A radiomics-based interpretable model integrating delayed-phase CT and clinical features for predicting the pathological grade of appendiceal pseudomyxoma peritonei

Abstract

Objective

This study aimed to develop an interpretable machine learning model integrating delayed-phase contrast-enhanced CT radiomics with clinical features for noninvasive prediction of pathological grading in appendiceal pseudomyxoma peritonei (PMP), using Shapley Additive Explanations (SHAP) for model interpretation.

Materials and methods

This retrospective study analyzed 158 pathologically confirmed PMP cases (85 low-grade, 73 high-grade) from January 4, 2015 to April 30, 2024. Comprehensive clinical data including demographic characteristics, serum tumor markers (CEA, CA19-9, CA125, D-dimer, CA-724, CA-242), and CT-peritoneal cancer index (CT-PCI) were collected. Radiomics features were extracted from preoperative contrast-enhanced CT scans using standardized protocols. After rigorous feature selection and five-fold cross-validation, we developed three predictive models: clinical-only, radiomics-only, and a combined clinical-radiomics model using logistic regression. Model performance was evaluated through ROC analysis (AUC), Delong test, decision curve analysis (DCA), and Brier score, with SHAP values providing interpretability.

Results

The combined model demonstrated superior performance, achieving AUCs of 0.91 (95%CI:0.86–0.95) and 0.88 (95%CI:0.82–0.93) in training and testing sets respectively, significantly outperforming standalone models (P < 0.05). DCA confirmed greater clinical utility across most threshold probabilities, with favorable Brier scores (training:0.124; testing:0.142) indicating excellent calibration. SHAP analysis identified the top predictive features: wavelet-LHH_glcm_InverseVariance (radiomics), original_shape_Elongation (radiomics), and CA-199 (clinical).

Conclusion

Our SHAP-interpretable combined model provides an accurate, noninvasive tool for PMP grading, facilitating personalized treatment decisions. The integration of radiomics and clinical data demonstrates superior predictive performance compared to conventional approaches, with potential to improve patient outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-025-01843-6.

Introduction

Pseudomyxoma peritonei (PMP) is a rare and clinically heterogeneous condition, with an estimated incidence of 1–3 cases per million individuals annually. Over 90% of PMP cases originate from ruptured appendiceal mucinous neoplasms (AMNs), which exhibit a wide spectrum of biological behavior, ranging from indolent mucinous ascites to aggressive peritoneal dissemination with poor prognosis [1, 2]. The 5th edition of WHO and the 8th and 9th editions of AJCC advocate a three-stage grading system for appendicular mucinous tumor-associated peritoneal disease: Grade 1(G1), Grade 2(G2), and Grade 3(G3), with significant prognostic implications [3, 2]. Notably, patients with low-grade PMP (G1-G2) have a 5-year survival rate of 63%, whereas those with high-grade PMP (G3, including signet-ring cell morphology) experience a drastically reduced survival of only 23% [4]. The grading of appendiceal and peritoneal diseases are the same in most cases, inconsistent grading can only be observed in rare cases [5, 6].

According to the Peritoneal Surface Oncology Group International (PSOGI) consensus, the standard of care for the treatment of low-grade appendiceal PMP is Cytoreductive Surgery (CRS) followed by Hyperthermic intraperitoneal chemotherapy (HIPEC), which has demonstrated favorable long-term outcomes [7]. This method was first popularized by Sugarbaker in the 1990s [8] and is now widely adopted. In contrast, the management of high-grade PMP remains controversial, with systemic chemotherapy often preferred due to the limited efficacy of CRS/HIPEC and the high likelihood of incomplete resection (10–33% of cases) [9–13]. Given these stark differences in treatment response and survival, accurate preoperative pathological grading is critical for optimizing therapeutic strategies and improving patient outcomes [14]. However, current diagnostic methods rely heavily on invasive surgical biopsies, which carry inherent risks and delays in clinical decision-making.

Contrast-enhanced computed tomography (CT) is the primary imaging modality for PMP assessment, but conventional radiological evaluation is limited to qualitative assessments of tumor morphology and distribution. Delayed-phase contrast-enhanced CT was prioritized for its superior ability to differentiate fibrovascular tumor stroma from hypodense mucin, enhancing boundary delineation and correlating with tumor cellular density in high-grade PMP.Delayed-phase CT, unlike arterial or venous phases, provides optimal visualization of PMP lesions due to progressive contrast retention in tumor stroma, facilitating accurate texture analysis and morphological assessment-key advantages for radiomics-based grading.

