Abstract
Purpose
To evaluate the value of a nomogram model incorporating clinical parameters, hematologic inflammatory biomarkers, and MRI radiomic features in predicting postoperative disease-free survival (DFS) in patients with cervical cancer and to explore the biologic mechanism underlying the radiomic signature.
Materials and Methods
This multicenter retrospective study enrolled 804 patients with cervical cancer (2016–2023), with a median follow-up of 43.7 months. Three-dimensional radiomic features were extracted from pretreatment MRI tumor and 5-mm peritumoral regions. Five machine learning algorithms were tested to construct the optimal radiomics score (Radscore). Multivariable Cox regression was used to develop the nomogram model, with bioinformatic analysis and in vitro experiments for biologic mechanism validation.
Results
In the clinical prediction cohort (n = 751; median age, 52 years [IQR, 47–58 years]), 164 patients (21.8%) experienced recurrence during follow-up. The random survival forest–based radiomics model achieved optimal predictive performance in the external test set, with 1-, 3-, and 5-year DFS areas under the receiver operating characteristic curves (AUCs) of 0.88 (95% CI: 0.83, 0.94), 0.82 (95% CI: 0.75, 0.90), and 0.89 (95% CI: 0.83, 0.95), respectively. International Federation of Gynecology and Obstetrics stage, squamous cell carcinoma antigen level, systemic inflammation response index, and Radscore were identified as independent prognostic factors for DFS, which were incorporated into the nomogram model, with 1-, 3-, and 5-year DFS AUCs of 0.93 (95% CI: 0.90, 0.97), 0.84 (95% CI: 0.78, 0.90), and 0.86 (95% CI: 0.80, 0.92) in the external test set. A radiomic signature-TRIM29-cell cycle regulatory axis was identified and validated via transcriptomic analysis and in vitro experiments.
Conclusion
The integrated clinical-radiomic nomogram developed in this study enables accurate noninvasive prediction of postoperative recurrence in patients with cervical cancer.
Keywords: MRI, Machine Learning, Radiomics, Prognosis and Prediction, Cervical Cancer, Recurrence, Inflammatory Markers
Supplemental material is available for this article.
© RSNA, 2026
Keywords: MRI, Machine Learning, Radiomics, Prognosis and Prediction, Cervical Cancer, Recurrence, Inflammatory Markers
Summary
A nomogram integrating clinical, hematologic, and intratumoral and peritumoral MRI radiomic features predicted postoperative disease-free survival in cervical cancer, and a radiomic-TRIM29-cell cycle regulatory axis was identified.
Key Points
■ Intratumoral and peritumoral radiomic features predicted tumor recurrence in cervical cancer (recurrence rate, 21.8%), with a 5-year disease-free survival (DFS) area under the receiver operating characteristic curve (AUC) of 0.89 in the external test set.
■ A machine learning–based radiomic signature combined with clinical predictors and systemic inflammation response index improved prognostic stratification, with the integrated model achieving a 5-year DFS AUC of 0.86 (P < .001 vs radiomic-only/clinical-only models) in the external test set.
■ The radiomic signature was negatively correlated with TRIM29 expression (r = −0.70, P < .001) and associated with cell cycle regulation in cervical cancer.
Introduction
Cervical cancer remains a formidable global health challenge for female individuals because of its high incidence and mortality rates, making it a priority in female reproductive system malignancy prevention (1,2). The widespread implementation of cervical cancer screening strategies has substantially increased early-stage detection rates, creating opportunities for improved patient outcomes (3).
In accordance with guidelines from authoritative institutions such as the National Comprehensive Cancer Network and the European Society for Medical Oncology, treatment strategies are stratified by disease stage: For early-stage cervical cancer (International Federation of Gynecology and Obstetrics [FIGO] stage IA–IIA), radical surgery with systematic lymphadenectomy (combined with adjuvant therapy when indicated) is the standard of care (4,5); in contrast, for locally advanced cervical cancer (FIGO IIB–IVA), concurrent chemoradiotherapy is recommended as the primary radical treatment (6,7). Notably, a subset of patients with locally advanced cervical cancer may undergo adjuvant surgery for residual lesions or salvage surgery for recurrent tumors, while some high-risk, early-stage patients will also require postoperative adjuvant therapy (radiation therapy/chemotherapy) (4,7). This stage-adapted, comprehensive therapeutic approach has substantially improved survival outcomes, with most early-stage patients achieving long-term survival and a 5-year survival rate reaching 90% (8,9).
However, a critical clinical dilemma persists across disease stages: Approximately 15%–30% of patients experience disease recurrence or metastasis after treatment, including after radical surgery (early stage) or concurrent chemoradiotherapy (locally advanced), or even with adjuvant therapy (10). More alarmingly, recurrent disease is associated with a dramatically worsened outcome, with the median survival rate decreasing to <20% at 1 year (9,11). Accurate prediction of individual recurrence risk is crucial for delivering precision intervention and optimizing treatment regardless of the initial therapeutic modality. For surgical patients, identifying high-risk individuals preoperatively can guide the intensity of adjuvant therapy (eg, omitting unnecessary treatment for low-risk patients or escalating therapy for high-risk patients); for patients with locally advanced cervical cancer undergoing concurrent chemoradiotherapy, predicting the potential for recurrence can help select candidates for intensified treatment or close surveillance while also providing prognostic guidance for those undergoing salvage surgery (6,7).
Current clinical practices rely primarily on conventional parameters such as the FIGO stage for risk assessment. The reported recurrence rates and 5-year survival rates for patients with stage I–IV disease are approximately 10% and 50%, 23% and 40%, 42% and 20%, and 74% and 1%, respectively (12,13). While such staging can provide basic risk stratification, it is inherently limited in capturing tumor biologic heterogeneity and host response complexity, resulting in suboptimal predictive performance and failure to meet personalized medical needs (7).
