The optimal SVM model from multi-model comparison outperforms conventional criteria in preoperative prediction of ovarian cancer residual disease Authors: Dongdong Ye, Shaodan Lin, Zhuyan Shao, Yunyun Liu, Bin Wu, Chuying Huo, Dongdong Xu, Suen Wai Pang, Chunxian Huang, Jingyan Li, Jiachen Liu, Lijuan Wang, Tao Zhu, Huaiwu Lu Published: 2026 ABSTRACT AbstractObjectiveTo develop an interpretable artificial intelligence (AI)-based machine learning model integrating 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) radiomics, metabolic parameters, and clinical biomarkers for predicting residual disease after primary debulking surgery of ovarian cancer.MethodsThis multicenter retrospective study included 203 epithelial ovarian cancer patients (2017–2024), including 153 in the training cohort and 50 in the validation cohort. Radiomic features and metabolic parameters were extracted from preoperative PET/CT, and clinical variables were from medical records. After Synthetic Minority Over-sampling Technique (SMOTE) balancing and minimum Redundancy Maximum Relevance (mRMR)-based feature selection, seven models were trained using nested 5 × 10 cross-validation. Performance was assessed by Area Under the Curve (AUC), Brier score, and decision curve analysis (DCA), with comparison to the conventional criteria such as the Suidan criteria or Fagotti criteria. Survival outcomes were analyzed using Kaplan-Meier and Cox regression.ResultsThe SVM model (constructed using the support vector machine method) achieved superior discrimination (AUC 0.896 [95% CI 0.842–0.937] for training; 0.872 [0.792–0.932] for validation), significantly outperforming the Suidan criteria (AUC 0.696, P < 0.001; DeLong’s test). Key predictors included tumor spatial heterogeneity (Elongation, Least Axis Length) and metabolic activity (SUVmax), alongside CA-125 and albumin. The model demonstrated excellent calibration (Brier score 0.119) and clinical net benefit across thresholds (4–85%). Moreover, high-risk patients had increased mortality risk (P < 0.001).ConclusionThe PET/CT-based model improves preoperative prediction of residual disease in ovarian cancer, potentially reducing nontherapeutic surgeries. Automated lesion segmentation and multicenter validation are needed for clinical translation.Trial registration NCT06533709 (ClinicalTrials.gov, Registration Date: 2024-08-01).Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13048-026-02176-w.Keywords: Ovarian cancer, Radiomics, Surgical outcomes, Debulking surgery, Artificial intelligence Introduction Ovarian cancer (OC) contributes to a considerable number of cancer fatalities in women. Nearly half of OC cases present with distant metastasis, and 22% exhibit regional metastasis at diagnosis [1]. The ovarian cancer treatment strategy typically includes an initial surgical procedure, which is also called primary debulking surgery (PDS), and subsequent platinum-based chemotherapy. The removal of all visible tumor lesions during PDS is a key factor in improving survival rate [2–4]. Complete resection during PDS can be difficult in patients who are older, have a poor performance status, or present with unresectable disseminated disease. In these cases, neoadjuvant chemotherapy prior to the surgery would provide a relevant prognosis and help to reduce the complications of PDS [5, 6]. Therefore, preoperative evaluation to assess the feasibility of achieving complete resection is crucial for selecting the optimal treatment strategy. Several methods have been developed to assess tumor load in ovarian cancer. Fagotti et al. [7] proposed laparoscopic exploration to assess the lesions involved in pelvic, abdominal, and peritoneal cavities. Suidan et al. [8] developed a non-invasive method using clinical and radiological criteria to evaluate the risk of residual disease. The accuracy of the existing model in predicting residual disease indicated that approximately one-third of patients are not predicted accurately, which leads to less effective clinical decision-making. Other scoring systems, such as the Eisenkop, PCI, and Aletti scores, are also used to assess tumor burden or surgery complexity [9]. Laparoscopic exploration may not be accepted by patients due to its invasiveness and associated costs compared with simple radiological examination. Imaging-based assessment models face issues with reliability, accuracy, and inter-observer variability in the prediction. Therefore, the development of more effective, sensitive, and objective non-invasive evaluation methods is crucial. Artificial intelligence (AI) has been utilized across multiple stages of medical imaging, such as preprocessing, feature extraction, detection of regions of interest (ROI), and model development. The integration of AI has enhanced the accuracy and efficiency of medical predictions, overcoming bias introduced by subjective judgment. PET/CT, a combination of PET metabolic imaging and traditional CT scans, has been helpful in assessing tumor burden, clarify the extent of lesions, and predict therapeutic efficacy [10–14]. AI combined with PET/CT has shown promise in prognosis evaluation across various cancers [15–18]. The role of applying AI and PET/CT-derived features in predicting tumor residuals in patients with ovarian cancer undergoing surgery remains unclear. Our research aims to develop a predictive model using radiomic and metabolic data extracted from