Radiomics analysis of dual-layer detector spectral CT-derived iodine maps for predicting Ki-67 PI in pancreatic ductal adenocarcinoma

Abstract

Objective

To evaluate the feasibility of radiomics analysis using dual-layer detector spectral CT (DLCT)-derived iodine maps for the preoperative prediction of the Ki-67 proliferation index (PI) in pancreatic ductal adenocarcinoma (PDAC).

Materials and methods

A total of 168 PDAC patients who underwent DLCT examination were included and randomly allocated to the training (n = 118) and validation (n = 50) sets. A clinical model was constructed using independent clinicoradiological features identified through multivariate logistic regression analysis in the training set. The radiomics signature was generated based on the coefficients of selected features from iodine maps in the arterial and portal venous phases of DLCT. Finally, a radiomics-clinical model was developed by integrating the radiomics signature and significant clinicoradiological features. The predictive performance of three models was evaluated using the Receiver Operating Characteristic (ROC) curve and Decision Curve Analysis. The optimal model was then used to develop a nomogram, with goodness-of-fit evaluated through the calibration curve.

Results

The radiomics-clinical model integrating radiomics signature, CA19-9, and CT-reported regional lymph node status demonstrated excellent performance in predicting Ki-67 PI in PDAC, which showed an area under the ROC curve of 0.979 and 0.956 in the training and validation sets, respectively. The radiomics-clinical nomogram demonstrated the improved net benefit and exhibited satisfactory consistency.

Conclusions

This exploratory study demonstrated the feasibility of using DLCT-derived iodine map-based radiomics to predict Ki-67 PI preoperatively in PDAC patients. While preliminary, our findings highlight the potential of functional imaging combined with radiomics for personalized treatment planning.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is a highly malignant disease with a 5-year survival rate of <12%. It is projected to become the second-leading cause of cancer-related deaths by 2030 [1, 2]. Ki-67 is a marker to distinguish between proliferating and non-proliferating cells [3]. A high Ki-67 proliferation index (PI) has been shown to correlate with reduced overall survival and recurrence-free survival [4, 5]. Preoperatively identifying PDAC patients with high Ki-67 PI and implementing neoadjuvant therapy may result in better clinical benefits than performing upfront surgery.

While immunohistochemical (IHC) examination is a standard method for assessing Ki-67 PI in PDAC, limitations such as small sample sizes in biopsies, tumor heterogeneity, and the operator’s technical proficiency and experience can affect the results [6, 7]. In addition, the invasive nature of obtaining biopsy and surgical specimens exacerbates these challenges, whereas non-invasive imaging techniques such as CT offer a promising alternative. If CT imaging modalities demonstrate an association between Ki-67 PI and PDAC characteristics, they could serve as indirect prognostic indicators. This could help identify PDAC patients who may benefit from more aggressive therapeutic interventions.

Radiomics, which enables the extraction of high-dimensional quantitative features from standard medical images, has shown great promise in tumor characterization, risk stratification, and prognosis prediction across various malignancies [8, 9]. However, most radiomics studies in PDAC rely on conventional CT or MRI, with limited functional information. The emergence of dual-layer spectral detector CT (DLCT) allows for material decomposition and the generation of iodine maps, providing functional insights into tumor perfusion and vascularity. DLCT-derived quantitative parameters have demonstrated utility in predicting tumor stage, lymph node (LN) metastasis, and histological differentiation in PDAC [10,11,12], while DLCT-based radiomics has shown potential in evaluating tumor heterogeneity and metastatic potential [10, 13].