Overcoming these limitations requires the integration of multidimensional biomarkers. As a hallmark of the tumor microenvironment, tumor-associated inflammation has profound effects on carcinogenesis, progression, and metastasis (14). Easily accessible hematologic indicators can provide noninvasive assessments of systemic inflammatory status, offering a macroscopic perspective on tumor biology (15). Moreover, medical imaging, particularly in the form of high-throughput radiomics, provides unique microscopic insights (16). Radiomics technology enables the deep mining and quantitative analysis of complex texture features from primary lesions that are strongly associated with intrinsic tumor heterogeneity, tumor aggressiveness, and even underlying gene expression patterns, thus demonstrating tremendous potential for predicting tumor behavior across treatment modalities (17). However, existing studies have not yet integrated clinical parameters, systemic inflammatory hematologic biomarkers, and combined intratumoral/peritumoral radiomic features into a single robust predictive model for cervical cancer recurrence, and few have validated the biologic mechanisms underlying radiomic signatures, leaving a critical unaddressed gap in prognostic model development and mechanistic interpretation of radiomic phenotypes.
The purpose of this study was to evaluate the value of a nomogram model incorporating clinical parameters, hematologic inflammatory biomarkers, and MRI intratumoral and peritumoral radiomic features in predicting postoperative disease-free survival (DFS) in patients with cervical cancer and to explore the biologic mechanism underlying the radiomic signature.
Materials and Methods
Study Sample and Design
This retrospective multicenter study included 804 patients with pathologically confirmed primary cervical cancer from five participating centers between December 2016 and December 2023, with follow-up until June 2025 (every 3–6 months for the first 2 years, every 6–12 months for the 3rd to 5th years, and annually thereafter). The study protocol was approved by the institutional review board of each center, and informed consent was waived due to the retrospective design.
Inclusion criteria were the following: (a) pathologically confirmed primary cervical cancer, (b) underwent radical hysterectomy plus pelvic lymphadenectomy, (c) underwent preoperative pelvic MRI within 2 weeks before surgery, and (d) no history of preoperative antitumor therapy.
Exclusion criteria were the following: (a) maximum tumor diameter < 1 cm, (b) poor MR image quality unfit for radiomic analysis, (c) loss to follow-up or incomplete clinical/imaging/survival data, and (d) concurrent other malignant tumors.
The primary end point was DFS, defined as the time from surgery to tumor recurrence, distant metastasis, all-cause death, or last follow-up. Patients were divided into two cohorts: (a) clinical prediction cohort (n = 751, centers 1–4), split into training set (center 1, n = 502), internal validation set (center 2, n = 126), and external test set (centers 3–4, n = 123) and (b) biological mechanism cohort (n = 53, center 5). Clinical, hematologic, and pathologic data were extracted from medical records, and systemic inflammatory indexes were calculated from preoperative blood tests. Postoperative pathologic findings were only used for outcome comparison, not model construction.
Image Analysis
Pretreatment axial T2-weighted images (Digital Imaging and Communications in Medicine format) were acquired with 1.5-T and 3.0-T MRI scanners (core parameters: repetition time, 3000–6000 msec; echo time, 80–120 msec; section thickness, 3–5 mm), imported into 3D Slicer (version 4.13.0), and processed with N4ITK bias field correction. Tumor boundaries and 5-mm peritumoral regions were manually delineated section by section by a senior radiologist [S.C., with >10 years of experience in gynecologic tumor imaging] and merged into three-dimensional volumes of interest. A total of 1409 three-dimensional texture features were extracted via PyRadiomics (version 3.0.1), followed by z score normalization to eliminate dimensional differences between features and avoid bias in model training (Fig S1).
For 100 randomly selected patients across the full cohort, volumes of interest were redrawn by the same radiologist (1-month interval, intraobserver variability) and a second radiologist (D.C., with >8 years of experience in gynecologic tumor imaging) (interobserver variability). Intraclass correlation coefficients (ICCs) were calculated, with ICC > 0.75 indicating excellent consistency. Detailed MRI acquisition parameters are provided in Appendix S1.
Construction of the Radiomic Model
Features with both intraobserver and interobserver ICC > 0.75 were retained. Redundant features were eliminated via Spearman correlation analysis (|r| > 0.80), with the feature showing stronger DFS prognostic correlation (smaller P value in univariable Cox regression in the training set) retained. Five machine learning algorithms (least absolute shrinkage and selection operator, random survival forest [RSF], gradient-boosting machine, extreme gradient boosting, and support vector machine) were used to construct models in the training set, with the internal validation set used for hyperparameter tuning. The final model output was defined as the radiomics score (Radscore). Full analysis code, along with a link to a permanent public GitHub repository, are provided in Appendix S1.
Incremental Value of Radiomics to the Clinical Model
The Radscore was integrated with clinical and hematologic markers to develop the comprehensive predictive model. Independent prognostic predictors were identified via multivariable Cox regression and incorporated into a visualized nomogram. The incremental value of the Radscore was evaluated via integrated discrimination improvement (IDI) and net reclassification improvement (NRI).
Elucidating the Biological Significance of the Radiomic Model
Radiomic features were extracted from 53 patients with cervical cancer in The Cancer Imaging Archive database. Patients were stratified by the optimal Radscore cutoff, and differentially expressed genes (DEGs) were identified via the limma package in R (version 4.3.1), with |logFC| > 1 and P < .05 set as the threshold. Functional enrichment analysis was performed via Gene Ontology and Kyoto Encyclopedia of Genes and Genomes pathway analysis. Recurrence-related DEGs were screened from the Gene Expression Omnibus (GEO) GSE44001 dataset, and overlapping genes between radiomic-related and recurrence-related DEGs were identified. Hub genes were screened via Cytoscape (version 3.10.1), and Pearson correlation analysis was used to assess the correlation between hub gene expression and Radscore. Single-cell RNA sequencing data and in vitro functional experiments were used to validate the hub gene function and its correlation with the Radscore. Detailed workflow is provided in Appendix S1.
