Introduction

Bladder cancer (BCa) ranks as one of the most prevalent urologic malignancies and is the 10th most common cancer globally [1, 2]. The incidence of BCa is higher in men, where it stands as the 6th most common cancer and the 9th leading cause of cancer-related deaths [2]. In developed nations, around two-thirds of BCa cases manifest as non-muscle-invasive BCa (NMIBC), with the remaining cases presenting as muscle-invasive BCa (MIBC) [3]. The extent of muscle invasion serves as an independent prognostic factor that greatly influences clinical management and survival outcomes. MIBC is a lethal malignancy, results in over 85% mortality within 2 years if left untreated or treated inadequately [4, 5]. Given these grave implications, the ability to predict muscle invasion could substantially enhance preoperative planning and patient outcomes.

Biopsy, being the gold standard for diagnosing muscle invasion, is an invasive method that struggles to definitively determine the extent of muscle layer infiltration due to challenges in obtaining a sufficient sample [3, 6]. This pitfall may pose serious problems, including the delay of radical treatment, disease progression and worsening oncological results [7]. Radiomics, as an emerging field, holds potential in converting digital imaging data into innumerable and mineable quantitative features that reveal pathophysiology [8]. Previous researches have shown that radiomics models utilizing conventional CT images can accurately forecast muscle invasion in BCa [9,10,11,12]. Dual-energy computed tomography (DECT) enables qualitative analysis of lesions and generation of material decomposition and virtual monochromatic images (VMIs) for providing additional value of disease diagnosis [13,14,15]. Material decomposition allows for the estimation of the concentration of specific elements within a region of interest, thereby distinguishing absorption characteristics of different elements [16,17,18]. Processed images such as VMIs have shown promising results in oncological imaging for tumor detection, characterization, and assessing therapy response [15]. Although radiomics, DECT material decomposition, and VMIs have demonstrated potential for characterizing various tumor types, their combined value in predicting muscle invasion in BCa remains unexplored. Consequently, this study aimed to investigate the optimal DECT-derived radiomics model for evaluating the muscle status of BCa and to explore the clinical significance of combining quantitative parameters and radiomics to preoperatively predict the muscle invasion.

Methods

Patients

This retrospective study was approved by the Institutional Ethics Committee, and the patient consent was waived. The study population was retrospectively enrolled patients who underwent dual-energy computed tomography urography (DECTU) for BCa staging purpose. Among these patients, those scanned between April 2021 and October 2023 were recruited. The inclusion criteria were confirmed by pathological examination as BCa and their CTU scans performed with dual-energy technique. The exclusion criteria were (I) incomplete imaging datasets, (II) incomplete clinicopathological diagnostic report, (III) inadequate image quality for analysis, (IV) lesions with a maximum diameter or thickness less than 10 mm, (V) radiation therapy or chemotherapy between CTU and pathological assessment. Ultimately, 126 patients were included. We randomly divided the cases into training (n = 88) and test (n = 38) cohorts with a ratio of 7:3. An overview workflow of this study is shown in Fig. 1.

Fig. 1
figure 1

Overview workflow of this study. DECT, Dual-energy CT; CTU, CT urography; MIBC, muscle-invasive bladder cancer; NMIBC, non-muscle-invasive bladder cancer; VMI, virtual monochromatic images; IMD, iodine material

Full size image

DECT image acquisition

All CTU examinations were performed on a 256-row CT scanner (Revolution CT, GE HealthCare, Milwaukee, WI, USA). The acquisition parameters were as follows: Gemstone Spectral Imaging (GSI) mode with fast tube voltage switching (between 80 and 140 kVp); automatic exposure mode (GSI Assist) for selecting tube current to achieve noise index of 10; detector width of 80 mm; pitch of 0.992:1; tube rotation speed of 0.6 s/r. The contrast injection, with the amount of iodine intake proportional to patient’s weight (500 mgI/kg) and at a maximum of 100 mL, was performed through a peripheral vein of the forearm at the injection speed of 3.0–4.0 ml/s, followed by appropriate amount of saline at the same rate to maintain a total injection time of 30 s. The non-ionic contrast agent (Ioversol, 350 mg/mL, Jiangsu Hengrui Pharmaceutical Co. Ltd., China) was used. The venous phase scan started at 70 s after administration of contrast medium. All venous phase images were reconstructed by using the adaptive statistical iterative reconstruction-V (ASIR-V) algorithm at 60% strength with slice thickness of 1.25 mm.