PET/CT scans to assess residual disease of the PDS, and to assist clinicians and radiologists in making decisions before initial treatment of the ovarian cancer. Method Patient selection This retrospective study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the ethics committees of Sun Yat-sen Memorial Hospital (SYSUMH, No: SYSKY-2024-503-01) and Zhejiang Cancer Hospital (ZJCH, No: IRB-2024-901). Inclusion criteria were as follows: (1) patients aged between 18 and 80 years; (2) histologically confirmed epithelial ovarian cancer; (3) underwent primary debulking surgery at either participating center; (4) availability of complete clinical, surgical, and pathological data; and (5) preoperative PET/CT performed within 4 weeks prior to the PDS. Exclusion criteria were: (1) patients who had received any disease-specific treatment (such as chemotherapy or radiotherapy) before PET/CT examination; (2) patients who did not undergo PDS, including fertility-sparing or palliative surgery; (3) patients presenting with infectious or inflammatory symptoms at the time of PET/CT examination; (4) patients co-suffering from other cancers; and (5) patients with severe comorbidities that could affect surgical outcomes. A total of 153 patients from SYSUMH and ZJCH were used for model development, while an additional independent cohort of 50 patients from ZJCH was included for external validation. Data collection Clinical data encompassed demographics, ASA status, preoperative labs, tumor markers, intraoperative details, surgical outcomes, recurrence, and survival data. Pathological staging followed 2014 International Federation of Gynecology and Obstetrics (FIGO) guidelines with World Health Organization (WHO)-defined histology. R0 (complete gross resection) indicates macroscopic tumor clearance after the surgery is completed, assessed by the surgeon. R1 (optimal resection) and R2 (suboptimal resection) denote residual disease of ≤ 1 cm and > 1 cm respectively. Specialized gynecological and general surgeons performed procedures with standardized protocols. Postoperative residual disease was documented in the medical record. Progression-free survival (PFS) spans from anti-tumor treatment initiation to progression; overall survival (OS) is defined as the period from diagnosis to death. PET/CT imaging was performed at two institutions using dedicated PET/CT systems. At SYSUMH, examinations were conducted using Siemens Biograph mCT or Biograph 64 scanners (Siemens Healthcare, Erlangen, Germany), whereas patients at ZJCH were scanned using a Discovery 710 PET/CT system (GE Healthcare, Milwaukee, WI, USA). All patients fasted for at least 6 h prior to imaging, and peripheral blood glucose levels were confirmed to be < 200 mg/dL before tracer administration. 18F-FDG was administered intravenously at a dose of approximately 3.7–5.55 MBq/kg. PET acquisition was initiated approximately 60 min after tracer injection. CT scans were performed for attenuation correction using standardized parameters (tube voltage 120 kV; slice thickness approximately 3–4 mm). Whole-body PET imaging covered the region from the skull base to the proximal thighs. PET images were reconstructed using iterative ordered-subset expectation maximization (OSEM) algorithms with CT-based attenuation correction. Vendor-specific reconstruction implementations were applied, including TrueX/TOF reconstruction for Siemens systems. Detailed PET/CT acquisition and reconstruction parameters for each center are summarized in Table S1. PET/CT images were obtained in anonymized DICOM format, and patients with incomplete or mismatched PET and CT images were excluded. Radiologic data were imported into 3D Slicer software (version 5.6.2) for image processing and feature extraction. Tumor segmentation was performed using a semi-automatic approach on PET images. The PETDICOMExtension plugin (version: 4bf67c6) of the 3D Slicer software was used to calculate the standardized uptake values (SUV). Segmentation included the primary ovarian tumor and all detectable metastatic lesions identified on PET/CT images. Volumetric regions of interest (ROIs) were initially generated using a fixed SUV threshold of 4.2. To reduce potential false-positive signals from physiological FDG uptake (e.g., brain parenchyma or urinary tract), two experienced radiologists with more than 10 years of experience in nuclear medicine imaging independently reviewed the PET/CT images and manually excluded non-pathological hypermetabolic areas with reference to the corresponding CT images. The remaining ROIs were kept and considered as tumor lesions, and radiomics features were extracted. The intra-class correlation coefficient (ICC) was evaluated for the consistency between two experts. PET-indiC (version: 5187e76) and SlicerRadiomics (version: 8426cdf) were two plugins integrated in the 3D Slicer software. The PET-indiC (v5187e76) automatically quantified metabolic parameters of the ROI, such as SUV maximum (SUVmax), metabolic tumor volume (MTV), and total lesion glycolysis (TLG). The SlicerRadiomics (v8426cdf) extracted IBSI-compliant radiomic features spanning four categories, including morphological (such as sphericity and elongation), intensity-based (such as kurtosis and skewness), textural (such as gray-level co-occurrence matrix contrast), and spatial heterogeneity metrics (such as gray-level size zone matrix zone variance). The tumor-to-liver