Despite these advancements, few studies have investigated the potential of DLCT-based radiomics for predicting Ki-67 PI in PDAC. Our group has previously demonstrated the feasibility of using both conventional CT radiomics and DLCT-derived parameters for this purpose [14,15,16]. Building on these efforts, we hypothesize that variations in Ki-67 PI reflect microscopic changes in tumor perfusion heterogeneity, which can be quantitatively captured by radiomics features extracted from DLCT-derived iodine maps. The arterial phase (AP) and portal venous phase (PVP) provide complementary information on tumor vascularity, and their combined analysis may enhance the assessment of perfusion patterns and tumor aggressiveness. Previous studies have confirmed that dual-phase radiomics improves prognostic prediction in PDAC [17,18,19]. Therefore, this study aims to investigate whether radiomics features derived from dual-phase DLCT iodine maps can serve as non-invasive biomarkers for preoperatively predicting Ki-67 PI in PDAC patients.

Materials and methods

Patients

The study was approved by the institutional review board of Chongqing General Hospital, with a waiver of written informed consent due to the retrospective study design. Between July 2021 and December 2023, patients diagnosed with pathologically confirmed PDAC were enrolled. As depicted in Fig. 1, patients were eligible for the study if they (a) had a histopathologically confirmed diagnosis of PDAC; (b) obtained Ki-67 PI through IHC; and (c) underwent DLCT scans within 2 weeks at our institution before the IHC. Patients were ineligible if they (a) had received any relevant treatment (radiotherapy, chemotherapy, or chemoradiotherapy) before IHC; (b) had nondiagnostic CT image quality; (c) had coexisting other primary malignancies; or (d) had partially missing DLCT images or clinicopathological data.

Fig. 1

Flowchart of the study population. PDAC, pancreatic ductal adenocarcinoma; IHC, immunohistochemistry; DLCT, dual-layer detector spectral computed tomography; PI, proliferation index

Immunohistochemical analysis of Ki-67 PI

Standard IHC examination was performed on specimens obtained from fine needle aspiration or surgery to detect Ki-67 PI in all 168 patients. Positive cells were identified as those with a brown nucleus. Ki-67 PI was assessed by calculating the percentage of positive cells among 1,000 malignant cells at 200 × magnification. In the absence of a predetermined optimal threshold, we adopted a threshold of 50%, as suggested by previous research [20]. In our study, PDAC was categorized into either the low Ki-67 PI group (<50%) or the high Ki-67 PI group (≥50%).

DLCT image acquisition

All patients underwent DLCT examinations at our institution, with detailed scan protocols provided in the Supplementary Materials.

Evaluation of clinical features and CT imaging signs

Clinical features, including gender, age, body mass index, carbohydrate antigen (CA) 19–9, CA125, and carcinoembryonic antigen (CEA), were retrieved for each patient from the hospital information system.

All CT imaging signs were independently analyzed by two abdominal radiologists (hereinafter referred to as Radiologists A and B), with 7 and 11 years of experience, respectively. They remained blinded to clinical and pathological details throughout the evaluation process. The final results were determined by consensus between the two radiologists. The candidate CT imaging signs of PDAC included CT-reported T stage, CT-reported regional LN status, vascular invasion, and extrapancreatic perineural invasion. The evaluation criteria for each imaging sign are described in the Supplementary Materials.

Volumes of interest (VOI) segmentation

The workflow for establishing the key steps of the radiomics signature is illustrated in Fig. 2. Radiologist A used 3D Slicer (open-source software, version 5.6.2, https://www.slicer.org/) to manually delineate tumor contours slice by slice on PVP images, then transferred the contours from PVP images to AP images. If there were contour mismatches due to respiratory motion, manual adjustments were made to align the replicas with the contours observed in the PVP. Two weeks later, 30 cases were randomly selected, and segmentation was repeated by two radiologists (A and B) to assess intra- and interobserver reproducibility. Neither radiologist was aware of the histopathological results. Features with inter- and intraobserver correlation coefficients (ICCs) greater than 0.75 were considered sufficiently consistent for further analysis.