Radiomics-a high-throughput extraction of quantitative imaging features—has emerged as a powerful tool for decoding tumor heterogeneity and predicting pathological characteristics noninvasively [15]. Only a few studies have analyzed the pathological grading of PMP. Recent studies [16, 17] have explored radiomics-based PMP grading, but most were constrained by small sample sizes, particularly for high-grade cases, and lacked robust interpretability. Zhou N et al. identified the pathological grade of PMP by contusing the plain CT scan omentula images of 88 patients combined with laboratory indicators, among which 28 cases were high-grade PMP. Pengtao Sun et al. analyzed CT signs of 75 patients with PMP, among which only 18 patients had high-grade PMP. Moreover, existing models predominantly rely on plain CT or clinical biomarkers alone, overlooking the potential synergistic value of combined radiomic-clinical analysis.To address these limitations, we developed an interpretable machine learning model integrating delayed-phase contrast-enhanced CT radiomics with key clinical biomarkers. Our approach leverages Shapley Additive Explanations (SHAP) to elucidate feature contributions, enhancing clinical translatability [18]. This study represents the largest radiomics investigation of PMP to date, with particular emphasis on high-grade cases, and introduces novel delayed-phase CT features to improve diagnostic accuracy. In our study, all patients underwent surgery without prior knowledge of their pathological grading, as the definitive diagnosis could only be obtained through histopathological examination post-surgery. The primary aim of our research is to develop a non-invasive method to assess PMP pathological grading preoperatively, thereby informing treatment decisions based on the predicted grade.

Materials and methods

Data collection

• This retrospective study was conducted at a tertiary referral center specializing in peritoneal surface malignancies. Between January 4, 2015 to April 30, 2024, we identified 200 consecutive patients who underwent surgical intervention for appendiceal pseudomyxoma peritonei (PMP). After applying stringent selection criteria, 158 patients formed our final study cohort (Fig. 1).The inclusion criteria were as follows: (1)Patients with appendiceal PMP confirmed by surgery and pathology with complete pathological results; (2)Enhanced CT scan was performed within 2 weeks before surgery; (3)CT enhancement images and clinical data were complete and clear; (4)The pathological grade of appendiceal lesions was consistent with PMP. Exclusion criteria were as follows: (1)patients who had previously received chemoradiotherapy or surgery; (2)Poor CT image quality or poor lesion display; (3)Patients with other malignant tumor diseases such as digestive tract or reproductive system; (4)Incomplete clinical or pathological data. The final cohort consisted of 158 patients with a median age of 58 years (range: 21–81), including 85 cases (53.8%) of low-grade PMP and 73 cases (46.2%) of high-grade PMP. Our study’s 158-patient cohort (including 73 high-grade cases) compares favorably with existing literature.While our Events Per Variable (EPV) ratio of 7.3 (73 high-grade cases/10 predictors) falls slightly below the ideal threshold of 10, this remains methodologically acceptable in rare disease research where EPV > 5 is often considered adequate.

• Pathological grading was performed for both peritoneal and appendiceal lesions. All cases selected in our study demonstrated consistent grading between the primary appendiceal tumor and peritoneal deposits. Grading followed the 5th edition of WHO and the 8th and 9th editions of AJCC three-tier system (G1 Well-differentiated, G2 Moderately differentiated, G3 Poorly differentiated), with G1-G2 classified as low-grade and G3 (including signet-ring cell morphology) as high-grade. Two pathologists with 10 years’ experience independently reviewed all cases.This study was conducted and reported in accordance with the CLEAR checklist for radiomics research.This study was designed, conducted, and reported in accordance with the METRICS checklist to ensure methodological rigor and transparency.The Aerospace Center Hospital Ethics Committee approved the retrospective study.

Fig. 1.

Patients selection flow chart

Clinical data acquisition and peritoneal disease assessment

A comprehensive panel of clinical parameters was prospectively collected for all enrolled patients, encompassing demographic characteristics (age and sex), routine hematological markers (white blood cell count, WBC [×10⁹/L]), and established serum tumor biomarkers including CA-125 (U/mL), CA19-9 (U/mL), CEA (ng/mL), D-dimer (µg/L), CA-724 (U/mL), and CA-242 (kU/L). Peritoneal disease burden was quantitatively assessed using the validated Sugarbaker Peritoneal Cancer Index (PCI) scoring system [19], which methodically evaluates tumor distribution across 13 distinct anatomic compartments (Fig. 2): nine abdominopelvic regions (R0-R8) and four small bowel segments (R9-R12). Two board-certified abdominal radiologists (with 7 and 10 years of specialized experience, respectively) independently scored each region according to standardized criteria: 0 (no macroscopic tumor), 1 (maximum tumor diameter < 0.5 cm), 2 (0.5–5.0 cm), or 3 (> 5.0 cm or confluent lesions), with a maximum cumulative score of 39. Discrepancies were resolved through consensus review with a third senior radiologist (15 years’ experience), achieving excellent inter-rater reliability (intraclass correlation coefficient = 0.91, 95% CI 0.88–0.93). This rigorous scoring protocol, performed on dedicated PACS workstations with optimized viewing parameters, ensured precise quantification of peritoneal disease extent for subsequent radiomics analysis.

