Fully end-to-end automated bladder segmentation and risk assessment based on bladder morphology features for predicting upper urinary tract dysfunction

Abstract

Background:

This study aims to analyze bladder morphology features for upper urinary tract dysfunction (UUTD) risk assessment and develop a fully end-to-end (E2E) automated segmentation and risk identification algorithm.

Materials and methods:

A retrospective cohort of 604 neurogenic bladder patients undergoing video-urodynamics was analyzed. Bladder regions of interest (ROIs) were manually and automatically segmented. Thirty-six morphology features, including shape, Laplacian of Gaussian histograms, and fractal characteristics, were extracted from ROIs. After LASSO regression, selected predictive features were integrated into machine learning models. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis.

Results:

Based on manually segmented ROI, 28 (77.78%) morphology features exhibited significant differences between different bladder grades. Shape_Perimeter_to_Surface_ratio emerged as the most significant predictor of UUTD (AUC = 0.722, 95% confidence interval [CI]: 0.674–0.769). The support vector machine (SVM) model demonstrated the highest performance, with an AUC of 0.860 (95% CI: 0.820–0.899) in the training cohort and 0.830 (95% CI: 0.744–0.915) in the test cohort, which was significantly higher than the performance of using single features. However, incorporating clinical information did not further enhance the overall predictive performance of the model. To realize a fully E2E approach, a fully convolutional network with a ResNet101 backbone (FCN-ResNet101) was adopted for automated bladder segmentation (dice similarity coefficient [Dice] = 0.9848, mean intersection over union [mIoU] = 0.9705). Morphology features extracted from automated and manual segmentations showed a strong correlation (all r > 0.90). Similarly, machine learning models were developed, with SVM achieving the highest performance (AUC = 0.843, 95% CI: 0.801–0.885 in the training cohort; AUC = 0.815, 95% CI: 0.729–0.900 in the testing cohort).

Conclusion:

The bladder morphology features are robust predictors of UUTD. This fully automated, E2E bladder segmentation and UUTD risk prediction model demonstrated superior performance compared to traditional clinical parameters, providing a quantitative and objective tool for early risk assessment. Large-scale, multicenter prospective studies are required to validate its clinical applicability and robustness in the future.

HIGHLIGHTS

• Bladder morphology features are associated with the upper urinary tract dysfunction risk in neurogenic bladder patients.

• This fully automated bladder segmentation and risk prediction model can offer valuable guidance for clinical decision-making.

Introduction

Neurogenic bladder is a condition resulting from neurological disorders that impair the function of the bladder and/or urethra, leading to a range of lower urinary tract symptoms^[1]. Prolonged bladder dysfunction can cause serious complications such as upper urinary tract dysfunction (UUTD), including vesicoureteral reflux (VUR), hydronephrosis, and ureter dilation/tortuosity, and even renal function impairment, which, in severe cases, may be life-threatening. Therefore, early identification of the risk of UUTD is essential.

Previous studies have indicated a correlation between bladder morphology and UUTD^[2–5]. Over the past few decades, there have been frequent attempts to characterize the morphology of the neurogenic bladder, including the classic subjective appearance of the Christmas tree^[6]. However, most descriptions of bladder morphology remain qualitative and subjective, lacking standardized quantitative metrics. Moreover, no studies have developed a predictive model for UUTD based on lower urinary tract morphology. In this study, we analyzed bladder morphology features for UUTD risk assessment and developed an automated segmentation and risk identification algorithm (Fig. 1). This end-to-end (E2E) fully automatic approach enables the early detection of UUTD, offering valuable guidance for clinical decision-making. The work has been reported in line with the TITAN criteria^[7].

Figure 1.

The process of feature extraction, selection, and signature development is based on manual and automatic segmentation of the bladder.

Materials and methods

Ethical approval of the study protocol

This study was conducted in accordance with the Declaration of Helsinki and received ethical approval from the institutional review board of our hospital (202307401). The study was conducted and reported in accordance with the Standards for Reporting Diagnostic Accuracy Studies guidelines^[8].

