We identified 181 consecutive women with LACC who underwent surgeries following pretreatment MRI using a 3.0-T scanner at our institution between January 2011 and February 2017 (to ensure a minimum follow-up of 3 years). The inclusion criteria were as follows ( Supplementary Figure S1 ): (i) patients who underwent radical hysterectomies and pelvic lymphadenectomies; (ii) patients who have not received treatment before surgery and the clinicopathological data are complete; and (iii) sagittal T2-weighted imaging (T2WI) and axial diffusion-weighted imaging (DWI) were performed less than the 2-week period before surgery. The exclusion criteria were as follows: (i) lesions invisible on axial DWI or sagittal T2WI; (ii) poor image quality due to the movement of the patient during examination or the chemical shift artifact of the gas in the colorectum; (iii) patients who underwent preoperative therapies; and (iv) patients with other cancers at the same time.

Information was collected from telephone consultations, outpatient medical records, and social security death indices. In our study, the endpoint event was PFS, which was defined as the time from the date of surgery until any recurrence (local or distant recurrence, metastasis). Clinical follow-up was every 3 months for the first 2 years, every 6 months for the next 3 years, and annually thereafter. Gynecological examinations, cervical cytology, and imaging tools such as computed tomography (CT), MR, and positron emission tomography (PET)/CT imaging were used to evaluate the patients during the follow-up.

All patients underwent pelvic 3.0-T MRI scans (Signa EXCITE; GE Medical Systems, 3200N, Grandview Blvd, Waukesha, WI 53188, USA; or Achieva, Philips Healthcare, The Netherlands), using a 16-channel phased-array encoding abdominal coil. The scanning range was set to cover the entire pelvis from the level of the anterior superior iliac spine to the inferior level of the symphysis pubis. Patients had to fast at least 6 h before the examination. The standard pelvic MR scan protocol was used in this retrospective study, including multiple b-value DWI (repetition time (TR)/echo time (TE) = 3,625/74 ms, field of view (FOV): 420 × 420 mm, matrix: 256 × 128, slice thickness/gap = 5 mm/6 mm, b-values: 0.700 s/mm²), sagittal T2-weighted fat suppression (FS) images: (TR/TE = 3,200/106 ms, FOV = 320 × 320 mm, matrix = 320 × 224, slice thickness/gap = 4 mm/5 mm); axial T2-weighted fat suppression (FS) images: (TR/TE = 3,600/104 ms, FOV = 380 × 380 mm, matrix = 320 × 224. slice thickness/gap = 5 mm/6 mm); coronal T2-weighted fat suppression (FS) images: (TR/TE = 3,400/107 ms, FOV = 360 × 360 mm, matrix = 320 × 192, slice thickness/gap = 4 mm/5 mm); and axial T1-weighted imaging (T1WI): (TR/TE = 245 ms/2 ms, FOV: 380 × 380 mm, matrix: 384 × 180, slice thickness/gap = 5 mm/6 mm).

Pelvic MRI Digital Imaging and Communications in Medicine (DICOM) original images of all patients were downloaded from the Picture Archiving and Communication System (PACS) and uploaded into the ITK-SNAP (open source software; www.itksnap.org) for three-dimensional manual segmentation of MR images (15). The regions of interest (ROIs) of the entire tumor were manually outlined layer by layer by a radiologist with 5 years of experience in gynecological imaging, and the results were verified by a senior radiologist with 15 years of work experience. The radiomics workflow is displayed in Figure 1 .

Before feature extraction, the image was preprocessed, including resampling the MR image to a 1 × 1 × 1-mm³ voxel size and normalizing the image gray scale to 0 to 255. The purpose of image preprocessing is to reduce heterogeneity bias caused by different equipments and scanning parameters. Radiomic feature extraction was implemented in Artificial Intelligence Kit Version 3.0.0.R, which is a commercial software of GE Healthcare. We extracted 396 radiomic features of tumor on T2WI and DWI, respectively, and a total of 792 quantitative features for each patient. The features were divided into four groups: (I) intensity histogram (n = 84), (II) morphology (n = 40) and (III) texture (n = 668), including (a) gray-level co-occurrence matrix (GLCM, n = 288), (b) gray-level run length matrix (GLRLM, n = 344), and (c) Haralick (n = 36).