ratio (TLR), comparing the SUV mean of primary ovarian lesions to the SUV mean of normal liver parenchyma, was used to normalize PET values [19]. Tumor SUVmax was measured from the PET slice showing the highest uptake within the primary lesion. The liver reference was obtained by placing a 3-cm circular ROI in the homogeneous parenchyma of the right hepatic lobe, avoiding hepatic margins, focal lesions, and adjacent structures to minimize partial-volume effects; measurements were repeated three times and averaged. The Suidan scores were derived from the CT component of the PET/CT scan by two experienced radiologists according to the original scoring criteria, while the Fagotti scores were assessed during diagnostic laparoscopy by gynecologic oncologists based on direct exploration of the pelvic and abdominal cavities. Data processing and model construction To tackle the issue of class imbalance in the training cohort, the Synthetic Minority Over-sampling Technique (SMOTE) was utilized. Feature selection was conducted using the Max-Relevance and Min-Redundancy (mRMR) algorithm [17], which optimizes feature relevance to residual disease (quantified via mutual information) while minimizing redundancy among features, resulting in the retention of the top 10 discriminative biomarkers (Fig. S2). A 5 × 10 nested cross-validation framework was employed to ensure unbiased model evaluation. In the outer loop, the dataset was stratified into five folds, with 80% (4 folds) allocated for model development and 20% (1 fold) reserved as a hold-out validation set in each iteration. Within the development subset, an inner 10-fold grid search systematically optimized algorithm-specific hyperparameters, iteratively refining model configurations to maximize generalization performance while mitigating overfitting. This dual-loop design rigorously decouples hyperparameter tuning from final validation, ensuring robust estimation of predictive accuracy across heterogeneous patient subgroups. Seven machine learning algorithms were evaluated: Least Absolute Shrinkage and Selection Operator Logistic Regression (LASSO, λ selected via 1SE rule), decision tree (DT, Gini impurity), k-nearest neighbors (kNN, k = 5), linear discriminant analysis (LDA), random forest (RF, 500 trees), support vector machine (SVM, RBF kernel), and extreme gradient boosting (XGBoost, max_depth = 3). Model performance was assessed by AUC, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). The model with the highest performance was chosen for additional analysis. Predictive performance was compared to existing R0 predicting criteria, including the Suidan criteria, PCI score, and Fagotti score, through DeLong’s method. Surgical complexity was ranked using the Aletti score. A Brier score based on the calibration curve was calculated, and decision curve analysis (DCA) was performed. The research also evaluated the model’s prediction for the survival. To enhance model interpretability, Shapley Additive Explanations (SHAP) [18], a game-theoretic approach, was applied. The overall process is illustrated in the flowchart (Fig. 1 and Fig. S1). Flow chart of the study. Ovarian cancer patients who underwent preoperative PET/CT imaging from two centers were retrospectively enrolled and divided into a training cohort and a validation cohort. Features were extracted from regions of interest (ROIs) on PET/CT images and medical records. After data preprocessing and feature selection, multiple machine learning models were constructed and evaluated through nested cross-validation. The best-performing model (SVM) was further evaluated using ROC analysis, calibration analysis, decision curve analysis and SHAP-based interpretability analysis. Finally, the clinical applicability of the model was assessed through comparison with existing clinical scoring systems and survival analyses. SMOTE, Synthetic Minority Oversampling Technique; mRMR, maximum relevance and minimum redundancy; LASSO, Least Absolute Shrinkage and Selection Operator Logistic Regression; KNN, k-nearest neighbors; LDA, Linear discriminant analysis; SVM, Support vector machine; XGB, Extreme gradient boosting; SMOTE, Synthetic Minority Over-sampling Technique; mRMR, Max-Relevance and Min-Redundancy; SHAP, Shapley Additive Explanations Statistical analysis The proportion of missing values for each variable is less than 5%, which is acceptable in predictive modeling. Continuous variables were compared using independent t-tests or Mann-Whitney U tests, while ordinal/categorical data were analyzed through chi-square or Kruskal-Wallis tests. Model performance was assessed through ROC curves (AUC). Survival outcomes (PFS/OS) were evaluated with log-rank tests and Cox regression and visualized through Kaplan-Meier curves. Analyses were performed in R (4.4.1 and 4.0.5), applying a significance threshold of p < 0.05 (two-sided tests). Results Patient characteristics Between July 2017 and February 2024, a total of 203 patients (103 from SYSUMH and 100 from ZJCH) with pathologically confirmed epithelial ovarian cancer who underwent PET/CT prior to primary debulking surgery were retrospectively included in the study. 153 patients were in the training group. Of these, 109 (71.2%) patients underwent R0 resection, 19 (12.4%) patients underwent R1 resection, and 25 (16.3%) patients underwent R2 resection. 