Subjects

This retrospective study was conducted at the Urological and Pelvic Floor Rehabilitation Center of our institution. A total of 604 patients with lower urinary tract dysfunction, who were initially diagnosed and underwent video-urodynamics (VUDS) between 2016 and 2023, were included in the analysis. Patients who had undergone interventions that could potentially affect lower urinary tract function, such as the use of anticholinergic medications, bladder augmentation, or intravesical botulinum toxin injections, were excluded from the study. Additionally, patients with concurrent bladder tumors or bladder stones were also excluded. Cases with insufficient medical history or incomplete clinical data were further eliminated to ensure the integrity and reliability of the study.

VUDS were performed using a Laborie TRITON multi-channel urodynamic system (Laborie Medical Technologies, Canada). The procedures were conducted in strict accordance with the Good Urodynamic Practice guidelines recommended by the International Continence Society^[9]. Sterile iopromide (15%–20%) was infused at a temperature of 37°C and a rate of 20–30 mL/min. The following parameters were recorded: maximum bladder capacity (MBC), maximum detrusor pressure (MDP), bladder compliance, and post-void residual urine volume (PVR). VUR was graded according to the International Reflux Grading System^[10], and upper urinary tract dilation was assessed using the magnetic resonance urography grading system proposed by Liao^[1,11,12]. The presence of VUR, upper urinary tract dilation, and/or renal functional decompensation was considered indicative of UUTD.

Bladder segmentation

The region of interest (ROI) for the bladder was defined using both manual and automated segmentation methods. Two independent researchers defined the ROI on fluoroscopic images captured at maximum bladder filling status. To achieve an E2E workflow, we implemented and compared several deep-learning models for bladder segmentation, including fully convolutional networks with ResNet50 and ResNet101 backbones (Fcn_resnet50, Fcn_resnet101), DeepLabV3 networks with ResNet50, ResNet101, and MobileNetV3-Large backbones (Deeplabv3_resnet50, Deeplabv3_resnet101, deeplabv3_mobilenet_v3_large), and U-Net. The samples were stratified into a training cohort (70%) and a test cohort (30%). During the training phase, we employed selective cropping of sub-volumes from the images and corresponding labels to balance the distribution of positive and negative labels. Specific spatial dimensions and sample quantities were defined to ensure a balanced dataset. To further enhance the diversity of the training dataset, we applied online data augmentation techniques, including spacing adjustments and random cropping. These methods ensured that each training iteration utilized a varied set of images, improving the model’s generalization capability. We employed the DiceCELoss as the loss function, a hybrid formulation that integrates Dice Loss and Cross-Entropy Loss.

$$Loss=\frac{-}{N}\sum _{i=1}^N\left(\alpha log\left(\frac{2}{\frac{1}{\in +Dic{e}_i}+\frac{1}{\in +C{E}_i}}\right)+\left(1-\alpha \right)log\left(\in +Dic{e}_i\right)\right)$$ (1)

This combined approach capitalizes on the complementary strengths of both loss functions, effectively addressing critical challenges such as class imbalance and segmentation accuracy. The Adam optimizer was utilized with an initial learning rate of 1 × 10⁻³. The model was trained for 90 epochs, with an early stopping mechanism applied to prevent overfitting. During the prediction phase, scattered points – representing erroneously identified voxels – may appear across the image space. To evaluate the segmentation accuracy, two key metrics were utilized: the mean intersection over union (mIOU) and the dice similarity coefficient (DICE).

$$\text{\$Dice( gt,Pred) = {{2are{a\_{GT}}{{mathop ,,cap ,,nolimits^ }^}are{a\_{Pred}}} over {are{a\_{GT}} + are{a\_{Pred}}}}}$$ (2)

$$\text{\$mIOU(gt,Pred) = {{are{a\_{GT}}{{mathop ,,cap ,,nolimits^ }^}are{a\_{Pred}}} over {are{a\_{GT}}{{mathop ,,cup ,,nolimits^ }^}are{a\_{Pred}}}}}$$ (3)

By integrating both the DICE and mIOU metrics, the model’s segmentation performance was thoroughly assessed, providing a robust and comprehensive evaluation of its accuracy in predicting ROI.