Fig. 2.

The peritoneal cancer index (PCI) illustration drawn by Harmon and Sugarbaker

CT Imaging Protocol

All patients underwent standardized multiphase contrast-enhanced abdominopelvic CT examinations within 7 days prior to surgery using either a GE LightSpeed VCT or Siemens SOMATOM Force dual-source CT scanner. Following an 8-hour fasting period and oral administration of 1500 mL iodinated contrast (500 mL at 21:00 the preceding day, 500 mL 2 h pre-scan, and 500 mL immediately before imaging), patients were positioned supine with arms elevated. Scanning extended from the diaphragmatic dome to the iliac crest, employing a standardized protocol: 120 kVp, 200–250 mA, 5 mm slice thickness, and 1.375:1 pitch.

After initial unenhanced acquisition, multiphase contrast-enhanced imaging was performed following intravenous administration of 75–100 mL non-ionic contrast medium (ioverol 350 or iopamidol 370) via antecubital vein injection at 2.5-3.0 mL/s using a dual-chamber power injector. Scan timing was optimized for arterial (25–30 s), portal venous (65–70 s), and delayed (180s) phases. All studies included isotropic 1.25 mm reconstructions in axial, coronal, and sagittal planes, reconstructed with both standard and sharp kernels for comprehensive evaluation.

The choice of delayed-phase imaging was prioritized due to its superior ability to differentiate fibrovascular tumor stroma from hypodense mucin, enhancing boundary delineation and correlating with tumor cellular density in high-grade PMP. This phase provides optimal visualization of PMP lesions through progressive contrast retention in the tumor stroma, facilitating accurate texture analysis and morphological assessment. Multiphasic contrast-enhanced CT remains irreplaceable by single-phase imaging for comprehensive PMP evaluation, as the sequential phases provide essential information about distant metastases while enabling precise assessment of vascular anatomy and potential vascular anomalies.

Image analysis and radiomics feature extraction

All delayed-phase contrast-enhanced CT images in DICOM format were retrieved from our institutional PACS system and processed using both the Deepwise Research Platform (v2.2) and the open-source PyRadiomics library (version 3.0.1, https://github.com/Radiomics/pyradiomics) to ensure comprehensive and reproducible radiomics analysis.Resampling was performed to normalize all images to an isotropic voxel size of 1.0 × 1.0 × 1.0 mm³ using cubic B-spline interpolation, preserving spatial resolution while harmonizing voxel dimensions.Subsequently, discretization of gray-level intensities was applied to reduce noise and enhance texture feature robustness. The pixel intensity values were discretized into 64 bins using a fixed bin width method.A rigorous multi-step segmentation protocol was implemented: first, a radiologist with 5 years of abdominal imaging experience (Reader 1) performed semi-automated volumetric segmentation of appendiceal lesions using a layer-by-layer approach with manual correction when necessary. These segmentations were then independently reviewed by a senior abdominal radiologist (Reader 2, 10 years’ experience), with consensus reached through joint evaluation of discordant cases. To assess inter-observer reliability, 50 randomly selected cases underwent repeat segmentation by Reader 2, demonstrating excellent reproducibility (intraclass correlation coefficient = 0.87, 95% CI0.82-0.91).

Prior to feature extraction, image intensity normalization was performed using Z-score standardization to minimize scanner-related variability. Our comprehensive radiomics pipeline incorporated ten advanced image preprocessing techniques: native resolution analysis, Laplacian of Gaussian filtering, wavelet decomposition, gradient computation, 2D/3D Local Binary Patterns, along with nonlinear transformations (squaring, square root, logarithmic, and exponential). Feature extraction focused on three key categories: (1) first-order statistics, (2) morphological descriptors, and (3) advanced texture features including Gray Level Co-occurrence Matrix (13 directions), Gray Level Run Length Matrix, Gray Level Size Zone Matrix, Neighborhood Gray-Tone Difference Matrix, and Gray Level Dependence Matrix.

