Development and Validation of a Predictive Model Combining [18F] FDG PET/CT and the Suidan Score for Primary Optimal Cytoreduction in Advanced High-Grade Serous Ovarian Cancer: A Retrospective Cohort Study

Introduction

Epithelial ovarian cancer (EOC) is one of the 3 most common gynecological malignancies and is the leading cause of cancer-related mortality among gynecological cancers, posing a significant threat to women’s health and lives. Due to its insidious onset, nonspecific symptoms, and lack of effective early screening methods, approximately 75% of patients are diagnosed at an advanced stage, resulting in poor prognosis, with a 5-year relative survival rate of 50% [1–3]. The standard treatment for ovarian cancer involves platinum-based chemotherapy combined with surgical intervention, with cytoreductive surgery being the standard procedure [4]. Optimal cytoreductive surgery, defined as achieving no macroscopic residual disease (R0), significantly improves both progression-free survival (PFS) and overall survival (OS) [5–8]. However, extensive peritoneal dissemination and lymph node metastasis often preclude achieving optimal cytoreduction with primary surgery.

For patients deemed unlikely to achieve optimal cytoreduction initially, neoadjuvant chemotherapy (NACT) can be administered to reduce tumor burden, followed by interval debulking surgery (IDS) to attain R0 resection [9,10]. Consequently, whether performing primary cytoreductive surgery or interval debulking surgery, achieving R0 resection is essential for optimal patient prognosis [11,12]. A previous study demonstrated that in patients undergoing primary cytoreductive surgery, the median OS was 58 months for those achieving R0, compared to 35 months with residual disease of 1–5 mm, 29 months with 6–10 mm residual disease, and only 22 months when residual disease exceeded 1 cm [13]. Therefore, accurate preoperative assessment is critical for determining the optimal treatment strategy and assessing prognosis in patients with advanced ovarian cancer.

The primary assessment tools currently used for ovarian cancer patients are the Suidan score, based on contrast-enhanced CT imaging, and the Fagotti score, based on laparoscopic evaluation [14,15]. The Suidan score has lower sensitivity for detecting micrometastases and disseminated lesions, potentially leading to false-negative results, and its accuracy is highly dependent on radiologist expertise, while the Fagotti score involves an invasive procedure that necessitates additional anesthesia and surgical interventions. Laparoscopy also has inherent limitations, including the lack of direct tactile assessment through palpation, potential obstruction of visualization in certain anatomical spaces by fixed masses or carcinomatous adhesions, and technical constraints.

While previous studies have utilized PET/CT to analyze survival outcomes following cytoreductive surgery for recurrent ovarian cancer, the predictive utility of PET/CT for achieving optimal cytoreduction during primary cytoreductive surgery in advanced ovarian cancer remains inadequately defined [16,17]. Consequently, this study aimed to develop and internally validate a PET/CT-based radiomics model to predict the likelihood of attaining optimal cytoreduction in patients undergoing cytoreductive surgery for ovarian cancer. This radiomics model was further integrated with the Suidan score to explore the potential of a combined predictive framework.

The critical clinical decision between attempting primary debulking surgery (PDS) or opting for NACT) hinges on accurately predicting the likelihood of achieving complete (R0) resection. Current decision-making relies on a multidisciplinary evaluation incorporating imaging and laparoscopic scores, which have limitations as described above. This retrospective study generated hypotheses by investigating whether a PET/CT radiomics model, alone or in combination with established clinical criteria like the Suidan score, can provide additional, objective information to aid this complex preoperative assessment. This model is intended solely to inform and supplement multidisciplinary team discussion, not to replace clinical judgment. Furthermore, this work represents a hypothesis-generating exploration; any proposed model requires rigorous prospective validation before it can be considered for clinical use.

Material and Methods

STATEMENT OF ETHICS:

The study protocol was reviewed and approved by the Ethics Review Board of the Second Affiliated Hospital of Dalian Medical University (approval number: 2023-170), which covered the use of imaging, surgical, and retrospective clinical data for this study. The study strictly followed the Declaration of Helsinki and the national ethics guidelines for biomedical research. Standardized informed consent procedures were implemented for all studies involving human subjects. Participants signed written authorization documents after being fully informed of the purpose of the study and the scope of data use. Patients’ personal information was anonymized using double encryption technology to ensure that the research data were only used for the purpose of this research. Informed consent was obtained from all participants.

