This study retrospectively collected 157 patients who underwent gynecological pelvic MRI in our hospital from January 2018 to December 2020. Relevant clinical data including age, squamous cell carcinoma antigen (SCC) value (1–2 days before surgery) and postoperative pathological results were recorded.

The inclusion criteria for patients were as follows: ① patients underwent gynecological pelvic MRI examination before operation which included T2WI, DWI (b value was 800 s/mm${}^2$) and contrast-enhanced T1WI sequence; ② patients were pathologically confirmed to have cervical cancer, and the interval between pathological examination and MRI examination was no more than 2 weeks.

The exclusion criteria for patients were as follows: ① any form of treatment before MRI examination, such as medication, surgery, radiotherapy and neoadjuvant chemotherapy; ② lesions too small to have been accurately identified on MRI; ③ poor quality MRI, such as images with significant artifacts; ④ incomplete imaging or clinical information.

As a result, 101 patients were included in this study and the patient enrollment process is shown in Fig. 1. We assigned 60% (n = 61) of the patients to the train set and 40% (n = 40) to the test set by stratified randomized sampling.

Fig. 1.

Flowchart of inclusion. LVSI, lymphovascular space invasion; MRI, magnetic resonance imaging; T2WI, T2 weighted imaging; DWI, diffusion weighted imaging; T1WI, T1 weighted imaging.

All MRI examinations were performed on a 3.0T magnetic resonance scanner (Ingenia3.0T; Philips Healthcare, Best, the Netherlands). Magnetic resonance imaging sequences included the following: axial and sagittal T2-weighted turbo spin-echo (T2W-TSE) sequence; axial T1 high resolution isotropic volume examination (THRIVE) sequence; axial and sagittal T2 spectral presaturation attenuated inversion recovery (SPAIR). The contrast medium was Omniscan (GE Healthcare Ireland, Shanghai, China), which was injected intravenously at a rate of 3.0 mL/s with the dose of 0.2 mL/kg body weight, followed by 15 mL saline flush. The high pressure injector used was Medrad Spectris Solaris EP MR Injector System (One Medrad Drive, Indianola, PA, USA). B values of 0 and 800 s/mm${}^2$ were used for DWI and apparent diffusion coefficient (ADC) maps were automatically generated and uploaded by post-processing software. The entire MRI scanning lasted approximately 15 min.

T2WI, DWI (b value = 800 s/mm${}^2$) and T1WI + C sequences were used for analysis in this study. Two radiologists with experience in gynecological imaging read the images independently. Depth of myometrial invasion ($\ge$1/2 cervical interstitium/1/2 cervical interstitium) were evaluated simultaneously. In the case of disagreement between the two radiologists, the lesion scope and depth of myometrial invasion were finally determined after joint discussion. The maximum diameter of the lesion on the crosssection and corresponding ADC value were recorded.

MRI images of all patients (including axial T2WI, DWI and T1WI + C sequences) were imported into ITK-SNAP 3.0 software (Cognitica, Philadelphia, PA, USA; http://www.itksnap.org) in digital imaging and communications in medicine (DICOM) format. The identified lesions were manually segmented layer by layer by a radiologist with 4 years’ experience in gynecological imaging and rechecked by a second radiologist with 15 years’ experience. One week later, T2WI, DWI and T1WI + C images of 25 patients were randomly selected and outlined again for intraclass correlation coefficient (ICC) feature screening.

The criteria for manual segmentation: (1) outline along the boundary of the lesion in order for the maximum volume of interest (VOI); (2) necrosis and bleeding area inside the lesion were included in order to reflect the heterogeneity of the tumor; (3) adjacent organs such as rectum and bladder were avoided.

The workflow of radiomics is shown in Fig. 2.

Fig. 2.

Workflow of this research. MRI, magnetic resonance imaging; SVM, support vector machine; ROC, receiver operating characteristic; AUC, area under the ROC curve.