50 patients were assigned into the validation group, in which 31 (62%) patients underwent R0 resection, 10 (20%) patients underwent R1 resection, and 9 (18%) patients underwent R2 resection. No missing value in the study cohort (Table 1). Table 1Clinical-pathologic characteristics of patientsR0 (n = 140)non-R0 (n = 63)P valueTraining cohort (n = 153)Validation cohort (n = 50)P valueGross Complete Resection0.22 R0109 (71.24)31 (62.00) non-R044 (28.76)19 (38.00)ASA0.1590.948 ≤II88 (62.86)33 (52.38)91 (59.48)30 (60.00) > II52 (37.14)30 (47.62)62 (40.52)20 (40.00)Histology0.7390.990 High Grade Serous Carcinoma85 (80.19)39 (82.98)82 (79.61)42 (84.00) Clear Cell Carcinoma9 (8.49)4 (8.51)9 (8.74)4 (8.00) Mucous Carcinoma4 (3.77)0 (0.00)3 (2.91)1 (2.00) Endometrioid Carcinoma4 (3.77)1 (2.13)4 (3.88)1 (2.00) Others4 (3.77)3 (6.38)5 (4.85)2 (4.00)FIGO Stage< 0.001< 0.001 I8 (5.71)0 (0.00)7 (4.58)1 (2.00) II11 (7.86)0 (0.00)11 (7.19)0 (0.00) III75 (53.57)25 (39.68)87 (56.86)13 (26.00) IV46 (32.86)38 (60.32)48 (31.37)36 (72.00)Progression0.098 No110 (78.57)39 (61.90)0.032109 (71.24)40 (80.00) Yes24 (17.14)17 (26.98)31 (20.26)10 (20.00) Unknown8 (5.84)5 (7.58)13 (8.50)0 (0.00)Status0.44 Alive122 (87.14)40 (63.49)< 0.001124 (81.05)38 (76.00) Dead18 (12.86)23 (36.51)29 (18.95)12 (24.00) Age (years)55.50 (50.00, 62.25)56.00 (51.00, 63.00)0.40156.00 (50.00, 63.00)55.00 (48.25, 61.00)0.414Time interval from PET/CT diagnosis to surgery (d)7 (5, 8)7 (5, 8)0.6307 (5, 8)7 (5, 8)0.947ALB (g/L)38.55 (34.90, 41.20)36.20 (30.60, 38.30)< 0.00137.10 (33.70, 40.40)39.25 (34.65, 42.02)0.041CA125 (U/ml)870.85 (286.25, 1761.82)1325.70 (554.85, 2714.45)0.007908.00 (296.00, 2112.00)1125.97 (476.77, 2066.96)0.462Long Run High Gray Level Emphasis2813.64 (503.93, 7754.23)6668.18 (1695.98, 11849.58)0.0114227.25 (648.47, 9009.65)3425.87 (744.79, 6875.01)0.516Major Axis Length175.05 (127.32, 288.61)316.57 (261.18, 383.10)< 0.001249.52 (151.06, 343.79)188.82 (124.28, 312.42)0.07Sphericity0.38 (0.30, 0.45)0.28 (0.22, 0.33)< 0.0010.34 (0.26, 0.43)0.36 (0.30, 0.45)0.18Maximum 2D Diameter Column384.87 (291.48, 542.60)569.68 (465.39, 703.37)< 0.001463.34 (335.82, 606.08)412.88 (293.05, 599.39)0.358SUVmax15.76 (10.15, 20.84)18.34 (14.47, 36.42)0.00216.22 (11.05, 23.19)16.46 (11.91, 22.94)0.707Elongation0.63 (0.44, 0.75)0.48 (0.35, 0.62)< 0.0010.57 (0.41, 0.71)0.61 (0.44, 0.73)0.648Cluster Tendency2.61 (1.34, 5.89)3.62 (1.97, 14.18)0.0082.80 (1.47, 7.32)2.75 (1.38, 6.13)0.569Least Axis Length74.52 (57.14, 92.77)101.94 (86.34, 117.07)< 0.00184.80 (65.43, 110.85)75.58 (57.32, 97.17)0.074Suidan score2 (1, 3)3 (2, 4)< 0.0012 (1, 3)2 (1, 3)0.958Fagotti score2 (0, 4)4 (2, 6)0.0014 (2, 6)2 (0.50, 6)0.405PCI score10 (7, 15)15 (9, 16)0.04213 (8, 16)10 (7, 14.75)0.074Aletti score5 (4, 8)6 (4, 9)0.6795 (4, 8)5 (4, 9)0.567Values are presented as frequency (percentage) or mean (standard error), or median (1st Quartile, 3st Quartile).ASA: American Society of Anesthesiologists; FIGO, International Federation of Gynecology and Obstetrics; HRD: homologous recombination deficiency; BMI: body mass index; ALB: albimin; CA-125, cancer antigen 125; TLR: tumor-to-liver ratio; PCI: Peritoneal Cancer Index Clinical-pathologic characteristics of patients Values are presented as frequency (percentage) or mean (standard error), or median (1st Quartile, 3st Quartile).ASA: American Society of Anesthesiologists; FIGO, International Federation of Gynecology and Obstetrics; HRD: homologous recombination deficiency; BMI: body mass index; ALB: albimin; CA-125, cancer antigen 125; TLR: tumor-to-liver ratio; PCI: Peritoneal Cancer Index A total of 118 features were incorporated, including 107 radiomics parameters, 6 metabolic parameters, TLR, and 4 clinical factors. Among the extracted features, 104 (91.2%) demonstrated good reproducibility (ICC ≥ 0.80), with a mean ICC of 0.94 (details in Table S2). The non-R0 group exhibited higher TLR (4.88 vs. 4.46, P = 0.003), Long Run High Gray Level Emphasis (2813.64 vs. 6668.18, P < 0.001), Major Axis Length (175.05 vs. 316.57, P < 0.001), Maximum 2D Diameter Column (384.87 vs. 569.68, P < 0.001), SUVmax (15.76 vs. 18.34, P = 0.002), Cluster Tendency (2.61 vs. 3.62, P = 0.008), lower Sphericity (0.38 vs. 0.28, P < 0.001), Elongation (0.63 vs. 0.48, P < 0.001). The non-R0 group also had higher score of the Suidan criteria (2 vs. 3, P < 0.001), Fagotti scores (2 vs. 4, P = 0.001) and PCI score (10 vs. 15, P = 0.042). No difference was tested in the ASA grade, histology, age and the time interval from PET/CT diagnosis to surgery. Patients in the training cohort and validation cohort had the same characteristics except for the FIGO stage and albumin (ALB). The validation cohort had higher rate of stage IV patients and higher ALB (37.10 vs. 39.25, P = 0.041). The median follow-up time was 24.0 months (IQR, 11.2–59.8 months). Data preprocessing and model construction The Synthetic Minority Oversampling Technique (SMOTE) was applied to balance the training cohort, resulting in 197 samples. The ratio of R0 to non-R0 resection patients was adjusted from 109:44 to 109:88. The distribution of the samples is uniformly distributed (Fig. 2A-B). Feature selection was performed using the Max-Relevance and Min-Redundancy (mRMR) algorithm, which identified the 10 most important features: Least Axis Length, Sphericity, Maximum 2D Diameter Column, SUV Max, ALB, Elongation, Cluster Tendency, Major Axis Length, Long Run High Gray Level Emphasis, and CA125 (Fig. 