PATIENTS:

This study specifically focused on high-grade serous ovarian carcinoma (HGSOC), the most common and aggressive histologic subtype. Non-HGSC histologies (eg, low-grade serous, endometrioid, clear cell carcinoma) were excluded to ensure a biologically homogeneous cohort, as their distinct molecular profiles and clinical behaviors could confound the analysis of treatment response and imaging biomarkers.

The sample size was calculated a priori based on the requirement for developing a predictive model with an expected R0 resection rate of 75% and 6 predictor variables. With 10 events per variable (EPV), a minimum of 60 R0 resection events were required in the training set. With an anticipated event rate of 75%, the target sample size for the training set was calculated to be 80 patients (60/0.75). Given a planned 2: 1 ratio for partitioning the cohort, the total target sample size was determined to be 120 patients, resulting in a validation set of 40 patients.

From October 2020 to October 2023, we retrospectively enrolled 217 patients with advanced ovarian cancer who underwent primary cytoreductive surgery. All patients underwent both positron-emission tomography/computed tomography (PET/CT) and contrast-enhanced CT (CECT) examinations prior to surgery. The inclusion criteria were as follows:

Inclusion criteria:

Exclusion criteria:

After applying the exclusion criteria, 48 patients with early-stage disease and 59 patients with non-HGSOC pathology was excluded. A final cohort of 110 eligible patients was enrolled. This cohort was subsequently divided into a training set of 70 patients and a validation set of 40 patients, which aligns closely with the calculated sample size targets (Figure 1).

CYTOREDUCTIVE SURGERY AND OUTCOME ASSESSMENT:

All cytoreductive surgeries were performed by a dedicated team of gynecologic oncologists at our institution, each with over 10 years of experience in advanced ovarian cancer surgery, in accordance with our institutional protocol, with complete resection (R0) as the primary surgical objective. The surgical procedures encompassed total hysterectomy, bilateral salpingo-oophorectomy, omentectomy, and systematic pelvic and para-aortic lymphadenectomy, supplemented by required resection of any metastatic lesions involving the peritoneum, bowel, spleen, or diaphragm to achieve optimal cytoreduction.

The extent of residual disease was assessed intraoperatively by the operating surgeon and documented in the surgical report. Residual disease was categorized according to the following standardized definitions:

Optimal cytoreduction was strictly defined as R0 resection, consistent with contemporary oncological standards. This assessment, while performed by the operating surgeon, was based on the above standardized, institution-wide definitions to ensure consistency across cases.

PET/CT IMAGE ACQUISITION AND PROCESSING:

All patients underwent a standardized preoperative ¹⁸F-FDG PET/CT examination using a Philips Ingenuity TF scanner according to an institutional protocol. ¹⁸F-FDG was produced by a Sumitomo HM-10 cyclotron and synthesized via a PET (Beijing) chemistry module, with radiochemical purity >95%. Patients fasted for at least 6 hours and then received an intravenous injection of ¹⁸F-FDG at a dose of 3.7–5.55 MBq/kg. Following a standardized uptake period of 60 minutes in a dim, quiet room, PET/CT imaging from the skull base to the mid-thigh was performed. First, a low-dose CT scan was acquired for attenuation correction and anatomical localization (parameters: 120 kV, 90 mA, 0.75 s/rotation, 512×512 matrix). Immediately thereafter, PET emission data were collected over 8–10 bed positions at 1 minute per bed position, using a 144×144 matrix. PET images were reconstructed using an ordered-subsets expectation maximization (OSEM) algorithm with CT-based attenuation correction and were automatically fused with the CT images on the vendor’s workstation for analysis.