Quantitative parameter measurement

Spectral images were generated using the Advantage Workstation 4.7 dedicated to postprocessing (GE HealthCare) to reconstruct VMIs at 3 different energy levels (40, 70, and 100 keV), iodine (water)-based material decomposition (IMD), fat (water)-based and effective atomic number (Zeff) images. Quantitative parameters were measured on three adjacent slices containing the largest area and the substantial part of the primary lesion, avoiding cystic area, calcification and necrosis. Two radiologists (with 2 and 10 years of experience in abdominal radiology, respectively) performed the quantitative measurements, unaware of the pathologic findings. The average values from the two radiologists were used for the final evaluation. The quantitative parameters from DECT images included the following: (I) iodine concentration (IC) of the lesion (IClesion) and IC of descending aorta (ICaorta) in the same layer, to calculate the normalized iodine concentration (NIC) = IClesion/ ICaorta; (II) fat-water concentration (FC); (II) Zeff.

Images segmentation

The 40, 70, 100 keV and IMD images were imported into the uAI Research Portal V1.1 software developed by Shanghai United Imaging Intelligence Co., Ltd. The tumor segmentation was delineated manually in three dimensions slice-by-slice on the 40 keV axial images until the whole lesion was captured by a radiologist (with 2 years of experience in the abdominal imaging diagnosis), who was unaware of the clinical data and pathological results. To ensure accuracy, another senior radiologist (with 20 years of experience in abdominal imaging) verified and corrected all segmentation masks. Disagreements were resolved with consensus-based discussion. Subsequently, the resulting volume of interest (VOI) was copied to other types of images.

Radiomics model selection

Feature extraction

To address the interference caused by the non-uniform spatial resolution of CT images, we normalized all images using Z-score and resampled the voxel size spaced to 1 × 1 × 1 mm through the B-spline interpolation, and gray values discretized into 25 bin-width. Features were extracted separately from each images using uAI Research Portal V1.1 software. These radiomic features were subdivided into the following classes: (1) first-order features; (2) shape features; (3) texture features, including gray level co-occurrence matrix (GLCM) features, gray level size zone matrix (GLSZM) features, gray level run length matrix (GLRLM) features, gray level difference matrix (GLDM) features, and neighboring gray tone difference matrix (NGTDM) features.

Feature selection and radiomics model construction

All radiomics features in the training cohort were standardized using the Z-score normalization method. To select the optimal radiomics features to construct the final model, a step-wise feature selection strategy was adopted. Firstly, we conducted the recursive feature elimination (RFE) via five-fold cross-validation to solve over-fitting and classification accuracy problems. Then the least absolute shrinkage and selection operator (LASSO) was used to select the optimal radiomics features. Support Vector Machine (SVM) was utilized to construct radiomics models individually for single images (at 40, 70, 100 keV and IMD images) and multi-images (combining 40, 70, 100 keV and IMD images), aiming to identify the optimal performance radiomics model both in the training and test cohorts. A radiomics workflow is shown in Fig. 2.

Fig. 2
figure 2

A flow chart illustrating the process for building radiomics models in this study. RFE, recursion feature elimination; LASSO, least absolute shrinkage and selection operator

Full size image

Models building

DECT quantitative parameters selected through univariable and multivariate logistic regression analyses were utilized in building the DECT model. Then the most effective radiomics model was identified as the final choice. Finally, a fusion model integrating the radiomics scores (rad-score) from the radiomics model with the independent predictors from DECT model was developed using logistic regression algorithm. The fusion model was visually represented as a nomogram for individualized estimating the probability of MIBC in the training cohort.

Statistical analysis

Statistical analysis was performed using MedCalc version 20.2 (MedCalc, Ltd, Ostend, Belgium), R software (version 4.2.1), and SPSS version 26.0 (SPSS Inc., Chicago, IL, USA). Quantitative measurements from the monochromatic images and MD images in DECT were initially checked with the Shapiro-Wilk test for normal distribution. The normally distributed parameters were presented as mean ± standard deviation (SD), while skewed variables were expressed as the median and inter quartile range (IQR). Variables were compared using independent sample t-test, Mann-Whitney U Test, chi-square tests, or Fisher exact test as appropriate. Variables with a p value < 0.05 in the univariate analysis were identified as candidate variables for the multivariable binary logistic regression analysis. Subsequently, candidate variables were included as independent variables in a forward stepwise binary logistic regression to determine the independent predictors. A p value less than 0.05 was considered statistically significant.

The receiver operating characteristic (ROC) curves were drawn, and sensitivity, specificity, accuracy, precision, and the area under the curve (AUC) were calculated respectively based on the R language to evaluate the diagnostic efficacy of the different models, and statistical differences between AUCs were compared using the DeLong test. Decision curves were utilized to compare the clinical usefulness of models.

Results

Patient characteristics

The clinical characteristics and DECT quantitative parameters of the 126 BCa patients are summarized in Tables 1 and 2, respectively. No significant differences were observed in age, gender, lesion location and diameter between the training and test cohorts (all p > 0.05). Moreover, regarding quantitative parameters, NIC showed significant difference between the MIBC and NMIBC in both cohorts with all p < 0.05, while IC was significantly different in the training cohort only with p < 0.05.