Fig. 2

Workflow of the key steps in conducting radiomics analysis of iodine maps. DCA, decision curve analysis; KW, Kruskal-Wallis; RFE, recursive feature elimination; ANOVA, analysis of variance; SVM, support vector machines; LDA, linear discriminant analysis; LR, logistic regression; LRLasso, lasso logistic regression; ROC, receiver operating characteristic; VOI, volumes of interest; LN, lymph node; CA19-9, carbohydrate antigen 19–9; AUC, area under the curve

Radiomics extraction and selection

All radiomics analyses were implemented with FeAture Explorer Pro (FAE, V 0.5.14) on Python (3.7.6) [21]. Both AP and PVP DLCT-derived iodine maps were used for radiomics feature extraction. Before feature extraction, the CT images were resampled to a standardized pixel dimension of 1.0 × 1.0 × 1.0 mm³ and normalized to a scale of 1000. Gray-level discretization was applied to the original intensities by resampling them into 25 fixed bins. The processed images were subjected to various transformations using wavelet, square root, square, logarithm, gradient, exponential, Laplacian of Gaussian, and local binary pattern (3D) filters. The extracted features in this study included shape-based features, first-order statistical features (e.g., maximum, skewness, mean, median), and multiple texture matrices, including the gray-level co-occurrence matrix (GLCM), gray-level dependence matrix (GLDM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighboring gray tone difference matrix (NGTDM).

First, Z-score normalization was applied to features with ICCs greater than 0.75. Second, the Pearson correlation coefficient (PCC) was used to measure the correlation between each pair of features and to reduce dimensionality. If the PCC exceeded 0.99, one of the redundant features was randomly removed. Finally, feature selection was conducted using techniques including Kruskal-Wallis, Relief, recursive feature elimination (RFE), and analysis of variance (ANOVA). All these methods (KW, Relief, RFE, ANOVA) were applied for feature selection. We constructed multiple model pipelines, where each pipeline consisted of: (1) a feature selection method (KW, Relief, RFE, ANOVA), (2) a machine learning model (support vector machines (SVM), linear discriminant analysis (LDA), logistic regression (LR), and lasso logistic regression (LRLasso)) (Table S1), and (3) varying numbers of features (from 1 to 15). The performance of each pipeline was evaluated using 10-fold cross-validation. The discrimination performance was evaluated using the receiver operating characteristic (ROC) curve, and the radiomics signature that demonstrated the best predictive performance in the validation set was chosen as the optimal signature (The detailed parameter configuration of ML algorithms can be found in the Supplementary Materials).

Development of the clinical model and radiomics-clinical nomogram

First, the clinicoradiological features (i.e., the clinical features and CT imaging signs) between the high and low Ki-67 PI groups were compared using the Chi-square test, Mann-Whitney U test, and two-sample t test. Then, significant factors with p < 0.05 from univariate analysis were included in multivariate logistic regression (LR) analysis to determine the independent factors for constructing the clinical model in the training set.

The radiomics-clinical nomogram was developed by integrating the radiomics signature from dual-phase iodine maps and significant clinicoradiological features using multivariate LR. The odds ratio (OR) and 95% confidence interval (CI) were calculated for each independent predictive factor.

Assessment of the performance of the different models

The area under the curve (AUC) with 95% CI, sensitivity, and specificity was used to assess the performance of the three models (clinical model, radiomics signature, and radiomics-clinical nomogram) in both the training and validation sets. The DeLong test was applied to compare the AUCs of the three models. Subsequently, a nomogram was constructed using the optimal model. Decision curve analysis (DCA) was conducted to evaluate the clinical usefulness of the three models by quantifying their net benefits across different threshold probabilities in the validation set. Finally, the calibration curve and Hosmer-Lemeshow test were used to evaluate the calibration performance of the nomogram in both the training and validation sets.