A total of 1158 features were extracted from T2WI, DWI and T1WI + C modes, including morphological features, first-order statistical features, and texture features (gray level co-occurrence matrix, GLCM; gray level run-length matrix, GLRLM; gray level region size matrix, GLSZM; gray level difference matrix, GLDM). Higher-order features were extracted after Laplacian Gaussian transform and Wavelet transform. Radiomics features were extracted using IBSI-compliant pyradiomics software (version 3.0.1) (https://www.radiomics.io/). Image preprocessing was performed before feature extraction, including resampling to a voxel size of 1 $\times$ 1 $\times$ 1, gray level discretization using binwidth of 5, and image normalization using z-score transformation. In order to evaluate the inter-observer agreement, 25 cases were randomly selected for segmentation and feature extraction. According to the 95% confidence interval estimated by intraclass correlation coefficient (ICC), it was classified as excellent (ICC = 0.90–1.00), good (ICC = 0.75–0.90), moderate (ICC = 0.50–0.75), or fair (ICC 0.50). ICC $>$0.75 was considered as good reproducibility, and features with ICC $\le$0.75 were excluded.

The 101 lesions were divided into the training set and the validation set according to time period. For the training set, features with high stability were first screened out according to the intra-group and inter-group consistency of feature extraction (ICC $>$0.75), and were then selected using the least absolute shrinkage and selection operator (LASSO) logistic regression classifier. Furthermore, the penalty parameters were adjusted by 10$\times$ cross validation, with features of non-zero coefficient being used as the correlation feature of cervical cancer lymphovascular space invasion (Fig. 3).

Fig. 3.

Selection of radiomics features using least absolute shrinkage and selection operator (LASSO) regression. (A–C) show the optimal penalty coefficient $\lambda$ (lambda) for DWI, contrast-enhanced T1 weighted imaging (TIWI + C), and T2WI modalities by adjusting different parameters using LASSO through 10-fold cross-validation, respectively. The minimum criterion and 1 standard error of the minimum criterion (1- standard error (SE) criterion) were used to draw vertical dashed lines at the optimal value. (D–F) respectively show the profile of the cable coefficients drawn according to the Log($\lambda$) sequence. Each curve represents the coefficient change trajectory of a feature. The abscissa is the parameter, and the ordinate is the coefficient.

Support vector machine (SVM) was used for the model establishment. SVM was implemented using the “e1071” package (https://CRAN.R-project.org/package=e1071) of R software (version 3.6.1, http://www.r-project.org). The type of SVM is “eps-regression”, of which the kernel function is radial basis. The total number of support vectors is 225.

All statistical analysis were performed using Rstudio (https://www.rstudio.com). Continuous quantitative data with a normal distribution were expressed as the mean $\pm$ standard deviation (SD). Categorical variables and continuous variables were compared by Chi-square test and Student t test respectively. Receiver operating characteristic (ROC) curve analysis and area under the ROC curve (AUC) were performed to detect the performances of the established models. The accuracy, sensitivity and specificity were calculated at the same time. A p value 0.05 was considered statistically significant.

A total of 101 patients were included in this study (mean age: 53.42 $\pm$ 10.01 years, range: 29–76 years) (Table 1). All the cases were pathologically confirmed as cervical cancer with the International Federation of Gynecology and Obstetrics (FIGO) stage ⅠA–ⅢC (cervical squamous cell carcinoma 84, adenocarcinoma 16 and mixed adenoendocrine carcinoma 1). 48 cases were pathologically confirmed with lymphovascular space invasion with the remaining 53 with no lymphovascular space invasion. SCC values of the enrolled patients ranged from 0.46 to 100 ng/mL, with an average of 8.66 $\pm$ 17.40 ng/mL. The average SCC value of the LVSI positive group (10.82 $\pm$ 20.11 ng/mL) was higher than that of the LVSI negative group (6.71 $\pm$ 14.45 ng/mL), however there was no significant statistical difference between the two groups.

Clinical parameters including age, preoperative SCC value and the depth of muscular invasion were collected for this study, with no significant difference being found between the LVSI positive and negative group.

SCC is a serine protease inhibitor, which inhibits cell apoptosis and participates in the metabolism, proliferation, as well as differentiation of tumor cells. The average SCC value of patients enrolled in this study was 8.66 $\pm$ 17.40 ng/mL, which was similar to Li et al. [16] result (9.7 $\pm$ 16.9 ng/mL) in patients with cervical cancer prior to treatment. Our study demonstrated that the mean value of SCC in the LVSI positive group (10.82 $\pm$ 20.11 ng/mL) was higher than that in the LVSI negative group (6.71 $\pm$ 14.45 ng/mL), but no significant difference was found between the two groups (p = 0.133). The study of Yuan et al. [17] also suggested similar results, with no significant correlation between SCC level and LVSI. However, a different conclusion was reported as positive SCC being significantly correlated with LVSI, which may be because the qualitative method used was different from the quantitative analysis method in this study [18]. Therefore, the assessment value of SCC level in evaluating LVSI of cervical cancer needs to be further explored in a larger sample size.