Statistical Analysis
All statistical analyses were performed with R (version 4.3.1) and SPSS (version 26.0), with a two-sided P < .05 set as the threshold for statistical significance. Optimal cutoffs of continuous variables were determined via X-tile (version 3.6.1). Kaplan-Meier analysis with log-rank test was used to assess the association between variables and DFS.
Multivariable Cox regression analysis was performed to identify independent prognostic factors for DFS, with the model including the independent variables of interest (Radscore, FIGO stage, squamous cell carcinoma antigen [SCC-Ag] level, systemic inflammation response index [SIRI]) and adjusted confounding variables (age, tumor diameter, adjuvant therapy). Hazard ratios (HRs) with 95% CIs were calculated via the Wald method. Time-dependent receiver operating characteristic curves were plotted, and area under the receiver operating characteristic curve with 95% CI was calculated via the bootstrap method (1000 resamples). Model calibration was assessed via calibration curves, and clinical utility was evaluated via decision curve analysis. Only the external test set was used for the final evaluation of model predictive performance and stability.
Results
Patient Characteristics
A total of 1044 patients with pathologically confirmed cervical cancer who underwent radical hysterectomy were initially identified across five participating centers between December 2016 and December 2023. After applying the prespecified eligibility criteria, 240 patients were excluded for the following reasons: maximum tumor diameter < 1 cm (n = 38), poor MR image quality unfit for radiomic analysis (n = 12), loss to follow-up with incomplete survival data (n = 164), and concurrent other malignant tumors (n = 26). Finally, 804 eligible patients were included in this study, with the complete enrollment workflow shown in Figure 1.
Figure 1:
Patient enrollment flowchart and study design overview.
All included patients were divided into two predefined cohorts based on the research objectives: (a) The clinical prediction model cohort (n = 751, from centers 1–4), used for model construction, internal validation, external testing, and survival analysis (Tables 1 and S1). This cohort was further split into three nonoverlapping datasets: training set (center 1, n = 502), internal validation set (center 2, n = 126), and external test set (centers 3 and 4, n = 123). (b) The biologic mechanism exploration cohort (n = 53, from center 5), used for analyzing the molecular mechanisms underlying the radiomic features identified in this study (Table S2).
Table 1:
Baseline Demographic and Clinical Characteristics of 751 Patients with Cervical Cancer in Clinical Prediction Cohort
| Parameter | Center 1 (Training Set, n = 502) | Center 2 (Internal Validation Set, n = 126) | Center 3 (External Test Set, n = 84) | Center 4 (External Test Set, n = 39) |
|---|---|---|---|---|
| Recurrence | 108 (21.5) | 27 (21.4) | 20 (23.8) | 9 (23.1) |
| Age (y) | ||||
| Median | 53 (48–59) | 51 (47–56) | 49 (45–54) | 50 (46–55) |
| Range | 27–83 | 26–79 | 28–76 | 27–78 |
| Female sex | 502 (100.0) | 126 (100.0) | 84 (100.0) | 39 (100.0) |
| Radiologic tumor diameter (cm) | 3.6 ± 1.4 | 3.5 ± 1.3 | 3.4 ± 1.3 | 3.5 ± 1.2 |
| Lymph node metastasis (imaging) | 48 (9.6) | 12 (9.5) | 8 (9.5) | 3 (7.7) |
| Parametrial invasion (imaging) | 68 (13.5) | 17 (13.5) | 10 (11.9) | 5 (12.8) |
| Histologic type | ||||
| Squamous carcinoma | 402 (80.1) | 101 (80.2) | 70 (83.3) | 32 (82.1) |
| Adenocarcinoma | 78 (15.5) | 19 (15.1) | 10 (11.9) | 5 (12.8) |
| Adenosquamous carcinoma | 15 (3.0) | 4 (3.2) | 2 (2.4) | 1 (2.6) |
| Other | 7 (1.4) | 2 (1.6) | 2 (2.4) | 1 (2.6) |
| FIGO stage (2018) | ||||
| IB1 | 65 (12.9) | 16 (12.7) | 10 (11.9) | 4 (10.3) |
| IB2 | 118 (23.5) | 26 (20.6) | 17 (20.2) | 9 (23.1) |
| IB3 | 58 (11.6) | 15 (11.9) | 9 (10.7) | 4 (10.3) |
| IIA1 | 93 (18.5) | 26 (20.6) | 18 (21.4) | 8 (20.5) |
| IIA2 | 58 (11.6) | 15 (11.9) | 8 (9.5) | 4 (10.3) |
| IIB | 62 (12.4) | 16 (12.7) | 14 (16.6) | 7 (17.9) |
| IIIC1 | 48 (9.6) | 12 (9.5) | 8 (9.5) | 3 (7.7) |
| Tumor differentiation | ||||
| Low | 42 (8.4) | 11 (8.7) | 6 (7.1) | 3 (7.7) |
| Medium | 182 (36.3) | 46 (36.5) | 31 (36.9) | 14 (35.9) |
| High | 278 (55.4) | 69 (54.8) | 47 (56.0) | 22 (56.4) |
| SCC-Ag level (ng/mL) | 2.1 (0.8–8.5) | 2.0 (0.7–5.2) | 1.7 (0.5–4.8) | 1.7 (0.5–6.3) |
| Postoperative adjuvant therapy | ||||
| Radiation therapy | 326 (64.9) | 83 (65.9) | 47 (56.0) | 21 (53.8) |
| Chemotherapy | 335 (66.7) | 87 (69.0) | 51 (60.7) | 22 (56.4) |
Note.—Categorical data are presented as numbers, with percentages in parentheses. Continuous data are presented as means ± SDs for normally distributed data or medians, with IQRs in parentheses, for skewed data. All units are listed alongside the corresponding parameters. The total study sample included 804 patients: 751 in the clinical prediction cohort (training/validation/test sets) and 53 in the biologic mechanism exploration cohort (demographics in Table S1). FIGO = International Federation of Gynecology and Obstetrics, SCC-Ag = squamous cell carcinoma antigen.