Statistical analysis

All statistical analyses and calculations were performed using R software (http://www.R-project.org), SPSS (version 26.0, IBM), and MedCalc (version 18.2.1, MedCalc Software). The Shapiro-Wilk test was used to assess data normality. Normally distributed data were presented as mean ± standard deviation (SD), while non-normally distributed data were expressed as median (25th, 75th percentiles). Categorical variables were analyzed using the Chi-square test, and continuous variables were assessed using either the Mann-Whitney U test or the two-sample t test. Statistical significance was defined as a two-sided p-value < 0.05.

Results

Patient characteristics

A total of 168 patients were analyzed, with 118 in the training set and 50 in the validation set. The rates of low and high Ki-67 were 70.3% (83/118) and 29.7% (35/118) in the training set and 70.0% (35/50) and 30.0% (15/50) in the validation set, respectively. There were no significant differences between the training and validation sets in any clinicoradiological features (Table 1; all p > 0.05).

Table 1 Clinicoradiological features of PDAC patients

Clinical model development

Table 2 presents the results of the univariate and multivariate LR analyses between the high and low Ki-67 PI groups in the training set. The univariate analysis showed significant associations of CA19-9, CT-reported regional LN status, and extrapancreatic perineural invasion with Ki-67 PI (p < 0.05). Furthermore, stepwise multivariate LR analysis identified CA19-9 and CT-reported regional LN status as independent predictors of Ki-67 PI (Table 3). Consequently, a clinical model was developed based on these two predictors.

Table 2 Univariate analysis to differentiate between high and low Ki-67 PI groups in the training set

Table 3 Multivariate logistic regression analysis in the training set

Feature selection and radiomics signature building

Among the 2,622 dual-phase radiomics features, 1,982 stable features (974 from AP and 1,008 from PVP) with an ICC greater than 0.75 were retained for further analysis (Fig. S1). In constructing the radiomics signature, a pipeline utilizing Z-score normalization, PCC dimension reduction, RFE feature selection, and LR classifier identified six features: two first-order statistical features and four textural features. These features were included in the radiomics signature, which demonstrated optimal predictive performance. The composition and contribution of these features are illustrated in Fig. 3.

Fig. 3

Radiomics feature selection results. AP, arterial phase; PVP, portal venous phase; GLCM, gray-level co-occurrence matrix; GLRLM, gray-level run length matrix; GLSZM, gray-level size zone matrix

Following the multivariate LR analysis, we found that the radiomics signature (OR = 2.975; 95% CI: 1.724–5.134; p < 0.001), CA19-9 (OR = 7.734; 95% CI: 1.053–56.793; p = 0.044), and CT-reported regional LN status (OR = 11.461; 95% CI: 1.992–65.948; p = 0.006) were independent predictors for Ki-67 PI (p < 0.05) in the training set (Table 3). All these predictors were incorporated into the radiomics-clinical model and visualized as a nomogram, as shown in Fig. 4.

Fig. 4

Radiomics-clinical nomogram developed in the training set, incorporating the radiomics signature, CA19-9, and CT-reported regional LN status. CA19-9, carbohydrate antigen 19–9; LN, lymph node