2C). Detailed information on all features is provided in Table S2. Data preprocessing and feature selection. A Distribution of samples before and after SMOTE balancing in the training cohort. Each point represents an individual patient plotted according to two representative features, and colors indicate the resection outcome (R0 vs. R1/R2). B Changes in the number of samples before and after SMOTE balancing. C Feature selection using the mRMR algorithm. The top ten features with the highest relevance to the outcome and lowest redundancy among variables were selected for model construction. SUV, standardized uptake value; CA125, cancer antigen 125; ALB, albumin Predictive performance of the SVM model Using the binary variable of whether there is a complete gross resection as the model task. Seven machine learning algorithms were employed to construct the predictive models. The final predictive performance of each model was summarized by the average AUC across all folds (Fig. 3A). The support vector machine (SVM) model exhibited the highest AUC, as well as relatively higher positive predictive value (PPV) and negative predictive value (NPV). Based on these results, the SVM model was selected for further analysis. Predictive performance of machine learning models. A Comparison of the predictive performance of seven machine learning algorithms evaluated using the mean area under the receiver operating characteristic curve (AUC) across nested cross-validation folds. B, C Receiver operating characteristic (ROC) curves comparing the performance of the SVM model with clinical scoring systems in the training cohort B and validation cohort (C). D Calibration curves evaluating the agreement between predicted probabilities and observed outcomes in the training and validation cohorts. The dashed diagonal line represents perfect calibration. E Decision curve analysis (DCA) illustrating the net clinical benefit of the SVM model compared with the “treat-all” and “treat-none” strategies across a range of threshold probabilities. NPV, Negative predictive value; PPV, Positive predictive value; PCI, Peritoneal carcinomatosis index The average AUC of the SVM model across all folds in the nested cross-validation was 0.896 (Fig. 3B, C), which was superior to the AUC values of the Suidan score (0.696, P < 0.0001), Fagotti score (0.658, P < 0.0001), PCI score (0.606, P < 0.0001), Aletti score (0.489, P < 0.0001), and CA125 (0.587, P < 0.0001). The calibration performance of the model was assessed using calibration plots and the Hosmer–Lemeshow goodness-of-fit test (Fig. 3D). The results indicated good agreement between predicted probabilities and observed outcomes, with non-significant Hosmer–Lemeshow test results in both the training cohort (P = 0.596) and validation cohort (P = 0.738). In addition, the model achieved Brier scores of 0.119 and 0.139 in the training and validation cohorts, respectively, indicating good overall accuracy of probabilistic predictions. Decision curve analysis (DCA) demonstrated that the SVM model outperformed both the “no treatment” and “universal treatment” strategies (Fig. 3E). Although the model demonstrates a statistical net benefit across a wide range of thresholds (4–85%), the practical impact of the model may be limited at very low thresholds where surgical treatment would generally be recommended regardless of model prediction. The model would therefore be most informative in intermediate or higher decision thresholds, where the choice between primary debulking surgery and neoadjuvant chemotherapy is less certain. Model interpretability SHAP-based model interpretability analysis was performed to address the “black-box” nature of machine learning models, which refers to their lack of transparency in decision-making. SHAP values quantify the contribution of each feature to the model prediction for an individual patient. A SHAP value close to zero indicated a minimal contribution of the feature, while values farther from zero indicated a greater contribution. Positive SHAP values indicate that the feature increases the predicted probability of residual disease, whereas negative values suggest a decrease in the predicted probability. The SHAP summary bar plot (Fig. 4B) illustrates the overall functionality of the model according to the mean absolute SHAP value. The most influential predictors were Elongation, Least Axis Length, Maximum 2D Diameter Column, ALB, Sphericity, Major Axis Length, SUVmax, CA125, Long Run High Gray Level Emphasis, and Cluster Tendency. The SHAP summary dot plot (Fig. 4A) visualizes the distribution of SHAP values for each feature across all patients, where each dot represents an individual case and the color reflects the relative magnitude of the feature value. The SHAP dependence plot (Fig. 4C) further illustrates how variations in a single feature influence the predicted probability of residual disease. Elongation, ALB, and Sphericity showed a negative correlation with the probability of residual disease, whereas Least Axis Length, Maximum 2D Diameter Column, SUVmax, and CA125 demonstrated positive associations. Interpretability analysis of the SVM model using SHAP values. A SHAP summary dot plot showing the distribution of SHAP values for each feature across all patients. Each dot represents an individual patient, and the color indicates the relative magnitude of the feature value. B SHAP feature