For the clinical prediction model cohort (centers 1–4), the median age was 52 years (IQR, 47–58 years; range, 26–83 years), with a median follow-up duration of 43.7 months (range, 1–87 months). During follow-up, 164 of 751 patients (21.8%) experienced recurrence, including 87 (11.6%) with local recurrence, 56 (7.5%) with distant metastasis, and 21 (2.8%) with both local recurrence and distant metastasis. The 1-, 3-, and 5-year DFS rates for this cohort were 92.1%, 85.2%, and 72.4%, respectively.
Additionally, we evaluated the concordance between MRI-diagnosed cervical cancer characteristics and the pathologic reference standard, with high consistency between radiologic and pathologic findings (Table S3).
Radiomic Model
A total of 1409 radiomic features were extracted from both intratumoral and peritumoral regions. After quality control, 802 features with high robustness (intraobserver and interobserver ICC > 0.75) were retained. Spearman correlation analysis further eliminated redundant collinear features (|r| > 0.80, retaining the feature with smaller P value in univariable Cox regression), yielding 233 intratumoral and 201 peritumoral features for model construction.
In the external test set, both intratumoral and peritumoral radiomic features showed independent prognostic value for cervical cancer recurrence (intratumoral RSF model 5-year DFS AUC = 0.78 [95% CI: 0.72, 0.84]; peritumoral RSF model 5-year DFS AUC = 0.76 [95% CI: 0.70, 0.82]), and the combination of both regions yielded significantly higher predictive performance (combined RSF model 5-year DFS AUC = 0.89 [95% CI: 0.83, 0.95]; DeLong test vs intratumoral model: P = .008, vs peritumoral model: P = .003). Among the five tested machine learning algorithms, the RSF-based model achieved the optimal predictive performance in the independent external test set, with 1-, 3-, and 5-year DFS AUCs of 0.88 (95% CI: 0.83, 0.94), 0.82 (95% CI: 0.75, 0.90), and 0.89 (95% CI: 0.83, 0.95), respectively (Table 2, Fig 2). Feature selection, model training, and hyperparameter tuning were performed exclusively in the training set (center 1, n = 502), with the internal validation set (center 2, n = 126) used only for hyperparameter optimization; the full performance results of the training and internal validation sets are provided in Table S4.
Table 2:
AUC Values of Different Machine Learning Classifiers for Predicting Disease-Free Survival in External Test Set
| Region and Time Point | LASSO | RSF | GBM | XGBoost | SVM |
|---|---|---|---|---|---|
| Intratumoral | |||||
| 1 year | 0.67 | 0.82 (0.76, 0.88) | 0.73 | 0.61 | 0.60 |
| 3 years | 0.59 | 0.80 (0.73, 0.87) | 0.62 | 0.54 | 0.60 |
| 5 years | 0.61 | 0.78 (0.72, 0.84) | 0.63 | 0.64 | 0.61 |
| Peritumoral | |||||
| 1 year | 0.61 | 0.79 (0.73, 0.85) | 0.65 | 0.60 | 0.62 |
| 3 years | 0.58 | 0.78 (0.71, 0.85) | 0.55 | 0.56 | 0.59 |
| 5 years | 0.58 | 0.76 (0.70, 0.82) | 0.55 | 0.62 | 0.62 |
| Combined | |||||
| 1 year | 0.72 | 0.88 (0.83, 0.94) | 0.78 | 0.61 | 0.65 |
| 3 years | 0.68 | 0.82 (0.75, 0.90) | 0.77 | 0.55 | 0.60 |
| 5 years | 0.67 | 0.89 (0.83, 0.95) | 0.72 | 0.66 | 0.64 |
Note.—Unless otherwise noted, all values are areas under the receiver operating characteristic curves (AUCs) for disease-free survival prediction in the independent external test set (n = 123), rounded to two decimal places. Values in parentheses are 95% CIs for the optimal random survival forest (RSF) model; full performance results of the training set and internal validation set are provided in Table S4. GBM = gradient-boosting machine, LASSO = least absolute shrinkage and selection operator, SVM = support vector machine, XGBoost = extreme gradient boosting.
Figure 2:
Machine learning–based radiomics model for disease-free survival (DFS) prediction. (A–C) Comparative analysis of composite AUC values (composite AUC = [1-year + 3-year + 5-year AUC]/3) across tumor, peritumoral, and combined models in all datasets demonstrates superior predictive performance of the combined model. (D–F) Comparison of AUC values for 1-/3-/5-year DFS prediction among different machine learning–integrated models confirms the optimal performance of the RSF-integrated model. (G–I) ROC curves of the RSF-integrated model for DFS prediction. (J, K) Feature importance ranking by importance scores and SHAP analysis in the RSF model identifies key predictive factors. AUC = area under the ROC curve, GBM = gradient-boosting machine, LASSO = least absolute shrinkage and selection operator, ROC = receiver operating characteristic, RSF = random survival forest, SHAP = SHapley Additive exPlanations, SVM = support vector machine, XGBoost = extreme gradient boosting.