Table 1 Clinical data of bladder cancer patients in training and test cohorts
Full size table
Table 2 Comparison of DECT quantitative parameters between MBIC and NMBIC
Full size table

Feature selection and optimal radiomics based model screening

A total of 104 radiomics features were extracted from each VOI based on the 40, 70, 100 keV and IMD images. Following feature selection using RFE and Lasso methods, the optimal radiomics features of each image set remained. The diagnostic performances of the 5 radiomics models in the training and test cohorts are shown in Table 3. Notably, the diagnostic performance based on multi-image was the best differentiating MIBC and NMIBC in the training (AUC = 0.943) and test cohorts (AUC = 0.867), respectively.

Table 3 The diagnostic performance of radiomics models
Full size table

Model building

DECT model

Among the four quantitative parameters (IC, NIC, FC and Zeff) analyzed, the univariable analysis revealed that IC and NIC were significant risk factors for predicting muscle invasion. Further multivariable logistic regression analysis showed that NIC (OR, 25.426; 95% CI: 5.458-118.434; P < 0.001) was the sole independent predictor (Table 4). Consequently, NIC was used for the DECT model.

Table 4 Independent factors for muscle invasion in bladder cancer
Full size table

Radiomics model

As previously mentioned, we selected the multi-image radiomics model based on the combination of the four sets of images (40, 70, 100 keV and IMD) as the optimal radiomics model.

Fusion model

We combined the independent predictor from the DECT model and the rad-score of the optimal radiomics model to construct the fusion model and develop a visual nomogram (Fig. 3).

Fig. 3
figure 3

The nomogram of the fusion model was constructed using the training cohort, incorporating NIC and rad-score of the multi-image radiomics model. NIC, normalized iodine concentration; Rad-score, radiomics scores; MIBC, muscle-invasive bladder cancer

Full size image

Overall performance assessment

Model comparison

The performance of three models in distinguishing between MIBC and NMIBC is presented in Table 5, and the DeLong test comparing the performance of these models is shown in Table 6. ROC curves illustrated that the radiomics model was superior to the DECT model in predicting the muscle invasion status in the training cohort (AUCradiomics= 0.943; AUCDECT =0.798) and the test cohort (AUCradiomics = 0.867, AUCDECT =0.704). Furthermore, the fusion model showed a slight additional improvement (AUCtraining = 0.943, AUCtest = 0.893) as depicted in Fig. 4. Decision curves revealed that both the radiomics model and the fusion model provided a greater net benefit compared to the DECT model (Fig. 5).

Table 5 Comparison of AUCs between DECT, radiomics, and fusion models
Full size table
Table 6 DeLong test for the comparison of AUC values between different models
Full size table
Fig. 4
figure 4

Receiver operating characteristic curves for DECT, radiomic, and fusion models in training (A) and test cohorts (B).

Full size image
Fig. 5
figure 5

Decision curves for DECT, radiomics, and fusion models in training (A) and test cohorts (B). Fusion model and radiomics model had higher net benefits than DECT model

Full size image

Diameter-stratified analysis

In the test cohort, for lesions larger than 3 cm, both the fusion model and the radiomics model achieved AUC values of 0.800. Conversely, for lesions smaller than 3 cm, the fusion model exhibited improved predictive ability with an AUC value of 0.911, although the difference compared to the AUC value of the radiomics model (0.884) was not statistically significant. Notably, the DECT model had the lowest AUC values in both groups. Specific advantages of the diagnostic performance are outlined in Table 7.

Table 7 The diagnostic performance of three models in diameter-stratified analysis
Full size table

Discussion

In this study, a fusion model integrating radiomics features and quantitative parameters derived from DECT was successfully developed to distinguish between MIBC and NMIBC. The results demonstrated that the fusion model exhibited the best predictive accuracy for muscle invasion, surpassing both the DECT and radiomics model. Additionally, this study highlighted the fusion model had potential advantages in identifying high-risk small lesions that may be challenging to identify. Our findings indicated that the fusion model held promise as a noninvasive tool for preoperative assessment of muscle status in BCa patients.

The iodine concentration (IC) derived from DECT enables the quantitative measurement of vascular enhancement in lesions, aiding in the differentiation of urothelial tumors from non-tumor lesions using IMD images [19, 20]. Our study demonstrated that the IC values of NMIBC were higher than those of MBIC, indicating discrimination potential. However, previous study had shown that IC values may be influenced by various factors, such as tumor size, blood volume, the degree of enhancement, and differences in hemodynamics among individuals [21]. On the other hand, NIC could mitigate variations resulting from differences in circulation statuses among individuals, providing a more accurate representation of tumor tissue blood supply. In contrary to IC, the NIC values of MBIC were significantly higher than those of NMBIC. Our logistic regression analysis concluded that only NIC was an independent predictor of muscle status, effectively distinguishing MBIC from NMBIC with an AUC of 0.704.