To optimize feature selection, we implemented a rigorous two-stage dimensionality reduction process: first removing redundant features (pairwise correlation coefficient > 0.8) while prioritizing those with stronger clinical associations, followed by application of six complementary selection methods (F-test, Pearson correlation, mutual information, L1 regularization, tree-based, and recursive feature elimination). A variety of feature screening methods were selected for simultaneous screening to retain common features. This multi-modal approach ensured retention of the most biologically and clinically relevant features while maintaining model generalizability, as detailed in our radiomics workflow (Fig. 3). All image processing and analysis steps adhered to the Image Biomarker Standardization Initiative (IBSI) guidelines to ensure methodological reproducibility.

Fig. 3.

The workflow of the radiomics process

Model building and interpretability analysis

The machine learning model was implemented using scikit-learn (version 0.22) with the following configuration: L2 regularization with a penalty strength of 12, an inverse regularization strength (C) of 1 (within a tunable range of 0.001-10), fit intercept enabled (True), and class weight set to balanced. To develop an effective and robust predictive model, We employed five-fold cross-validation with stratified sampling to ensure balanced class distribution in each fold. The dataset was randomly partitioned into five equal subsets (seed = 5 for reproducibility), with four folds used for training and one for validation in each iteration. Following feature selection, a logistic regression classifier was utilized to construct a radiomics-based prediction model for the pathological grading of appendiceal pseudomyxoma peritonei (PMP). The multivariate logistic regression model was constructed using the following formula: logit (P) = ln [probability (event)/probability (no event)] = a + b1 * variable1 + b2 * variable2…, a: constant, b1, 2: coefficient. Probability was calculated as follows: probability = 1/(1 + e^{− logit model)}.Logistic regression, a probabilistic classification model, employs a parameterized logistic function to estimate conditional probabilities, thereby enabling binary outcome prediction.

Clinically significant indicators, identified through univariate and multivariate analyses, were incorporated into a separate logistic regression model to establish a clinical prediction model. Subsequently, to enhance predictive performance, we integrated the optimal radiomics features with clinical variables, constructing a combined clinical-imaging model using logistic regression.

Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, with the area under the curve (AUC) quantifying predictive accuracy on both training and testing datasets. Statistical comparison of AUC values among the clinical, radiomics, and combined models was conducted using DeLong’s test. Decision curve analysis (DCA) was further applied to assess the net clinical benefit across varying threshold probabilities, providing insights into the models’ clinical utility. Additionally, the Brier score was employed to evaluate calibration accuracy, where scores ranging from 0 to 0.25 indicate strong predictive performance. A score closer to 0 reflects higher agreement between predicted probabilities and observed outcomes, signifying superior model performance.

Given the multifactorial nature of appendiceal PMP pathological grading, we performed interpretability analysis using the SHAP (SHapley Additive exPlanations) method. This data-driven approach quantifies variable importance by assessing each feature’s contribution to model predictions, thereby elucidating the relative influence of clinical and radiomics features on pathological grade determination.

Statistical analysis

Continuous variables following a normal distribution were presented as mean ± standard deviation and compared using independent sample t-tests. For non-normally distributed continuous variables, data were expressed as median (interquartile range) [M (Q1, Q3)], and group comparisons were performed using the Wilcoxon rank-sum test. Categorical variables were summarized as frequencies and percentages, with intergroup differences assessed via the chi-square test.

All statistical analyses were conducted using IBM SPSS Statistics 25 and R software (version 4.3.0). The intraclass correlation coefficient (ICC) was computed to evaluate intra- and interobserver agreement. A two-tailed p-value < 0.05 was considered statistically significant.

Results

Patient characteristics

A total of 200 patients were retrospectively reviewed, of whom 42 were excluded due to prior chemoradiation or surgical treatment (n = 15), suboptimal CT image quality or poor lesion visualization (n = 10), concurrent tumors at other sites (n = 10), or incomplete clinical/pathological data (n = 7). The final cohort comprised 158 patients with appendiceal pseudomyxoma peritonei (PMP), including 75 males (47.5%) and 83 females (52.5%), aged 21–81 years (mean: 58.22 ± 10.15). Based on pathological grading, 73 cases (46.2%) were classified as high-grade PMP, while 85 (53.8%) were low-grade PMP. Clinical parameters analyzed included age, gender, tumor markers (CEA, CA199, CA125, CA-724, CA-242), D-dimer levels, and preoperative CT-based peritoneal cancer index (CT-PCI), as detailed in Table 1.

Table 1.