Three-dimensional regions of interest (3D-ROIs) encompassing primary tumors were manually delineated slice-by-slice on the fused PET-CT images by 2 board-certified nuclear medicine physicians. Representative examples of PET/CT images with delineated ROIs, illustrating different metabolic burdens (Figure 2). The fused images were generated by integrating PET data with low-dose CT and CT-based attenuation correction (CTAC) images. Following consensus-based delineation, the integrated workstation software (MedEx) automatically calculated key quantitative metabolic parameters from the PET data within the ROIs, including maximum standardized uptake value (SUVmax) and total lesion glycolysis (TLG). For CT assessment, the same 2 physicians independently evaluated the contrast-enhanced CT component according to the established Suidan scoring criteria; morphological assessment included recording the maximum tumor diameter (Size) of the largest metastatic lesion. These imaging biomarkers were integrated with clinically established prognostic variables: patient age at diagnosis, serum CA-125 levels (U/mL) obtained within 1 week of imaging, and FIGO staging (stage IIIC/IV) determined by surgical exploration. These imaging biomarkers, together with clinically established prognostic variables (patient age at diagnosis, serum CA-125 levels, and FIGO stage), were preprocessed and integrated as input features for the development of a random forest-based PET/CT radiomics prediction model (Figure 3).

SUIDAN SCORE EVALUATION:

The Suidan score was applied to preoperative contrast-enhanced CT (CECT) images for all patients to provide a standardized assessment of disease burden predictive of cytoreductive outcome. CECT images were independently reviewed and scored by 2 experienced radiologists specializing in gynecologic oncology imaging, both of whom were blinded to the patient’s surgical outcome (R0/R1 status), PET/CT findings, and final clinical data to eliminate assessment bias. Any initial discrepancy in scoring was resolved through a consensus review involving a third senior radiologist.

FEATURE SET AND PREPROCESSING:

Predictor variables included clinical parameters (age, CA125, FIGO stage), key PET/CT-derived metrics (SUVmax, TLG, tumor size), and the Suidan score. Given the limited number of clinically relevant and interpretable predictors (n <10), and to prioritize model transparency and avoid overfitting inherent to high-dimensional feature selection on small samples, no post hoc automated feature selection algorithm was used. All pre-specified features were retained for model construction.

MODEL DEVELOPMENT, HYPERPARAMETER TUNING, AND CLASS IMBALANCE:

Two distinct random forest models were developed using the randomForest package (v4.7-1.1) in R: 1) a PET-CT radiomics model using the baseline feature set, and 2) a combined model incorporating all baseline features plus the Suidan score.

Model hyperparameters were tuned via a grid search combined with 10-fold cross-validation on the training set only. The final configuration was: number of trees (ntree)=500; number of variables randomly sampled at each split (mtry)=√p (the square root of the total predictors, default); minimum node size (nodesize)=1. To address the class imbalance between R0 and R1 outcomes in the training set, the classwt parameter was set to be inversely proportional to the class frequencies.

VALIDATION AND PERFORMANCE ASSESSMENT STRATEGY:

To robustly assess model performance and mitigate the risk of overfitting – a critical concern in radiomics studies – we implemented a stringent temporal hold-out validation framework. The cohort was partitioned into a training set (n=70, comprising earlier consecutive cases) and a temporally separate validation set (n=40, comprising later consecutive cases). This design simulates a prospective clinical application scenario, ensuring that the validation data is completely unseen during model training, thereby providing an unbiased estimate of future performance and offering a stronger guard against optimistic bias compared to resampling methods (eg, repeated cross-validation) on the entire dataset. All performance metrics reported in this study (eg, AUC, sensitivity, specificity) were derived from a single evaluation on this held-out validation set. The internal cross-validation mentioned above was used exclusively for hyperparameter tuning within the training set and did not contribute to the final performance scores.

PERFORMANCE ASSESSMENT:

Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC), with its 95% confidence interval (CI) calculated via 2000 bootstrap replicates using the pROC package to quantify estimation uncertainty. Model calibration (agreement between predicted probabilities and observed outcomes) was assessed using the Brier score (a lower score indicates better calibration). The optimal probability threshold for classification was determined by maximizing the Youden index. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated accordingly. Variable importance was quantified through the percentage increase in mean squared error (%IncMSE).

LIMITATIONS:

This study did not include external validation in an independent cohort, which is a limitation. The primary objective was the internal development and validation of a novel predictive framework within a retrospective, single-center cohort, serving as a proof-of-concept for future prospective, multi-center studies.

STATISTICAL ANALYSIS:

All statistical analyses were performed using R software (version 4.3.0). Descriptive statistics are reported as medians with interquartile ranges(IQR) for non-normally distributed continuous variables, means with standard deviations for normally distributed continuous variables, and counts with percentages for categorical variables. Differences between patient groups (eg, R0 vs R1) were assessed using the Wilcoxon rank-sum test for continuous variables and the chi-square or Fisher’s exact test for categorical variables, as appropriate. All tests were two-sided, with a P value <0.05 considered statistically significant.