The existing CT-based radiomics models were based on conventional CT images to differentiate whether the tumor has invaded the muscular layer [9,10,11]. These studies demonstrated that radiomics models were effective in evaluating the invasiveness of muscle in BCa and outperformed visual assessments by radiologists. This underscores the potential of CT-based radiomics for evaluating muscle invasion in BCa. However, the impact of DECT on predicting muscle status remains unclear. Previous studies have highlighted the advantages of multi-energy models derived from DECT. Al Ajmi et al. [22] introduced a multi-energy texture analysis utilizing VMIs ranging from 40 to 140 keV, capturing energy-dependent tissue attenuation changes. Using the same population, they improved the diagnostic accuracy of benign parotid tumors in comparison to the single-energy level of 65 keV (75–92%). Additionally, Forghani et al. [23] demonstrated that multi-energy texture analysis outperformed single-energy methods in evaluating lymph node metastasis in head and neck squamous cell carcinoma. Li et al. [24] utilized VMIs at 40, 65, and 100 keV obtained from both the arterial and venous phases to predict lymph node metastasis in gastric cancer. Their results showed that a nomogram based on multi-energy data exhibited superior predictive performance and added value compared to single-energy and clinical models. Considering CT physics and algorithms, low-energy images (40–70 keV) better reflect tissue enhancement features, whereas high-energy images (100–140 keV) represent non-enhanced tissue characteristics [25]. In our study, we selected the 40 keV to represent low-energy images, 100 keV for high-energy images, and 70 keV as an equivalent to a standard 120-kVp CT acquisition [26, 27]. IMD images selectively depict tissue perfusion and permeability, crucially linked to tumor angiogenesis and aggressiveness [28]. They not only facilitate the quantitative analysis of IC but also find successful application in radiomics, offering valuable insights for tissue characterization, diagnosing metastases, and predicting survival outcomes [29,30,31,32,33]. Furthermore, there is a notable deficiency in radiomics utilizing IMD images for muscle invasion characterization in BCa, which may enhance the predictive model performance by incorporating information on blood supply variations. Building on this premise, we incorporated IMD images into a multi-image analysis. The multi-image radiomics model incorporated features from images at 40,70,100 keV as well as IMD images, creating a holistic image dataset of the lesion that overcomed the constrains of single-energy imaging techniques. The results of our study indicated that the multi-image radiomics model outperformed radiomics models based solely on single images in distinguishing between MIBC and NMIBC. This suggested that multi-image radiomics model effectively captured variations in energy-dependent attenuation among different tissues and variations in blood perfusion, providing more comprehensive information than single image set alone.

In our study, the multi-image radiomics model exhibited superior predictive efficacy compared to the DECT model in forecasting the invasiveness of muscle. While quantitative parameters can capture microscopic details like blood flow and material composition, radiomics delves into both intra-tumor and inter-tumor heterogeneity by analyzing the entire lesion and its microenvironment without invasive conditions [34, 35]. The combination of these two aspects enhances the accuracy in revealing the fundamental characteristics of the lesion. Thus, we conducted a novel fusion model that combines the rad-score of radiomics model with NIC for the first time to preoperatively forecast muscle invasiveness of BCa. The fusion model showcased improved accuracy from 82 to 87%, with the sensitivity and specificity increasing from 73% and 87–80% and 91%, respectively. Tumor size holds prognostic and predictive significance in NMIBC and is considered as a critical feature in determining muscle invasion. Then, we separately evaluated lesions stratified by diameter [36, 37]. Our findings showed that the fusion model had the largest AUC compared to both DECT and radiomics models in lesions smaller than 3 cm. These results proposed that the fusion model may offer significant value in the assessment of small lesions.

This study has several limitations. Firstly, it depended on manual tumor segmentation, a process susceptible to subjectivity introduced by radiologists. Secondly, the adoption of DECT was lower than that of conventional CT, leading to a decreasing number of patients receiving both pathological biopsies and DECT. Consequently, our imaging options were limited to single-mode techniques. Thirdly, the utilization of only venous-phase CT images may limit the diagnostic performance of the models, suggesting potential enhancement by incorporating multi-phase CT images.

In conclusion, the multi-image radiomics model outperformed radiomics models based on the individual 40, 70, 100 keV and IMD images in distinguishing between MBIC and NMBIC. The fusion model, which combines the radiomics model and NIC, demonstrated improved predictive performance for muscle invasion compared to both the DECT and radiomics models, although the difference was not statistically significant. This innovative method holds promise in unraveling the heterogeneity of muscle status in BCa and in providing personalized medical care for MBIC patients efficiently and scientifically.