Baseline characteristics of patients between high-grade PMP and low-grade PMP

	Low-grade PMP (n = 85)	High-grade PMP (n = 73)	P value	Multivariate odds ratio (95% CI)	P value
Gender
Female	43 (50.6%)	40 (54.8%)	0.598
Male	42(49.4%)	33 (45.2%)
Age (years)	57.39 ± 9.65	59.19 ± 10.70	0.087
WBC (10^9/L)	6.19 ± 2.39	6.68 ± 3.08	0.484
CA125 (U/ml)	82.49 ± 135.26	207.34 ± 269.78	0.001*	1.082(1.036,1.130)	0.001*
CA19-9 (U/ml)	173.68 ± 435.09	1323.97 ± 2852.07	0.001*	1.059(1.017,1.102)	0.001*
CEA (ng/ml)	319.20 ± 1276.23	1768.44 ± 3570.40	0.001*	1.007(0.971–1.045)	0.017*
D- Dimer (ug/l)	661.86 ± 979.66	993.89 ± 1123.50	0.002*
CA-724 (U/ml)	85.55 ± 96.35	154.26 ± 144.64	0.002*
CA-242 (kU/L)	151.86 ± 224.15	153.78 ± 130.51	0.111
CT-PCI (n)	23.20 ± 11.13	29.88 ± 8.93	0.001*

*Signifcant P values, P<0.05

Model building

Clinical model

Univariate logistic regression identified six significant predictors (P < 0.05): CA 19 − 9, CEA, CA-125, D-dimer, CA 72 − 4, and CT peritoneal cancer index. Subsequent multivariate logistic regression with forward selection retained three serum markers (CA 19 − 9, CEA, and CA-125) for the final clinical model. The model demonstrated strong discrimination in the training cohort (AUC, 0.82; accuracy, 73%; sensitivity, 70%; specificity, 76%), with maintained performance in the testing cohort (AUC, 0.78; accuracy, 73%; sensitivity, 70%; specificity, 76%).The final results presented are from the best-performing fold in the 5-fold cross-validation.The differences between folds were minimal and statistical tests showed no significant differences (p > 0.05).

Radiomics model

The tumor segmentation masks were meticulously delineated across all slices on delayed-phase contrast-enhanced CT images of appendiceal lesions using a semi-automated volumetric approach with manual corrections when necessary(Fig. 4). From 2153 initially extracted radiomics features, features with correlation coefficients below 0.80 were eliminated. Seven optimal features were selected through joint feature selection, comprising one morphological feature, one first-order feature and five texture features. The logistic regression classifier constructed with these features achieved an AUC of 0.84 (accuracy, 75%; sensitivity, 74%; specificity, 75%) in the training set and 0.80 (accuracy, 73%; sensitivity, 70%; specificity, 76%) in the testing set.The final results presented are from the best-performing fold in the 5-fold cross-validation.The differences between folds were minimal and statistical tests showed no significant differences (p > 0.05).

Fig. 4.

Multiphasic CT demonstration of appendiceal PMP in a high-grade case. (A) Unenhanced phase, (B) Arterial phase, (C) Venous phase, and (D) Delayed phase. Delayed-phase imaging (D) optimally delineates tumor-mucin interfaces (arrows) and enhances stromal textures.Appendiceal wall enhancement(circle) is more prominent in the delayed phase than in the arterial/venous phase

Combined model

The integration of optimal clinical predictors (CA 19 − 9, CEA, and CA-125) with selected radiomics features (n = 7) through logistic regression yielded a combined model with superior diagnostic performance. In the training cohort, the model demonstrated excellent discrimination (AUC, 0.91; 95% CI: [0.86–0.95]), with an accuracy of 84% (sensitivity, 81%; specificity, 86%), and good calibration (Brier score, 0.124). These results were maintained in the independent testing cohort (AUC, 0.88; 95% CI: [0.83–0.93]; accuracy, 77%; sensitivity, 75%; specificity, 79%; Brier score, 0.142), indicating robust generalizability.The final results presented are from the best-performing fold in the 5-fold cross-validation.The differences between folds were minimal and statistical tests showed no significant differences (p > 0.05).

Model performance analysis

As detailed in Table 2; Fig. 5, comprehensive evaluation revealed the combined clinical-radiomics model demonstrated superior diagnostic performance compared to individual models. The integrated model achieved an AUC of 0.88 in the testing cohort, significantly outperforming both standalone models (DeLong test: combined vs. clinical model, Z = 2.06, P = 0.041; combined vs. radiomics model, Z = 2.76, P = 0.006). This performance advantage was consistent in the training cohort (combined vs. clinical: Z = 2.50, P = 0.012; combined vs. radiomics: Z = 2.51, P = 0.012). Notably, no significant difference existed between the radiomics and clinical models in either cohort (P > 0.05 for all comparisons).