For the predictive models, performance metrics including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated using standard definitions. The area under the ROC curve (AUC) with its 95% confidence interval was derived via bootstrapping with 2000 replicates. Confidence intervals for sensitivity, specificity, PPV, NPV, and accuracy were calculated using the exact Clopper-Pearson method. Data visualization (including violin plots and bar charts) was performed using ggplot2 (version 3.4.4) and ggpubr (version 0.6.0) packages.

Results

CHARACTERISTICS OF PATIENTS:

From October 2020 to October 2023, 110 patients met the inclusion criteria. Baseline clinical characteristics of the training (n=70) and validation (n=40) cohorts are presented in Table 1. No statistically significant differences were observed between cohorts for key baseline variables. Median age was 58 years (range: 37–77) in the training cohort and 61.5 years (range: 43–75) in the validation cohort (P=0.083). FIGO stage distribution was comparable: 52.86% (37/70) vs 42.5% (17/40) for stage IIIC (P=0.296), and 47.14% (33/70) vs 57.5% (23/40) for stage IV. All tumors exhibited serous histology. Good nutritional status (mean serum albumin: 36.73±4.86 g/L training, 37.70±6.06 g/L validation; P=0.362) was prevalent. Optimal cytoreduction (R0: no macroscopic residual disease) was achieved in 72.86% (51/70) of training and 75.0% (30/40) of validation patients, while suboptimal cytoreduction (R1: macroscopic residual disease) occurred in 27.14% (19/70) and 25.0% (10/40), respectively (P=0.806). Median preoperative CA125 levels were 734.4 U/mL (training) and 619.65 U/mL (validation) (P=0.294). Significant inter-cohort difference was noted only for intraoperative blood loss (P=0.015) (Table 1).

VALIDATION OF THE PREDICTING MODELS:

Figure 4A, 4B display violin-boxplot hybrid figures comparing prediction scores of the radiomics model and Suidan score model across the training and validation cohorts. In the validation cohort, the PET-CT radiomics model achieved an AUC of 0.808 (95% CI, 0.665–0.951), while the Suidan score achieved an AUC of 0.773 (95% CI, 0.615–0.931) (Figure 4C). Feature importance analysis (Figure 4D) quantified the percentage increase in mean squared error (%IncMSE) for each variable in the radiomics model: tumor size (10.8%), CA125 (9.8%), SUVmax (6.9%), and TLG (5.8%). Age exhibited a negative value (−1.9% IncMSE).

COMBINED MODEL FOR PPTIMAL CYTOREDUCTION:

A combined model was developed by integrating the PET-CT radiomics features with the Suidan score. Figure 5A presents the distribution of its prediction scores. This model achieved an AUC of 0.836 (95% CI, 0.706–0.966) in the validation cohort (Figure 5B). The importance of features within the combined model is shown in Figure 5C. Comprehensive performance metrics for all models are presented in Table 2. For the combined model in the validation set, sensitivity was 81.8%, specificity was 83.9%, positive predictive value (PPV) was 64.3%, negative predictive value (NPV) was 92.9%, and accuracy was 83.3%.

Discussion

In this retrospective, single-center study, we developed and evaluated a predictive model that integrates FDG PET/CT radiomics with Suidan scores to identify suboptimal cytoreduction in patients with advanced, high-grade serous ovarian carcinoma (HGSOC). The combined model demonstrated promising discriminative capacity (AUC 0.836) compared to Suidan scoring alone (AUC 0.773), with notable improvement in specificity (83.9% vs 38.7%) and negative predictive value (92.9%).

The pursuit of complete cytoreduction (R0) is the cornerstone of primary treatment for advanced ovarian cancer, with contemporary evidence firmly establishing no macroscopic residual disease as the primary surgical goal [18–20]. Achieving R0, however, remains a significant challenge at initial surgery due to frequent presentation with extensive disease. While neoadjuvant chemotherapy (NACT) followed by interval surgery improved R0 rates compared to primary debulking surgery (PDS) in some trials [11,21,22], contemporary studies employing maximal surgical effort have demonstrated that PDS itself can achieve R0 rates exceeding 85% in specialized centers [23,24].The recently presented preliminary results of the TRUST trial [25], mandating high surgical standards, confirmed that PDS could achieve a markedly higher R0 rate (68%) with a progression-free survival benefit. This reaffirms that achieving R0 resection is the key prognostic determinant, regardless of the pathway (PDS or NACT), and underscores the critical importance of accurate preoperative identification of patients likely to achieve complete resection.