Comparison of the predictive performance among models

The predictive performance of the three models is summarized in Table 4, including AUC, sensitivity, specificity, and DeLong test results. ROC analyses conducted to distinguish high and low Ki-67 PI in both the training and validation sets for the three models are shown in Fig. 5. The clinical model exhibited good diagnostic effectiveness in predicting Ki-67 PI among PDAC patients, achieving an AUC of 0.805 (95% CI: 0.724–0.886) in the training set, while demonstrating moderate effectiveness with an AUC of 0.720 (95% CI: 0.574–0.866) in the validation set. Both the radiomics signature and the radiomics-clinical model demonstrated excellent diagnostic effectiveness in predicting Ki-67 PI in PDAC, with AUCs of 0.957 (95% CI: 0.926–0.988) and 0.979 (95% CI: 0.961–0.998) in the training set, respectively. In the validation set, the radiomics signature and the radiomics-clinical model achieved AUCs of 0.901 (95% CI: 0.817–0.985) and 0.956 (95% CI: 0.904–1.000), respectively. The DeLong test indicated a significant improvement in the AUC of the radiomics-clinical model compared to the clinical model in both the training set (p < 0.001) and the validation set (p = 0.002). Additionally, a significant difference was observed between the radiomics signature and the clinical model in the training set (p < 0.001), but this difference was not statistically significant in the validation set (p = 0.057). In terms of clinical benefit, DCA demonstrated that the radiomics-clinical model provided a greater net benefit across a threshold probability range of 0.05–0.95, compared to both the clinical model and radiomics signature, indicating superior clinical utility (Fig. 6). The calibration curves for the radiomics-clinical model demonstrated strong alignment between predicted and observed outcomes in both sets (Fig. 7). The Hosmer-Lemeshow test showed no significant difference (p = 0.952 for the training set and p = 0.728 for the validation set), indicating that the nomogram was well-calibrated without significant deviation from the ideal fit.

Table 4 Predictive performance of the clinical model, radiomics signature, and radiomics-clinical models

Fig. 5

ROC curves depicting the predictive performance of the clinical model, radiomics signature, and radiomics-clinical models for Ki-67 PI in PDAC (a, b). AUC, area under the curve; PI, proliferation index; PDAC, pancreatic ductal adenocarcinoma; ROC, receiver operating characteristic

Fig. 6

DCA results for the clinical model, radiomics signature, and radiomics-clinical models (a, b). DCA, decision curve analysis

Fig. 7

Calibration curves of the radiomics-clinical nomogram (a, b)

Discussion

In this retrospective study, we built upon our team’s prior investigations to develop a radiomics signature based on dual-phase DLCT-derived iodine maps for predicting Ki-67 PI in PDAC patients. Furthermore, the radiomics-clinical model, integrating radiomics signature and independent clinicoradiological features, demonstrated superior predictive performance compared to either the radiomics signature or clinical model alone. This combined approach offers an effective and non-invasive method for the preoperative prediction of Ki-67 PI in PDAC patients.

The present study builds upon our previous work investigating imaging-based prediction of Ki-67 PI in PDAC. Li et al. demonstrated the feasibility of conventional CT radiomics, while subsequent studies explored the predictive value of DLCT-derived quantitative parameters [14,15,16]. Extending these efforts, the current research integrates dual-phase (arterial and portal venous) iodine maps derived from DLCT with advanced radiomics analysis to more comprehensively characterize tumor heterogeneity. Compared with conventional CT, DLCT iodine maps provide functional insights by separating iodine content from tissue attenuation, enabling a more accurate evaluation of tumor vascularity and perfusion [14, 22]. Dual-phase imaging captures complementary perfusion features—arterial phase reflects early vascularization, while portal venous phase offers additional information on tissue perfusion—thereby enhancing the assessment of tumor aggressiveness. Furthermore, unlike previous DLCT-based studies that did not incorporate clinical variables, our model integrates key clinicoradiological features such as CA19-9 and CT-reported regional LN status [15, 16], thereby improving the predictive performance and clinical applicability of the model.

Building on our prior findings, we further validated CA19-9 and CT-reported regional LN status as independent predictors of Ki-67 PI through multivariate analysis [14]. CA19-9 is a recognized prognostic marker in PDAC, with elevated levels linked to poor differentiation and outcomes [23, 24]. Ki-67 PI reflects tumor proliferative activity, with higher values indicating greater aggressiveness and worse prognosis [5, 25]. In this study, the high Ki-67 PI group showed significantly higher CA19-9 levels and CT-reported regional LN positivity (72.0% vs. 32.2%), consistent with reports associating high Ki-67 expression with increased regional LN metastasis [26, 27]. The strong correlation between CT-reported and pathological LN status further supports the reliability of imaging for LN assessment [4, 13, 28]. This association likely reflects the role of Ki-67 in tumor proliferation and lymphatic invasion, indicating that CT-reported regional LN status may serve as a non-invasive surrogate for Ki-67 PI. The clinical model incorporating CA19-9 and LN status achieved good performance in the training set (AUC = 0.805), but only moderate performance in the validation set (AUC = 0.720), suggesting that combining tumor markers and morphological features alone may be insufficient, and that integrating perfusion or intratumoral characteristics could improve predictive performance.