Feature importance ranking via importance scores and SHapley Additive exPlanations (ie, SHAP) analysis identified the key radiomic parameters driving model performance. The full results of feature importance ranking and SHAP analysis are provided in Table S5. The quality of radiomic analysis was evaluated via the CheckList for EvaluAtion of Radiomics (ie, CLEAR), Radiomics Quality Score, and METhodological RadiomICs Score (ie, METRIC), with scores of 31 of 36 for Radiomics Quality Score and 93.1% for METRICS (Tables S6–S8).
Comprehensive Predictive Model
Univariable log-rank tests showed significant associations between DFS and SCC-Ag levels (HR = 1.32 [95% CI: 1.15, 1.69], P = .001), parametrial invasion (HR = 1.25 [95% CI: 1.11, 1.58], P = .004), lymph node metastasis (HR = 2.43 [95% CI: 1.18, 4.97], P = .035), FIGO stage (P = .030), tumor differentiation (P = .019), postoperative radiation therapy (HR = 1.44 [95% CI: 1.02, 1.82], P = .043), SIRI (HR = 1.30 [95% CI: 1.16, 1.54], P = .001), and Radscore (HR = 3.83 [95% CI: 1.94, 7.21], P < .001). Subsequent univariable and multivariable Cox regression analyses identified SCC-Ag levels, FIGO stage, SIRI, and Radscore as independent prognostic factors for postoperative DFS (Table 3), which were incorporated into the final nomogram model (Fig 3).
Table 3:
Univariable and Multivariable Cox Regression Analyses of Factors Associated with Postoperative Disease-Free Survival
| Parameter | Univariable Cox Regression | Multivariable Cox Regression | ||
|---|---|---|---|---|
| Hazard Ratio | P Value | Hazard Ratio | P Value | |
| SCC-Ag level (ng/mL) | 1.32 (1.15, 1.69) | .001 | 1.84 (1.22, 2.78) | .004 |
| Parametrial invasion | 1.25 (1.11, 1.58) | .004 | … | … |
| Lymph node metastasis | 2.43 (1.18, 4.97) | .035 | … | … |
| FIGO stage | ||||
| I | Reference | … | Reference | … |
| II | 1.23 (1.13, 1.50) | .041 | 1.35 (0.88, 2.07) | .17 |
| III | 2.04 (1.42, 3.82) | .030 | 2.18 (1.05, 4.52) | .04 |
| Tumor differentiation | ||||
| Low | Reference | … | … | … |
| Medium | 0.44 (0.24, 0.76) | .019 | … | … |
| High | 5.31 (3.60, 7.78) | .061 | … | … |
| Postoperative radiation therapy | 1.44 (1.02, 1.82) | .043 | … | … |
| SIRI | 1.30 (1.16, 1.54) | .001 | 1.69 (1.11, 2.57) | .01 |
| Radscore | 3.83 (1.94, 7.21) | <.001 | 2.71 (1.74, 4.23) | <.001 |
Note.—Values in parentheses are 95% CIs. FIGO = International Federation of Gynecology and Obstetrics, Radscore = radiomics score, SCC-Ag = squamous cell carcinoma antigen, SIRI = systemic inflammation response index.
Figure 3:
Nomogram model for predicting postoperative disease-free survival (DFS). (A) The nomogram incorporates squamous cell carcinoma (SCC) antigen levels, International Federation of Gynecology and Obstetrics (FIGO) stage, systemic inflammation response index (SIRI), and radiomics score (Radscore). (B) Forest plot displays hazard ratios (HRs) with 95% CIs for each predictor. (C) Calibration curves for 1-/3-/5-year DFS demonstrate excellent model fit. (D) Decision curve analysis shows superior net clinical benefit of the nomogram compared with individual factors/clinical models. (E) Clinical impact curve confirms high predictive accuracy. (F–H) Receiver operating characteristic curves of the model’s 1-/3-/5-year DFS prediction performance across training/validation/test sets. AUC = area under the receiver operating characteristic curve, OS = overall survival.
In the independent external test set, the comprehensive nomogram achieved 1-, 3-, and 5-year DFS AUCs of 0.93 (95% CI: 0.90, 0.97), 0.84 (95% CI: 0.78, 0.90), and 0.86 (95% CI: 0.80, 0.92), respectively. The full performance results of the nomogram in the training and internal validation sets are provided in Figure S2.
In the external test set, calibration curves showed excellent agreement between the nomogram-predicted and actual 1-, 3-, and 5-year DFS rates (Hosmer-Lemeshow-Doornbos test, P = .87, P = .82, and P = .79 for 1-, 3-, and 5-year DFS, respectively). Decision curve analysis indicated that the nomogram provided superior net clinical benefit compared with the clinical-only model across the majority of threshold probabilities (threshold probability 0.1–0.8, net benefit difference 0.12–0.28, P < .05 for all comparisons). Clinical impact curves confirmed that the nomogram had high predictive accuracy for recurrence events, with 92.3% concordance between predicted and observed high-risk patients.
Incremental value analyses confirmed that the integrated model significantly improved the predictive capability of the clinical-only model in the external test set, with 1-year DFS IDI of 0.026 (P = .003), NRI of 0.120 (P = .028); 3-year DFS IDI of 0.046 (P = .001), NRI of 0.233 (P = .041); and 5-year DFS IDI of 0.026 (P = .001), NRI of 0.020 (P = .013). DeLong test results further confirmed that the integrated nomogram had significantly higher 1-year DFS AUC than the radiomic-only model (0.93 vs 0.88, P = .042) in the external test set (Tables S9, S10).