importance ranking based on the mean absolute SHAP value. C SHAP dependence plot illustrating the relationship between a representative feature and the predicted probability of residual disease, showing how variation in the feature value influences the model output. SHAP, Shapley Additive Explanations; LRHGLE, LongRunHighGrayLevelEmphasis Extensional implementation of the model To enhance the clinical applicability of the model, we stratified patients into high-risk and low-risk groups based on the optimal cutoff value determined by the Youden index. We then compared the predictive performance of the model with the existing Suidan scoring system and the combined Suidan & Fagotti criteria for residual tumor detection. Overall, the model achieved a sensitivity of 0.756, specificity of 0.893, and accuracy of 0.856, outperforming the Suidan criteria (sensitivity 0.634, specificity 0.589, accuracy 0.601) and the combined Suidan & Fagotti criteria (sensitivity 0.120, specificity 0.924, accuracy 0.731). Positive and negative predictive values further reflected this trend, with the model yielding a PPV of 0.721 and NPV of 0.909, versus 0.361/0.815 and 0.333/0.768 for the Suidan and combined criteria, respectively (Fig. 5A-D). McNemar’s paired comparison confirmed that the differences were statistically significant, with the SVM model outperforming the Suidan criteria (P = 0.012) and the combined Suidan & Fagotti criteria (P < 0.0001). Clinical application and prognostic evaluation of the SVM model. A Comparison of predictive performance metrics (sensitivity, specificity, accuracy, positive predictive value, and negative predictive value) among the SVM model, Suidan criteria, and the combined Suidan–Fagotti criteria. (B–D) Confusion matrices illustrating prediction results for the SVM model (B), Suidan score (C), and the combined Suidan–Fagotti criteria (D). E Heatmap displaying details for the most important radiomic features identified by the model and selected clinical variables. F Kaplan–Meier analysis of overall survival stratified by model-predicted risk groups in the training cohort. G Univariate and multivariate Cox proportional hazards regression analyses evaluating predictors of overall survival. H Kaplan–Meier analysis of overall survival stratified by model-predicted risk groups in the validation cohort. TP, True Positive; TN, True Negative; FP, False Positive; FN, False Negative To further evaluate the prognostic implications of the model, patients were stratified into high-risk and low-risk groups based on the optimal cutoff value determined by the Youden index derived from the ROC curve. Patients with predicted probabilities above this threshold were classified as the high-risk group, whereas those below the threshold were assigned to the low-risk group. This risk stratification was subsequently used for survival analyses. Kaplan-Meier survival analysis revealed a significant association between the predicted model and overall survival in ovarian cancer (P < 0.0001, and P = 0.042 in the training cohort and validation cohort, Fig. 5F, H). Furthermore, multiple Cox proportional hazards regression analysis (Fig. 5G) confirmed that the model is an independent predictor of overall survival. Discussion This study developed a comprehensive model for the prediction of residual disease in PDS of ovarian cancer, based on the metabolic and radiomic data extracted from PET/CT scans and clinical variables. Among the seven machine learning models evaluated, the SVM model achieved the best performance, outperforming the commonly used Suidan criteria [20] and Fagotti score. The model emphasized the significance of key predictors, offering a valuable tool to aid clinical decision-making. SHAP values were used to analyze the contributions of 10 features to the final prediction. A positive SHAP value for CA125 suggests that increased CA125 levels are linked to a higher likelihood of residual disease, while a negative SHAP value for ALB implies that lower levels of albumin may be associated with worse postoperative recovery, thereby decreasing the probability of achieving R0 resection. The validation cohort contained a higher proportion of stage IV patients and slightly higher albumin levels compared with the training cohort. Such distributional differences may reflect real-world clinical heterogeneity between patient populations. Despite these differences, the model maintained stable predictive performance in the validation cohort, suggesting that the selected features captured clinically relevant information. The seven radiomics features obtained from the regions of interest (ROI) included in the model represent different tumor characteristics. Elongation was defined inversely to the actual elongation for computational efficiency and reflects the relationship between the two principal components of the ROI. Sphericity measures how close the ROI is to a perfect sphere. Major axis length, maximum 2D diameter column, and least axis length described the range and shape of lesions at different levels. The SHAP analysis showed elongation and sphericity were negatively correlated with residual disease, while broader or more irregularly shaped lesions (greater major axis length, maximum 2D diameter column, and least axis length) were positively correlated with residual disease. This aligns with clinical understanding that broader, irregularly shaped lesions, often