The maximum diameter and ADC value of the tumor were measured in this study, with no significant difference being confirmed between the two groups (p = 0.651 and 0.349 respectively). Malek et al. [19] showed that the mean ADC value of the LVSI positive group (0.83 $\pm$ 0.17 $\times$ 10${}^{-3}$ mm${}^2$/s) was lower than that of LVSI negative group (1.05 $\pm$ 0.25 $\times$ 10${}^{-3}$ mm${}^2$/s), which was similar with our results (Table 1). LVSI is an indicator of tumor invasiveness. A tumor with high invasiveness tends to grow faster with higher cell density and less extracellular space, leading to limited diffusion of water molecules, resulting in a decreased ADC value [20, 21]. Their study also showed significant differences in ADC values between the two groups, suggesting that ADC values can reflect the LVSI status of cervical cancer. However, our study did not find a significant statistical difference between the two groups, suggesting that future studies are needed for validation.

Compared with the clinical and traditional imaging parameters used in this study, radiomics parameters showed great potential in the preoperative LVSI evaluation of cervical cancer (Figs. 4,5). T2_wavelet.LHH_glcm_MC, T2_wavelet.LHH_glszm_SmallAreaLowGrayLevelEmph- asis, DWI_wavelet.HHL_glcm_Imc2 and T1C_wavelet. HLH_glcm_Imc2 showed higher weights during the selection of features. GLCM contains spatial information about the relationship of pixel pairs with similar or specific intensity within an image, while GLSZM measures groups of pixels with the same intensity regard- less of direction. Both of them can reflect the heterogeneity of the tumor. Peeken et al.’s [22] research showed similar results as the wavelet features could reflect multifrequency information of spatial heterogeneity of tumors. Among the features, wavelet.HLH_glcm_Imc2 were negatively correlated with LVSI in DWI and T1WI models, indicating lower wavelet.HLH_glcm_Imc2 and less linear structures in lesions with LVSI [23].

Numerous studies have applied radiomics in the evaluation of LVSI in cervical cancer. For example, Wu et al. [24] showed that radiomics features of T2WI, T2 with fat suppression (T2-FS), DWI and dynamic contrast enhanced (DCE) sequences can help in evaluating LVSI of cervical cancer with AUC values ranging from 0.659 to 0.814, among which, DCE sequence showed the best performance. Similar results were produced in our study, as AUC values of the single-modality radiomics parameters ranged from 0.845 to 0.929 in the training set and 0.563 to 0.739 in the validation set, of which the T1WI + C sequence showed the best performance. Similar findings may suggest that radiomics features related to tumor perfusion are better at evaluating LVSI than the diffusion-related radiomics features. Although the assessment efficiency of DCE sequence used by Wu et al. [24] is slightly superior to that of T1WI + C sequence used in this study, it requires a more rigorous operation and longer scanning time than T1WI + C sequence. Therefore, the routine sequence (T1WI + C) was eventually chosen in this study.

In addition, our study further established the radiomics model combined three different modality parameters, including T2WI, DWI and T1WI + C. The results demonstrated that performance of the comprehensive model in evaluating LVSI was significantly better than that of the single-mode radiomics model, with an AUC value of 0.990 in the training set and 0.832 in the validation set (Figs. 5,6). The research of Huang et al. [25] also showed similar results, with their AUC values of single-mode radiomics model ranged from 0.681 to 0.878, while AUC value of the comprehensive model was 0.922. This suggests that with more parameters involved, more accurate information of tumor internal characteristics can be provided, resulting in better prediction efficiency.

This study has some limitations. First, the results were established in a single center with limited sample size, and this study did not review external validation datasets. Future studies should recruit a larger cohort of patients in multi-center study to confirm our results. Second, manual segmentation of the region of interest (ROI) is a time-consuming process and more subjective, automatic segmentation is crucial and objective for promoting the feasibility of radiometric measures. Third, because of limited cases and subtypes, we did not compare the depth of myometrial invasion subgroups. Nevertheless, our results suggest that the benefits of a radiomics model to suspect the presence of LVSI of cervical cancer.