Biological Mechanisms Underlying the Radiomic Model
Using the optimal Radscore cutoff (4.59) for the 53 TCIA patients, we identified 1615 radiomic feature–associated DEGs (1203 upregulated, 412 downregulated, |logFC| > 1, P < .05), which were enriched in cell cycle regulation, ubiquitination, and proliferation pathways (Fig 4). Analysis of 300 patients with recurrence data in the GEO GSE44001 dataset identified recurrence-associated DEGs (311 upregulated, 43 downregulated) with consistent pathway enrichment patterns.
Figure 4:
Radiomics-associated differentially expressed genes and functional enrichment analysis. (A, B) Volcano plot and functional enrichment of Radscore-stratified DEGs. (C, D) Volcano plot and enrichment of recurrence-stratified DEGs. (E) Venn diagram of overlapping DEGs of Radscore and recurrence. (F) Top 10 hub genes. (G) Heatmap of correlations between hub genes and Radscore. (H, I) Validation of TRIM29 expression in cervical cancer and its prognostic value for disease-free survival. DEG = differentially expressed gene, Radscore = radiomics score.
Intersection analysis identified 61 overlapping genes between radiomic-associated and recurrence-associated DEGs. DSC2, JUP, KRT6A, DSG3, TRIM29, DSP, DSC3, PKP1, TP63, and PERP were identified as the top 10 hub genes via Cytoscape. TRIM29 expression was negatively correlated with the Radscore (r = −0.70, P < .001) and was overexpressed in cervical cancer tissues. TRIM29 expression was associated with DFS in patients with cervical cancer (P < .05).
Validation of Hub Genes
Single-cell RNA sequencing data from GEO dataset GSE279998 (16 cervical cancer samples) showed that TRIM29 was predominantly expressed in malignant tumor cells (Fig 5). Functional experiments in SiHa and HeLa cells showed that TRIM29 overexpression inhibited cell proliferation and G2/M phase transition (P < .05), while there was no evidence of an effect on cell migration or invasion.
Figure 5:
Functional and correlational validation of hub genes. (A) Single-cell analysis demonstrates predominant TRIM29 expression in malignant cells. (B–E) CCK-8 and colony formation assays confirm TRIM29-mediated proliferation inhibition. (F, G) Cell cycle analysis shows TRIM29-induced G2/M phase arrest. (H–M) Scratch wound and Transwell assays indicate minimal effects of TRIM29 on migration/invasion. (N–P) Immunofluorescence staining findings validate the negative TRIM29-Radscore correlation (ρ = −0.711, P < .001) and its association with prolonged disease-free survival (DFS). HR = hazard ratio, Radscore = radiomics score.
Immunofluorescence analysis of 80 cervical cancer specimens from center 1 confirmed a strong negative correlation between TRIM29 expression and the Radscore (ρ = −0.71, P < .001). Clinically, patients with high TRIM29 expression had prolonged DFS (Fig S4). A complete assessment of a patient with cervical cancer is shown in Figure S5.
Discussion
In our study, data from 804 patients with cervical cancer across five multicenter cohorts were collected to construct and validate a nomogram model integrating radiomic features and clinical factors for predicting postoperative DFS in patients with cervical cancer. Against the clinical dilemma of 15%–30% postoperative recurrence in cervical cancer, this study aimed to establish a noninvasive predictive model integrating multidimensional biomarkers. Key findings with statistically significant differences are as follows: (a) intratumoral and peritumoral radiomic features had synergistic prognostic value for recurrence; (b) the RSF-based radiomic model achieved a 5-year DFS AUC of 0.89 (95% CI: 0.83, 0.95) in the external test set; (c) the integrated nomogram had incremental predictive value over the clinical-only model (IDI and NRI P < .05), with a 5-year DFS AUC of 0.86 (95% CI: 0.80, 0.92) in the external test set; and (d) the radiomic signature was significantly correlated with TRIM29-mediated cell cycle regulation (r = −0.70, P < .001). To our knowledge, our study represents the first comprehensive "imaging feature–prognostic model–driver gene–functional mechanism" cascade analysis in the field.
Traditional morphologic imaging is limited in capturing intratumoral heterogeneity, while radiomics enables noninvasive assessment of tumor biologic characteristics via quantitative texture feature mining, with three-dimensional radiomic features consistently showing superior performance in capturing spatial tumor heterogeneity and survival prediction (18–23). Specifically, multiregion radiomic profiling enables precise delineation of tumor habitat heterogeneity, including both intratumoral and periturmoral microenvironments. Unlike conventional radiomics studies that focus only on intratumoral regions, this research incorporated peritumoral radiomic features, which reflect the tumor microenvironment closely related to recurrence and metastasis (24,25). Among the five tested machine learning algorithms, the RSF-based model showed the most stable predictive performance across datasets, which may be attributed to its ability to preserve recurrence time information and handle nonlinear interactions among features (26). This is consistent with evidence from multicenter radiomic studies adopting interpretable machine learning frameworks that integrate imaging signatures with clinical parameters (27–29). These findings provide a robust methodological basis for noninvasive recurrence prediction in cervical cancer.
Beyond predictive accuracy, the integrated nomogram showed favorable clinical translational potential. IDI and NRI analyses confirmed that the Radscore substantially enhanced the predictive power of the conventional clinical model, indicating that radiomics complements rather than replaces current clinical risk stratification systems. The nomogram integrating Radscore, FIGO stage, SCC-Ag level, and SIRI showed robust and stable predictive performance, which may aid in identifying high-risk patients preoperatively, guiding personalized adjuvant therapy decisions, and optimizing postoperative surveillance strategies for patients with cervical cancer.