Table 2.

Comparison of diagnostic efficiency of different models for appendiceal PMP

Model	vs.Combined model P value	AUC (95% CI)	Accuracy	F1 score	Sensitivity	Specificity	PPV	NPV
Training set
Clinical model	0.012*	0.82(0.75–0.88)	0.73	0.71	0.70	0.77	0.72	0.75
Radiomics model	0.012*	0.84(0.78–0.90)	0.75	0.73	0.74	0.75	0.72	0.77
Combined model	/	0.91(0.86–0.95)	0.82	0.81	0.80	0.85	0.82	0.83
Testing set
Clinical mode	0.041 *	0.78(0.71–0.86)	0.73	0.71	0.70	0.77	0.72	0.75
Radiomics model	0.006 **	0.80(0.73–0.86)	0.73	0.71	0.70	0.77	0.72	0.75
Combined model	/	0.88(0.83–0.93)	0.77	0.75	0.74	0.79	0.75	0.78

* p < 0.05, ** p < 0.01 PPV: Positive Predictive Value NPV: Negative Predictive Value

Fig. 5.

The ROC curves of the three models for predicting pathological grade of appendiceal PMP in testing set

The calibration curves (Fig. 6) demonstrated excellent agreement between predicted probabilities and observed outcomes for the combined model, with superior goodness-of-fit compared to individual models. Decision curve analysis (Fig. 6) revealed all models provided clinical net benefit across threshold probabilities of 0.08–0.89, with the combined model offering the highest net benefit throughout this clinically relevant range. These results collectively indicate that the integrated approach provides more accurate risk stratification while maintaining clinical utility for appendiceal PMP grading.

Fig. 6.

The calibration curves and decision curve analysis of the three models for predicting pathological grade of PMP in testing set

Model interpretability analysis

To enhance the clinical interpretability of our optimal combined model, we performed SHAP analysis to quantify feature importance and elucidate decision-making patterns within the logistic regression classifier. As demonstrated in Fig. 7, the beeswarm plot ranks variables by their mean absolute SHAP values, revealing that wavelet-LLL_firstorder_Median (a radiomics feature reflecting tumor heterogeneity) and serum biomarkers CA 19 − 9 and CA-125 were the strongest positive predictors of high-grade PMP (P < 0.01).Conversely,

Fig. 7.

Influence of SHAP value on model output. Note Each row of the figure represents a variable, and each dot represents the SHAP value of the corresponding variable of a sample. The redder the color, the larger the variable value; The bluer the color, the smaller the variable value

wavelet-LHH_InverseVariance exhibited an inverse relationship with disease grade (β=-0.32, P = 0.02).The five most influential features were.

wavelet-LHH_glcm_InverseVariance (SHAP value: 0.41), Original_shape_Elongation (0.38), square_glszm_SizeZoneNonUniformityNormalized(0.35), wavelet-LLL_firstorder_Median (0.33), CA 19 − 9 (0.31).

This analysis demonstrates that while established serum biomarkers maintain diagnostic relevance, advanced radiomics features quantifying tumor morphology (elongation) and textural heterogeneity (wavelet transforms) provide complementary predictive value for PMP grading. The SHAP visualization offers clinicians intuitive insight into how specific imaging characteristics and laboratory values collectively contribute to individualized risk assessment.

Discussion

Pseudomyxoma peritonei (PMP) is an extremely rare condition (1–3 cases/million annually) with variable biological behavior, from indolent mucin accumulation to aggressive carcinomatosis [2]. Despite CRS/HIPEC, prognosis depends heavily on histopathological grade [20], making accurate preoperative grading crucial for treatment decisions. High-grade PMP often requires maximal cytoreduction (CC-0) or neoadjuvant chemotherapy due to high recurrence risk, while low-grade cases may be managed with minimally invasive laparoscopic approaches relative to conventional open techniques [21]. For high-grade PMP cases, knowing the grade in advance can help plan more aggressive surgical approaches to achieve optimal cytoreduction.Compared to prior radiomics-based studies, our model demonstrated notable improvements in diagnostic accuracy. Zhou et al. [16] utilized plain CT omentum images and clinical biomarkers to predict PMP grading in 88 patients (28 high-grade cases), reporting an AUC of 0.79. In contrast, our combined model, leveraging delayed-phase CT radiomics and clinical features, achieved an AUC of 0.88 in the testing cohort. Similarly, Sun et al. [17] analyzed CT signs in 75 PMP patients (18 high-grade cases) but relied solely on qualitative imaging features. Our radiomics approach quantified tumor texture and morphology, yielding a more objective and reproducible assessment.

