of stage I CC.

The diagnostic performance of the six radiomics models based on different ML algorithms is shown in Table 2 . The receiver operating characteristic (ROC) curves of these models in the training, internal validation, and external validation sets are shown in Figure 2 . Among them, the Xgboost model performed the best, with area under the curve (AUC) values in the training, internal validation, and external validation sets of 0.778, 0.751, and 0.751, respectively.

The combined model nomogram was established by combining the rad-score and clinical characteristics ( Figure 4A ). The calibration curve of the combined model all showed good agreement between the predicted and actual stage I CC in the three sets ( Figures 4B–D ).

The differential validity of the three diagnostic models (clinical, US-radiomics, and combined) is shown in Table 3 . The ROC curves of the three models in the training, internal validation, and external validation sets are shown in Figure 5 . The combined model had the best diagnostic performance, with AUC values in the training, internal validation, and external validation sets of 0.837, 0.828, and 0.839, respectively.

The Shapley summary diagram in Figure 6A showed the contribution of three factors (L, rad-score, and SCCa) in predicting the CC stage for each case. The greater the absolute distribution range of Shapley values, the more important for the diagnosis of the CC stage. The SHAP force plot can explain the assessment of each case, and each feature is a force for the visualized Shapley value, which either increases (positive value) or decreases (negative value) the prediction from baseline. The baseline is the mean Shapley value for all the predicted features. The length of the arrow in Figure 6B indicates the contribution of a feature to the Shapley value, whereas the red and blue arrows indicate positive and negative, respectively (23).

Two CC patients ( Figure 6B ), a 53-year-old woman with stage I CC and a 47-year-old woman with stage II CC, were selected for Xgboost model analysis. The cervical lesions of patient 1 had a short length (18 mm), low rad-score (0.214), and low SCCa (0.3 ng/mL), indicating a low SHAP value (−2.0), which strongly suggested that this patient was stage I CC and was consistent with the final pathological results. The cervical lesions of patient 2 had a long length (46 mm), high rad-score (0.521), and high SCCa (2.2 ng/mL), indicating a high SHAP value (1.44), which strongly suggested that this patient was non-stage I CC and was also consistent with the final pathological findings.