Furthermore, this study bridged radiomic phenotypes with molecular mechanisms, which is a key gap in existing radiomics research. Cross-analysis of TCIA and GEO GSE44001 datasets identified TRIM29 as the core hub gene correlated with the radiomic signature, which was validated via transcriptomic, single-cell sequencing and in vitro functional experiments. Consistent with previous studies, TRIM29 acts as an E3 ubiquitin ligase regulating cell cycle progression and proliferation in cervical cancer (30–32), which provides a biologic explanation for the prognostic value of the radiomic signature.
This study had some limitations. First, this study was a retrospective multicenter analysis, which may have been subject to inherent selection bias; prospective, multicenter clinical trials are needed to further validate the clinical utility of the model in future studies. Second, owing to the limited sample size of the TCIA database, the radiomic-transcriptomic correlation analysis may have been underpowered, and the TRIM29-mediated regulatory network needs further in-depth exploration. Third, a unified multivendor MRI acquisition protocol was lacking in this study; although image correction and feature standardization were performed, a standardized imaging pipeline is still needed to improve the generalizability of the radiomics model in clinical practice.
In conclusion, this multicenter study developed and validated an integrated clinical-radiomic nomogram for accurate noninvasive prediction of postoperative recurrence in patients with cervical cancer, with multilevel validation of its underlying biologic mechanism. This model has the potential to be translated into a clinical decision-support tool to guide personalized treatment and surveillance for patients with cervical cancer. Future research will focus on the prospective validation of the model in multicenter clinical cohorts and the exploration of radiomic biomarkers for immunotherapy response prediction in cervical cancer.
Supplemental Files
Acknowledgments
Acknowledgments
All authors sincerely thank two senior radiologists, Fei Gao, MD, and Chao Wei, MD, for delineating lesions and image analysis for this study. Yaqiong Ge, MPH, and Tienan Fang, MD, kindly provided statistical advice for this manuscript.
S.C. and Y.Z. contributed equally to this work.
Funding: This research was funded by Machine Learning-Based Radiomics Model for Predicting Disease-Free Survival and Immune Status in Endometrial Cancer Patients (no: 2024FYNUFN17).
Abbreviations:
- AUC
- area under the receiver operating characteristic curve
- DEG
- differentially expressed gene
- DFS
- disease-free survival
- FIGO
- International Federation of Gynecology and Obstetrics
- GEO
- Gene Expression Omnibus
- HR
- hazard ratio
- ICC
- intraclass correlation coefficient
- IDI
- integrated discrimination improvement
- NRI
- net reclassification improvement
- Radscore
- radiomics score
- RSF
- random survival forest
- SCC-Ag
- squamous cell carcinoma antigen
- SIRI
- systemic inflammation response index
Disclosures of conflicts of interest
Please see ICMJE form(s) for author conflicts of interest. These have been provided as supplemental materials.
References
- 1. Voelker RA . Cervical cancer screening . JAMA 2023. ; 330 ( 20 ): 2030 . [DOI] [PubMed] [Google Scholar]
- 2. Francoeur AA , Monk BJ , Tewari KS . Treatment advances across the cervical cancer spectrum . Nat Rev Clin Oncol 2025. ; 22 ( 3 ): 182 – 199 . [DOI] [PubMed] [Google Scholar]
- 3. Lycke KD , Kahlert J , Petersen LK , et al . Untreated cervical intraepithelial neoplasia grade 2 and subsequent risk of cervical cancer: population based cohort study . BMJ 2023. ; 383 : e075925 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Wu T , Lucas E , Zhao F , Basu P , Qiao Y . Artificial intelligence strengthens cervical cancer screening - present and future . Cancer Biol Med 2024. ; 21 ( 10 ): 864 – 879 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Ramirez PT , Robledo KP , Frumovitz M , et al . LACC Trial: Final Analysis on Overall Survival Comparing Open Versus Minimally Invasive Radical Hysterectomy for Early-Stage Cervical Cancer . J Clin Oncol 2024. ; 42 ( 23 ): 2741 – 2746 . [DOI] [PubMed] [Google Scholar]
- 6. McCormack M , Eminowicz G , Gallardo D , et al. ; INTERLACE investigators . Induction chemotherapy followed by standard chemoradiotherapy versus standard chemoradiotherapy alone in patients with locally advanced cervical cancer (GCIG INTERLACE): an international, multicentre, randomised phase 3 trial . Lancet 2024. ; 404 ( 10462 ): 1525 – 1535 . [Published correction appears in Lancet 2025;405(10477):468.] [DOI] [PubMed] [Google Scholar]
- 7. Xu W , Xu M , Wang L , et al . Integrative analysis of DNA methylation and gene expression identified cervical cancer-specific diagnostic biomarkers . Signal Transduct Target Ther 2019. ; 4 : 55 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Kechagias KS , Kalliala I , Bowden SJ , et al . Role of human papillomavirus (HPV) vaccination on HPV infection and recurrence of HPV related disease after local surgical treatment: systematic review and meta-analysis . BMJ 2022. ; 378 : e070135 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Wei G , Guo F , Qu A , et al . PD-1 Inhibitor Maintenance Therapy Combined Iodine-125 Seed Implantation Successfully Salvage Recurrent Cervical Cancer after CCRT: A Case Report . Curr Oncol 2021. ; 28 ( 6 ): 4577 – 4586 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Zhang Y , Zhang K , Jia H , et al . IVIM-DWI and MRI-based radiomics in cervical cancer: Prediction of concurrent chemoradiotherapy sensitivity in combination with clinical prognostic factors . Magn Reson Imaging 2022. ; 91 : 37 – 44 . [DOI] [PubMed] [Google Scholar]