High-dimensional extraction produced various radiomic features, but not all of them were significantly associated with PFS in LACC. To develop the radiomics model, we designed a two-step procedure for dimensionality reduction and selection of radiomic features. We applied the least absolute shrinkage and selection operator (LASSO) algorithm jointly with the Cox survival analysis to select the importantly prognostic features in the training cohort (16, 17). Then, the multiple-feature-based radiomic signatures were created for predicting patients’ PFS in the training cohort. The LASSO Cox regression model analysis was completed by the “glmnet” package (18, 19).

The radiomics score (Rad-score, which was defined as the radiomics signature in the current study) was computed in the training cohort by the LASSO Cox regression model analysis. Above all, we evaluated the potential relationship of the Rad-score with PFS in the training cohort and then tested it in the validation cohort. In addition, all patients were divided into high-risk and low-risk groups according to the optimum cutoff value of the Rad-score by applying X-tile software (20). The relationship of the radiomic risk score with PFS was assessed in the training cohort and then verified in the validation cohort by using Kaplan–Meier analysis and log-rank test.

The univariate Cox analysis was used to determine clinicopathologic factors associated with PFS in all patients (n = 181). LASSO Cox regression model analysis was applied to select variables with a p value < 0.1 in the univariate Cox analysis. Except the Rad-score, a clinical model which incorporated only the independent clinicopathologic risk factors was also built to predict the 3-year PFS. The prognostic performance of the models was measured by Harrell’s concordance index (C-index), C-index = 0.5 describes a random prediction, and C-index = 1.0 implies a perfect prediction ability (21). Since the C-index method may be wrong when predicting a fixed time point (22), we also use the time-dependent area under the receiver operating characteristic (ROC) analysis to assess the t-year risk of an event (23, 24).

To provide the clinician with a quantitative approach to predict patients’ probability of 3-year PFS, and to show the incremental value of the Rad-score to the clinicopathologic risk factors, we further built a combined nomogram (combined model) as an individualized PFS prediction model that incorporated both the Rad-score and clinicopathologic risk factors for PFS prediction. The prognostic performance of the nomogram was estimated by C-index and ROC analysis. Besides, calibration curves were used to compare the predicted PFS with the actual PFS.

Statistical analysis was executed by R software (version 3.6.3) and SPSS (version 23.0). Rad-scores were divided into two groups according to the cutoff value selected by X-tile. In addition, the continuous clinical variables were converted into categorical variables on the basis of cutoff values, which were determined by ROC analysis or routine cutoff points (for size). Differences in distributions between the training and validation cohorts were assessed with the chi-squared test as appropriate. A quantitative comparison of the area under the curve (AUC) was made with the Delong test (25). The glmnet package was adopted for running LASSO–Cox. The survival package was adopted for building the Cox proportional risk model, drawing the Kaplan–Meier analysis, and calculating the C-index. The rms package was used for nomograms and calibration curves. All two-sided p values less than 0.05 were considered significant.

The clinicopathologic factors of all patients are summarized in Table 1 . Of the 181 patients included in the study, the mean age of the patients was 50.67 ± 10.81 years. Disease recurred in 33 of 181 patients (18.2%). The median follow-up time was 50.5 months (40.75–66.0) for the training cohort and 48.0 months (40.0–66.0) for the validation cohort. There were no significant differences between the training cohort and the validation cohort in clinicopathologic features (p = 0.146–0.714).

The LASSO Cox regression model was used to build a prognostic Rad-score. Eight potential predictors, six features from T2WI, and two features from DWI were included in the training cohort. Finally, the Rad-score was constructed based on the eight features, and the calculation formula is as follows: Rad-score = -0.253 × T2WI_MinIntensity + 0.210 × T2WI_ClusterShade + 0.076 × T2WI_GLCM-IDM + 0.201 × T2WI_RLM-LongRunEmphasis × T2WI_RLM-LongRunHighGreyLevelEmphasis + 0.391 × T2WI_LowIntensityLargeAreaEmphasis + 0.198 × DWI_HighIntensitySmallAreaEmphasis + 0.493 × DWI_SmallAreaEmphasis. Distributions of the Rad-score in the training and validation cohorts are displayed in Supplementary Figure S2 .