This study introduces a novel application of delayed-phase contrast-enhanced CT in radiomics analysis of PMP, demonstrating significant clinical value. Compared to conventional CT protocols, delayed-phase imaging offers superior visualization of appendiceal lesions through three distinct advantages: (1) enhanced differentiation between progressively enhancing fibrovascular tumor stroma and hypodense free mucin, significantly improving tumor boundary delineation; (2) the contrast retention effect correlates with tumor cellular density in high-grade PMP with desmoplastic reaction, providing a pathological basis for texture feature analysis; and (3) a standardized imaging protocol that ensures feature stability. While acknowledging the increased radiation exposure from multiphase imaging, we maintain this protocol’s necessity as simplified approaches like single-phase imaging cannot adequately assess high-grade features, vascular involvement, or distant metastases-all critical for surgical planning. Current efforts to mitigate radiation risks include exploring MRI alternatives and investigating advanced techniques like AI-enhanced low-dose CT, aiming to preserve diagnostic accuracy while reducing cumulative dose exposure for patients requiring longitudinal monitoring.

Liang et al. [22] found that elevated serum and ascites abnormal biomarkers were independent risk factors for pathological grade of PMP. The higher the serum and ascites abnormal biomarkers CEA, CA19-9, CA125, the higher the pathological grade. Blaj S et al. [23] found that preoperative serum CA19-9 and CEA levels were significantly correlated with high-grade AMNs, but could also predict the recurrence of patients with low-grade histology. Biomarkers such as CEA and CA19-9 can inhibit cell differentiation and promote the formation of cell adhesion, thereby preventing cell death through apoptosis and enhancing their metastatic capabilities [24]. CA125 is involved in the process of peritoneal metastasis by mediating the adhesion between peritoneal free tumor cells and the surface of peritoneal mesothelial cells. Increased levels of these biomarkers often signify the presence of peritoneal metastasis or ascites. These biomarkers play an important role in leukocyte adhesion and extravasation. The results of this study showed that CA-199, CA-125 and CEA were independent risk clinical factors for pathological grade of PMP, which was consistent with the results of previous studies.

The model’s incorporation of PCI metrics acknowledges the well-documented correlation between CT-PCI and surgical PCI assessments (r > 0.900) [25], while addressing the unique challenges of PMP imaging interpretation through delayed-phase CT enhancement protocols that improve lesion delineation from mucinous ascites through characteristic mild late-phase enhancement patterns. However, it was found that CT-PCI was not an independent risk factor for pathological grade of PMP. The main reason may be that the mucus of appendiceal mucinous tumors has a secretory function, and after the rupture of the tumor, it is easy to implant in the abdominal cavity to form PMP. At the same time, the lesions are mainly distributed by gravity and involve the pelvic cavity, retrohepatic space, left paracolonic sulcus, hepatoduodenal ligament, etc. Therefore, the PCI score may not have an obvious relationship with pathological grade. In the future, we will increase the sample size to further analyze this result.

Radiomic feature analysis revealed particularly strong predictive value from wavelet-derived texture parameters and morphological characteristics, with texture features constituting 71% of the final feature set, reflecting their capacity to quantify tumor heterogeneity and organizational architecture through Gray Level Co-occurrence Matrix (GLCM) analysis and other advanced textural metrics that capture critical aspects of tumor biology not apparent on conventional imaging evaluation [26, 27].These features are not only related to the histological features of the tumor, but may also be related to the biological behavior of the tumor, such as cell proliferation, angiogenesis, and metabolic activity. The model’s superior performance (AUC 0.88) relative to standalone clinical or radiomics approaches and excellent calibration (Brier scores 0.124–0.142) validates this integrative methodology while addressing the limitations of single-modality prediction systems that fail to meet precision medicine standards [28, 29].

In this study, considering the importance of the generalization and interpretability of the model in practical applications, a five-fold cross-validation method was employed to enhance the model’s generalization capability and robustness. Machine learning models often suffer from the ‘black box effect’, resulting in a lack of interpretability in their decision-making processes [30]. Nohara and colleagues [31] further confirmed that SHAP values have better interpretability than the coefficients of generalized linear regression models, as evidenced by a comparative analysis of the interpretation results with other existing methods. Moreover, they found that SHAP summary maps provide more effective visualization results than feature importance maps.