- 11. Saraswati W , Dahlan EG , Saputra K , Sutrisno TC . Effect of Electroacupuncture on Natural-Killer Cells and Tumor Size in Patients with Cervical Squamous-Cell Carcinoma: A Randomized Controlled Trial . Med Acupunct 2019. ; 31 ( 1 ): 29 – 36 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Kim HJ , Kim LK , Lee SH , Park SA , Eoh KJ , Kim YT . Expression levels of the long noncoding RNA steroid receptor activator promote cell proliferation and invasion and predict patient prognosis in human cervical cancer . Oncol Lett 2018. ; 16 ( 4 ): 5410 – 5418 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Zheng X , Gu H , Cao X , et al . Tislelizumab for cervical cancer: A retrospective study and analysis of correlative blood biomarkers . Front Immunol 2023. ; 14 : 1113369 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Bao L , Wu Y , Ren Z , et al . Comprehensive pan-cancer analysis indicates UCHL5 as a novel cancer biomarker and promotes cervical cancer progression through the Wnt signaling pathway . Biol Direct 2024. ; 19 ( 1 ): 139 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Li J , Cao Y , Liu Y , et al . Multiomics profiling reveals the benefits of gamma-delta (γδ) T lymphocytes for improving the tumor microenvironment, immunotherapy efficacy and prognosis in cervical cancer . J Immunother Cancer 2024. ; 12 ( 1 ): e008355 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Qi YJ , Su GH , You C , et al . Radiomics in breast cancer: Current advances and future directions . Cell Rep Med 2024. ; 5 ( 9 ): 101719 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Huang H , Feng YL , Wan T , et al . Effectiveness of Sequential Chemoradiation vs Concurrent Chemoradiation or Radiation Alone in Adjuvant Treatment After Hysterectomy for Cervical Cancer: The STARS Phase 3 Randomized Clinical Trial . JAMA Oncol 2021. ; 7 ( 3 ): 361 – 369 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Li Y , Guo Y , Zhang K , et al . Cell Death Pathway Regulation by Functional Nanomedicines for Robust Antitumor Immunity . Adv Sci (Weinh) 2024. ; 11 ( 3 ): e2306580 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Yan X , Li S , Yan H , Yu C , Liu F . IONPs-Based Medical Imaging in Cancer Care: Moving Beyond Traditional Diagnosis and Therapeutic Assessment . Int J Nanomedicine 2023. ; 18 : 1741 – 1763 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Zhou TH , Zhou XX , Ni J , et al . CT whole lung radiomic nomogram: a potential biomarker for lung function evaluation and identification of COPD . Mil Med Res 2024. ; 11 ( 1 ): 14 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Meng L , Dong D , Chen X , et al . 2D and 3D CT Radiomic Features Performance Comparison in Characterization of Gastric Cancer: A Multi-Center Study . IEEE J Biomed Health Inform 2021. ; 25 ( 3 ): 755 – 763 . [DOI] [PubMed] [Google Scholar]
- 22. Ferretti M , Corino VDA . Integrating radiomic and 3D autoencoder-based features for Non-Small Cell Lung Cancer survival analysis . Comput Methods Programs Biomed 2025. ; 258 : 108496 . [DOI] [PubMed] [Google Scholar]
- 23. Prior O , Macarro C , Navarro V , et al . Identification of Precise 3D CT Radiomics for Habitat Computation by Machine Learning in Cancer . Radiol Artif Intell 2024. ; 6 ( 2 ): e230118 . [Published correction appears in Radiol Artif Intell 2024;6(3):e249001.] [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Barkley D , Moncada R , Pour M , et al . Cancer cell states recur across tumor types and form specific interactions with the tumor microenvironment . Nat Genet 2022. ; 54 ( 8 ): 1192 – 1201 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Lei G , Zhuang L , Gan B . The roles of ferroptosis in cancer: Tumor suppression, tumor microenvironment, and therapeutic interventions . Cancer Cell 2024. ; 42 ( 4 ): 513 – 534 . [DOI] [PubMed] [Google Scholar]
- 26. Sohn JH , Fields BKK . Radiomics and Deep Learning to Predict Pulmonary Nodule Metastasis at CT . Radiology 2024. ; 311 ( 1 ): e233356 . [DOI] [PubMed] [Google Scholar]
- 27. Liu RX , Li H , Towbin AJ , et al . Machine Learning Diagnosis of Small-Bowel Crohn Disease Using T2-Weighted MRI Radiomic and Clinical Data . AJR Am J Roentgenol 2024. ; 222 ( 1 ): e2329812 . [DOI] [PubMed] [Google Scholar]
- 28. Sun Z , Zhang T , Ahmad MU , et al . Comprehensive assessment of immune context and immunotherapy response via noninvasive imaging in gastric cancer . J Clin Invest 2024. ; 134 ( 6 ): e175834 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Liu Z , Luo C , Chen X , et al . Noninvasive prediction of perineural invasion in intrahepatic cholangiocarcinoma by clinicoradiological features and computed tomography radiomics based on interpretable machine learning: a multicenter cohort study . Int J Surg 2024. ; 110 ( 2 ): 1039 – 1051 . [Published correction appears in Int J Surg 2024;110(5):3138.] [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Jiang T , Xia Y , Li Y , et al . TRIM29 promotes antitumor immunity through enhancing IGF2BP1 ubiquitination and subsequent PD-L1 downregulation in gastric cancer . Cancer Lett 2024. ; 581 : 216510 . [DOI] [PubMed] [Google Scholar]
- 31. Huang W , Hu X , He X , et al . TRIM29 facilitates gemcitabine resistance via MEK/ERK pathway and is modulated by circRPS29/miR-770-5p axis in PDAC . Drug Resist Updat 2024. ; 74 : 101079 . [DOI] [PubMed] [Google Scholar]
- 32. Wang J , Lu W , Zhang J , et al . Loss of TRIM29 mitigates viral myocarditis by attenuating PERK-driven ER stress response in male mice . Nat Commun 2024. ; 15 ( 1 ): 3481 . [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.