DLCT enhances CT diagnostics by enabling functional imaging through iodine maps, which accurately depict lesion vascularity and improve contrast between hypoattenuating tumors and normal parenchyma [29]. Radiomics extracts high-throughput quantitative features that reflect intratumoral perfusion heterogeneity, leveraging DLCT’s functional information [30, 31]. Prior studies have applied DLCT-based radiomics to predict LN metastasis, tumor stage, and differentiation in PDAC [10,11,12,13], highlighting its clinical potential. Unlike these studies, our work focuses on Ki-67 PI, a key marker of tumor proliferation. We developed a radiomics signature based on dual-phase iodine maps, achieving excellent performance in both training (AUC = 0.957) and validation sets (AUC = 0.901). Among six selected features, five were from the PVP and one from the AP, aligning with findings that PVP images better capture PDAC biology and enhance tumor delineation [32,33,34]. Texture features predominated, consistent with prior studies [14], likely due to their ability to quantify spatial heterogeneity. Specifically, GLCM captures local texture variation, GLRLM reflects fine texture continuity, and GLSZM characterizes larger structural patterns [35,36,37]. These features provide a comprehensive view of tumor heterogeneity relevant to proliferation and prognosis. Notably, first-order features, while fewer, showed higher importance, offering direct intensity-based information with strong reproducibility across imaging conditions [38,39,40]. The integration of texture and first-order features balances biological insight and robustness, supporting clinical applicability.

To enhance predictive performance, the radiomics signature was combined with CA19-9 and CT-reported regional LN status to construct a radiomics-clinical model, which achieved the highest AUCs in both training (0.979) and validation (0.956) sets, outperforming the radiomics and clinical models alone. The improvement over the clinical model was statistically significant (p < 0.001 and p = 0.002, respectively). This enhancement likely stems from the model’s ability to capture tumor aggressiveness and heterogeneity via iodine map-derived features. DCA demonstrated superior net benefit of the radiomics-clinical model, supporting its clinical utility. Calibration curves showed good agreement between predicted and observed probabilities, with a nonsignificant Hosmer-Lemeshow test (p = 0.384), indicating satisfactory model fit. These results suggest that integrating radiomics and clinical features offers complementary value and significantly improves the prediction of Ki-67 PI in PDAC.

Despite promising results, this single-center study should be considered preliminary. As one of the few studies focusing on preoperative Ki-67 prediction in PDAC, it provides a foundation for future multicenter validation. Expanding to multiple institutions and assessing radiomics feature stability across different scanners, acquisition protocols, and reconstruction settings will be key to confirming model reproducibility and generalizability. Additionally, we are actively working to expand collaborations with other hospitals to promote Ki-67 IHC testing and DLCT scanning, which are essential for validating our approach. These efforts will facilitate the broader application of our model and support its prospective validation in independent cohorts, ensuring its clinical reliability and practical utility. Thirdly, future work may explore automated segmentation using deep learning combined with expert refinement to enhance reproducibility while maintaining clinical interpretability.

Conclusion

In conclusion, the radiomics-clinical model which integrates the radiomics signature from DLCT-derived iodine maps with clinicoradiological features exhibited excellent performance in preoperatively predicting Ki-67 PI in PDAC patients. This may help clinicians identify PDAC patients with high Ki-67 PI and facilitate personalized treatment strategies. However, it is important to note that these findings are preliminary and require external validation through multi-center studies to confirm their applicability and robustness in clinical practice.