found in the peritoneum or upper abdomen, are harder to resect. Cluster tendency measures how gray levels are spatially grouped and oriented in a specific direction, providing insights into the tissue’s structural organization and texture. The long run high gray level emphasis feature identifies the stretches of high gray values, providing insights into the lesions’ texture and heterogeneity. The two features imply that tumors in patients with incomplete resection disease may exhibit increased internal heterogeneity, which has been proved to be relevant to poorer clinical outcomes [21, 22]. Numerous previous studies have explored the role of PET/CT in predicting the likelihood of achieving complete cytoreduction in ovarian cancer. Csikos et al. [23] performed a meta-analysis and reported a summary RO of 0.781 for predicting primary cytoreductive outcomes. Shim et al. [24] developed a logistic regression model incorporating five PET/CT-derived features and achieved AUCs of 0.881 and 0.861 in the training and validation cohorts, respectively. However, this model relied on expert radiological interpretation, which may limit its reproducibility in routine clinical practice. Laios et al. [25] proposed a kNN-based model using clinical variables such as age, performance status, and surgical complexity, achieving an average predictive accuracy of 66%. Nevertheless, the interpretability of such machine-learning approaches remained limited. The present study employed a semi-automatic segmentation approach and systematically evaluated multiple machine-learning algorithms to construct the predictive model. Furthermore, SHAP-based interpretability analysis was incorporated to elucidate feature contributions, thereby improving the transparency and clinical interpretability of the model. Several studies have also highlighted the potential value of 18F-FDG PET/CT metabolic parameters in predicting surgical outcomes. Wang et al. [10] demonstrated that 18F-FDG PET/CT outperformed CT when applied to evaluate the Suidan criteria for predicting R0 resection. Similarly, Kim et al. [26] reported that metabolic tumor volume (MTV) and SUV-based parameters were significantly associated with residual disease during primary debulking surgery, with AUC values ranging from 0.653 to 0.749 for MTV and 0.680–0.737 for SUVmax. These findings support the potential utility of PET/CT-derived metabolic features in assessing tumor resectability. SUVmax ranked seventh in feature importance based on SHAP values. The relatively modest contribution of SUVmax may reflect known limitations of 18F-FDG PET in detecting small-volume peritoneal disease. Although FDG-PET/CT has demonstrated advantages in identifying lymph node metastases and peritoneal involvement, the detection of very small peritoneal implants remains challenging for current imaging modalities [27–29]. Small or hypometabolic lesions may fall below the spatial resolution or metabolic detection threshold of PET imaging, thereby weakening the association between SUV-derived metrics and the overall disease burden. Nevertheless, SUVmax may still capture aspects of tumor metabolic aggressiveness that are not fully reflected by morphological features [30–32]. The retention of SUVmax in the final model suggests that the metabolic activity of tumor lesions may provide complementary biological information beyond disease extent when estimating the likelihood of achieving complete cytoreduction. The use of fixed SUV thresholds to delineate metabolically active tumor regions is a common strategy in PET-based tumor segmentation [33, 34]. Dabir et al. [35] analyzed imaging data from 155 cancer patients and reported that the mean SUVmax of benign lesions was approximately 4.2, whereas malignant lesions demonstrated significantly higher uptake. ROC analysis further suggested that an SUVmax threshold of around 4.2 could effectively differentiate benign from malignant lesions with relatively high specificity. Similarly, Mansour et al. [34] reported that fixed SUV thresholds of 4.2 could achieve satisfactory lesion detection in PET imaging and capture a majority of metabolically active lesions. A SUV threshold around this range has been widely adopted in clinical and research settings as a practical reference for identifying metabolically active tumor regions. This study has several limitations. First, although the model demonstrated promising predictive performance, the sample size was relatively limited. Future studies incorporating larger multi-center datasets will be necessary to further confirm the generalizability of the proposed model. Second, the radiomic features were extracted from PET/CT images acquired using different scanners, and no formal cross-center harmonization procedures were performed. Multicenter harmonization strategies may help reduce potential imaging variability in future studies. Third, this study used a fixed SUV threshold (SUV > 4.2) for semi-automatic lesion segmentation. It may not fully capture the imaging characteristics of all tumor lesions. In particular, small-volume or low-metabolic peritoneal disease may be underrepresented. Future studies incorporating adaptive segmentation strategies or multimodal imaging approaches may further improve lesion characterization. Fourth, although the model was further evaluated using survival analysis, the survival data are still maturing. Longer follow-up will be required