In the study, wavelet-LHH_glcm_InverseVariance, original_shape_Elongation, square_glszm_SizeZoneNonUniformityNormalized, wavelet-LLL_firstorder_Median, and CA-199 emerged as the top five features in terms of importance, with the texture feature GLCM taking the lead. Theoretical and experimental evidence has established the effectiveness of GLCM in texture analysis, a method widely used to translate gray values into detailed texture information, offering significant clinical insights [32]. The most important clinical index is CA-199. In 2013, they reported the prognostic significance of preoperative CA 125, CEA and CA19-9 in 150 patients with appendiceal PMP over a 15-year period. CA19-9 was the best independent predictor not only for survival but also for recurrence in patients with PMP. Many studies have supported this conclusion [33]. The results of this study are fully consistent with previous studies, with CA19-9 having the highest prediction coefficient.

Our model demonstrates clinically significant performance in PMP grading (79% specificity, 74% sensitivity), with important management implications. False positives (21%) primarily stem from borderline lesions (38%), mucin-obscured tumors (29%), and intermediate histology (33%), addressed through probabilistic scoring (including an “indeterminate” category), MRI-DWI (ADC < 1.1 × 10⁻³ mm²/s for high-grade detection), and multidisciplinary review. Conversely, false negatives (26%) predominantly involve mucin-dominant cases (42%) or extensive ascites (31%), mitigated by ascites CEA analysis (> 5 ng/mL), targeted biopsies, and enhanced surveillance (3-month imaging). Current efforts focus on multicenter validation, advanced imaging integration (whole-body DWI-MRI/FDG-PET/CT), and radiomic-molecular correlation via liquid biopsies to refine diagnostic accuracy and clinical utility. These initiatives aim to optimize risk stratification while addressing real-world diagnostic challenges in PMP management.

Study limitations including its retrospective design, single-center patient cohort, and sample size constraints inherent to this rare disease entity underscore the need for future multicenter validation studies incorporating prospective designs and advanced artificial intelligence methodologies to further refine model performance and generalizability. Nevertheless, the current investigation makes significant contributions through its novel application of delayed-phase CT radiomics, development of an interpretable combined prediction model, and compilation of one of the largest high-grade PMP cohorts in the literature, establishing a foundation for improved personalized treatment strategies in this challenging disease.

We acknowledge the limitations of forward stepwise regression, including its sensitivity to initial variable selection and potential for local optima, particularly given our study’s modest EPV ratio of 7.3. However, we selected this approach for its clinical interpretability and demonstrated robustness in small-sample settings. We performed variance inflation factor (VIF) analysis, which yielded all well below the conservative threshold of 5, and five-fold cross-validation showed consistent performance. Prior studies support stepwise regression’s utility in low-dimensional clinical data [34, 35]. As part of ongoing work, we plan multicenter validation and comparison of different regression approaches in expanded datasets.

• In conclusion, our interpretable combined model integrates delayed-phase CT radiomics and clinical biomarkers to noninvasively predict PMP grading with high accuracy (AUC 0.88). The SHAP framework elucidates feature contributions, enhancing clinical trust and applicability. Key limitations include retrospective design and single-center data; future multicenter prospective studies should validate these findings and explore integration with artificial intelligence for further refinement. This approach aligns with precision oncology goals, offering a practical tool for preoperative decision-making in PMP management.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Zhiqun Wang and Zhhuozhao Zheng, as the corresponding author of the group, has played a leading role throughout the project. They have been responsible for coordinating the overall research direction and ensuring the smooth progress of the project. His extensive experience and vision have provided critical guidance for the team. Dong Bai has been the backbone of the technical implementation and written the paper. She has been responsible for the project’s technical development and problem-solving. His expertise and attention to detail have ensured the project’s technical feasibility and advancement. Guanjun Shi has been in charge of the project’s experimental design and data collection. He has meticulously planned and executed experiments, providing reliable data support for the research. Yuanzi Liang has been responsible for the project’s literature review and background research. She has gathered and organized a wealth of relevant materials, providing solid theoretical support for the research. Fang Li has been the project’s document manager. She has been responsible for organizing and managing the project’s documents, ensuring the completeness and accuracy of project materials.

Funding

This research received no specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The datasets used and analyzed during the current study are not publicly available due to patient privacy protections but are available from the corresponding author upon reasonable request. All data are fully anonymized to comply with ethical guidelines.

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Aerospace Center Hospital Ethics Committee Review Board. Due to the anonymized and retrospective nature of the data, the requirement for informed consent was waived by the ethics committee. All procedures followed were in accordance with the ethical standards of the responsible committee and with the Helsinki Declaration (as revised in 2013).

Consent for publication

Not applicable. This study utilized anonymized pathological and clinical data without identifiable personal information, images, or videos.

Data overlap

The datasets used in this study represent unique patient cohorts with no overlap with prior publications.

Competing interests

The authors declare no competing interests.