The optimum cutoff generated by X-tile was 0.543. According to the optimal cutoff value, patients were classified into high-risk (Rad-score ≥ 0.543) and low-risk (Rad-score < 0.543) groups. A Kaplan–Meier analysis ( Figure 2 ) was demonstrated that the patients in the high-risk group had shorter PFS than did the low-risk group in the training [hazard ratio (HR) = 3.279, 95% CI: 2.865–3.693, p < 0.0001] and validation cohorts (HR = 2.247, 95% CI: 1.735–2.759, p < 0.0001). Figure 3 shows typical patients who had similar clinicopathological characteristics, but their PFS time was significantly different (9 vs. 37 months). The Rad-score of patient 2 was significantly higher than that of patient 1 (1.380 vs. 0.029). In the LASSO Cox regression analysis, the Rad-score yielded a C-index of 0.778 [95% confidence interval (CI): 0.699–0.858] for the training cohort ( Table 2 ). The favorable prognostic performance of the Rad-score was further confirmed in the validation cohort (C-index, 0.816; 95% CI: 0.673–0.958).

In addition to Rad-score, we included a total of 12 clinicopathologic factors into the univariate analysis ( Table 3 ), and the results showed that age (HR = 0.511, 95% CI: 0.247–1.053, p = 0.069), LVSI (HR = 2.286, 95% CI: 1.151–4.538, p = 0.018), squamous cell carcinoma antigen (HR = 0.427, 95% CI: 0.182–1.005, p = 0.051), albumin (HR = 0.530, 95% CI: 0.252–1.113, p = 0.091), hemoglobin (HR = 5.393, 95% CI: 2.560–11.363, p < 0.0001), white blood cell (WBC) (HR = 2.233, 95% CI: 1.128–4.420, p = 0.021), and Rad-score (HR = 2.647, 95% CI: 2.002–3.501, p < 0.0001), would be identified as candidate risk factors (p < 0.1) into the LASSO Cox regression analysis.

The LASSO Cox regression model analysis showed that age, LVSI, hemoglobin, and WBC were finally selected and integrated into a clinical model ( Figure 4 ). The performance of the clinical model for 3-year PFS prediction yielded a C-index value of 0.778 (95% CI: 0.699–0.858) in the training cohort and 0.816 (95% CI: 0.673–0.958) in the validation cohort ( Table 2 ). The combined nomogram ( Figure 4 ) integrating the radiomic score and the above four clinicopathologic factors demonstrated a better discrimination both in the training (C-index, 0.821; 95% CI: 0.746–0.896) and in the validation cohort (C-index, 0.829; 95% CI: 0.699–0.959) when compared with radiomic score or clinical model alone ( Table 2 ).

In the training cohort ( Table 2 and Figure 5A ), the radiomics score (AUC, 0.804 [95% CI: 0.693–0.914]) showed a comparable prognostic performance with the clinical model (AUC, 0.774 [95% CI: 0.673–0.875]), with a p value of 0.719 ( Figure 6 ). The combined nomogram was shown to be with the highest AUC value (0.879, [95% CI: 0.811–0.947]), demonstrating a significant improvement in PFS prediction compared to the clinical model or the radiomics score (p = 0.038, p = 0.043, respectively) ( Figure 6 ). In the validation cohort ( Table 2 and Figure 5B ), the clinical model yielded an AUC of 0.707 (95% CI: 0.513–0.902). Although the radiomics score (AUC, 0.795 [95% CI: 0.653–0.937]) showed improvement compared with the clinical model, the Delong test found that no significant difference was shown between the AUCs (p = 0.458) ( Figure 6 ). The combined nomogram (AUC, 0.820 [95% CI: 0.668–0.971]) also showed improvement over the clinical model or the radiomics score for the PFS prediction in the validation cohort. However, the difference was not significant (p = 0.150, p = 0.684, respectively) ( Figure 6 ).