to fully assess the prognostic implications of the model. Finally, some variables in the workflow still rely on manual verification and dedicated image-processing software, which may limit immediate clinical implementation. Integration with automated segmentation algorithms and user-friendly software platforms may enhance the clinical applicability of this approach. Conclusion In conclusion, the study developed and validated an interpretable SVM-based machine learning model integrating preoperative clinical variables, metabolic parameters, and radiomic signatures from [18] F-FDG PET/CT, which demonstrated superior predictive accuracy for complete cytoreduction and prognostic stratification in ovarian cancer. The incorporation of tumor spatial heterogeneity metrics (e.g., lesion elongation) and metabolic activity patterns (SUVmax) provides a novel paradigm for preoperative decision-making. Future implementation of automated lesion segmentation and real-time R0 prediction algorithms could enhance clinical utility, potentially reducing futile laparotomies and guiding personalized therapy selection. Supplementary Information Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Acknowledgements The authors would like to thank all the patients and clinical staff at the participating centers for their contribution to this study. Abbreviations Artificial Intelligence Albumin American Society of Anesthesiologists Area under the curve Confident interval Cross-validation Decision curve analysis Digital Imaging and Communications in Medicine Decision tree Fluorodeoxyglucose International Federation of Gynecology and Obstetrics False Negative False Positive Interval debulking surgery k-nearest neighbors Least Absolute Shrinkage and Selection Operator Logistic Regression Linear discriminant analysis Machine learning Max-Relevance and Min-Redundancy Metabolic tumor volume Neoadjuvant chemotherapy Negative predictive value Ovarian cancer Overall survival Peritoneal carcinomatosis index Primary debulking surgery Positron Emission Tomography/Computed Tomography Progression-free survival Positive predictive value Random forest Receiver operating characteristic Regions of interest Shapley Additive Explanations Synthetic Minority Over-sampling Technique Standardized uptake values Support vector machine Sun Yat-sen Memorial Hospital Tumor-to-liver ratio True Negative True Positive World Health Organization Extreme gradient boosting Zhejiang Cancer Hospital Authors’ contributions Dongdong Ye : Writing-original draft, Writing-review & editing, Visualization, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Shaodan Lin and Zhuyan Shao : Writing-original draft, Writing-review & editing, Validation, Methodology, Formal analysis, Data curation. Yunyun Liu : Writing-review & editing, Validation, Supervision, Resources, Data curation. Bin Wu and Chuying Huo : Validation, Supervision, Resources, Data curation. Dongdong Xu and Suen Wai Pang : Validation, Methodology, Data curation. Chunxian Huang : Methodology, Supervision. Jingyan Li and Jiachen Liu :Resources, Data curation. Lijuan Wang and Tao Zhu : Writing-review & editing, Supervision, Project administration, Methodology, Conceptualization. Huaiwu Lu : Writing-review & editing, Validation, Supervision, Resources, Project administration, Methodology, Funding acquisition, Data curation, Conceptualization. Funding 1. National Natural Science Foundation of China (82573397, 82303860, 82503446); 2. Guangdong Basic and Applied Basic Research Foundation (2026A1515012551, 2024A1515013255); 3. Guangzhou Clinical High-Tech, Major, and Characteristic Technology Project (2026P-TS003); 4. Sun Yat-Sen Memorial Hospital Clinical Research 5010 Program (SYS-5010-202605); 5. Beijing Kanghe Public Welfare Foundation (KHKYP016); 6. China Postdoctoral Science Foundation (2025M782390); 7. Sun Yat-sen Pilot Scientific Research Fund (YXQH202606). Data availability The data used in this research are not publicly available due to the PET/CT images and clinical data contain information that could compromise the privacy of research participants. the privacy rights of human subjects have been observed and that informed consent was obtained for experimentation with human subjects. Declarations Ethical approval and consent to participate This retrospective study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the ethics committees of Sun Yat–sen Memorial Hospital (SYSUMH, approval number: SYSKY–2024–503–01, date: July 8, 2024) and Zhejiang Cancer Hospital (ZJCH, approval number: IRB–2024–901, date: October 6, 2024). Due to the retrospective nature of the study and the use of anonymized data, the requirement for written informed consent was waived by the Ethics Committees. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Dongdong Ye, Shaodan Lin, Zhuyan Shao and Yunyun Liu contributed equally to this work. Contributor Information Lijuan Wang, Email: wanglij9@mail.sysu.edu.cn. Tao Zhu, Email: zhutao@zjcc.org.cn. Huaiwu Lu, Email: luhuaiwu@mail.sysu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Data Availability Statement The data used in this research are not publicly available due to the PET/CT images and clinical data contain information that could compromise the privacy of research participants. the privacy rights of human subjects have been observed and that informed consent was obtained for experimentation with human subjects.