Several preoperative models aim to predict cytoreductive outcomes. The most established non-invasive tool is the Suidan score, a model refined from earlier versions [26] to incorporate clinical indicators and key CT imaging variables [14]. For direct surgical assessment, the Fagotti score, based on diagnostic laparoscopy, offers valuable insights but is invasive and can be technically limited in advanced disease [15]. Beyond anatomic assessment, FDG PET/CT has shown utility in predicting chemosensitivity and the feasibility of surgery in recurrent ovarian cancer [17,27–30]. However, its specific role in the preoperative stratification for primary cytoreductive surgery remains less clearly defined.

Building upon recent advances in radiomics [31,32], while addressing its standardization challenge in surgical prediction [33,34], our study contributes to the growing body of literature investigating PET/CT for predicting cytoreductive outcomes in ovarian cancer [35]. Recent efforts have focused on developing PET/CT-based predictive models. For instance, a multi-center study developed a model using metabolic tumor volume and lesion count, reporting an AUC of 0.771 upon external validation [36]. Another study directly compared a simpler PET/CT-based score (Boria score) with the Suidan score, finding comparable accuracy (AUC 0.738 vs 0.790) [37]. Our findings align with the premise that PET/CT provides valuable prognostic information, but our methodological approach differs significantly. Unlike models that primarily use volumetric or count-based PET parameters [37], we employed a radiomics framework to extract high-dimensional, quantitative data on tumor phenotype (eg, heterogeneity, intensity). Furthermore, while prior comparative studies have evaluated PET/CT models against established scores [36,37], our work is novel in its explicit integration of PET/CT radiomics features with the Suidan score itself, creating a combined model. This integrative approach may explain why our combined model achieved a higher AUC (0.836) in our internal validation cohort compared to the Suidan score alone (0.773) or the PET-only radiomics model (0.808). However, it is crucial to contextualize this performance. The reported sensitivity of PET/CT for detecting sub-centimeter miliary disease, a common cause of suboptimal resection, can be as low as 40% to 42% [38]. This inherent limitation of metabolic imaging underscores why a combined model that has both anatomic detail and metabolic heterogeneity may offer a more comprehensive assessment than either modality in isolation, potentially improving specificity for identifying truly inoperable disease and refining patient selection for primary surgery.

Several limitations and methodological considerations of this study should be acknowledged. First, and most importantly, our predictive models were developed and validated within a single-center, retrospective cohort, which inherently limits generalizability. We explicitly recognize the elevated risk of bias and irreproducibility associated with radiomics/ML studies. To mitigate these risks, we implemented key methodological safeguards: (1) using a small, pre-defined set of clinically interpretable features to avoid overfitting from data dredging; (2) ensuring blinded, independent assessment of the clinical imaging score (Suidan score) to minimize observer bias; and (3) using a rigorous temporal hold-out validation scheme to obtain a realistic, unbiased performance estimate. The absence of external validation means that the reported performance metrics are promising yet preliminary. Therefore, the model is not ready for clinical application, and prospective multi-center validation is an essential next step. Second, despite a priori sample size calculation, the cohort size remains modest for ML, warranting caution in interpretation. Third, fundamental differences exist between PET/CT and contrast-enhanced CT or laparoscopy in parameters and evaluation methods. While this highlights the complementary role of metabolic imaging, it also indicates that our model offers a distinct perspective. Given the retrospective design and the inherent limitations of the methodology, our conclusions are intentionally framed with caution and transparency. This study should be viewed as a robust internal validation and a necessary proof-of-concept, establishing a foundation for future validation in independent, prospective cohorts

Conclusions

The integration of FDG PET/CT radiomics with the Suidan score showed promising potential for improving preoperative assessment of cytoreductive outcomes in advanced HGSOC. However, the promising performance of this model must be validated through multi-center, prospective studies before any consideration can be given to its clinical implementation.