Morphology features

A total of 36 quantitative features were extracted from the ROI to characterize the bladder morphology. These features were categorized into three groups: (1) shape features, (2) histogram features after applying Laplacian of Gaussian (LoG) filters, and (3) fractal-based features. Shape features represent the most fundamental category, characterizing the morphological characteristics of the bladder^[13]. The histogram features characterize the distribution and frequency of pixel intensities within the ROI. The LoG method involves the application of a Gaussian smoothing filter to mitigate noise in the images, followed by a Laplacian filter to enhance regions exhibiting rapid intensity variations. Consequently, this approach is widely employed for edge detection tasks. In this study, LoG filters were implemented at multiple scales (1.0, 1.5, 2.0, 2.5, 3.0, and 3.5 voxels) to ensure comprehensive analysis across varying spatial resolutions^[13]. Fractal-based features, encompassing fractal dimension, fractal signature dissimilarity, and lacunarity, are capable of characterizing irregular and self-similar structural patterns. These features collectively provide a comprehensive mathematical framework to quantify the intrinsic morphological complexity of the bladder. Detailed mathematical definitions of the adopted features can be found in previous studies^[13–16].

Signature construction

Morphology features were extracted from both manually and automatically segmented ROIs. For the manually segmented ROIs, only features exhibiting excellent stability (intraclass correlation coefficient > 0.75) were retained for further analysis. The feature selection process was conducted as follows: First, all features were normalized using the z-score standardization method. Subsequently, correlations between features were evaluated using Spearman’s correlation coefficient. For feature pairs with a correlation coefficient greater than 0.9, only one feature was retained to avoid redundancy. Then, the least absolute shrinkage and selection operator (LASSO) regression model was applied for morphology signature construction. To determine the optimal penalty parameter (λ), 10-fold cross-validation with minimum criteria was performed, where the final λ value corresponded to the minimum cross-validation error. Features with nonzero coefficients were retained and used to fit the regression model. These selected features were combined into a morphology score through a linear combination weighted by their respective model coefficients.

After performing LASSO feature selection, the final selected features were input into various machine learning models, including support vector machine (SVM), k-nearest neighbors, random forest, extra trees, extreme gradient boosting (XGBoost), light gradient boosting machine, multilayer perceptron, adaptive boosting, and decision tree. A 10-fold cross-validation approach was employed for constructing the risk signature model. The development process of the clinical signature followed a similar methodology to that of the morphology signature. Ultimately, a combined nomogram model incorporating both the clinical and morphology signatures was established for comprehensive interpretation and analysis. The performance of these models was evaluated in both the training and test cohorts. The predictive performance of the models was assessed using receiver operating characteristic (ROC) curves, while the DeLong test was employed to compare the area under the curve (AUC) among the models. The calibration efficiency of the nomogram was evaluated through calibration curves, and the Hosmer-Lemeshow test was used to assess its goodness of fit. Additionally, decision curve analysis was performed to determine the clinical utility of the model.

Results

Patient characteristics

A total of 604 participants were included in this study. Among them, 169 patients presented with UUTD, including 105 with VUR, 113 with upper urinary tract dilation, and 18 with impaired renal function. Baseline characteristics are summarized in Table 1. Significant between-group differences were observed in disease course (P = 0.025), MBC (P = 0.028), MDP (P < 0.001), and compliance (P < 0.001). However, no significant differences were observed in age, gender, or PVR between the two groups.

Table 1.

Baseline characteristics of enrolled patients

	All	Without UUTD	With UUTD	P
Predictors
Age	41.16 ± 16.14	40.91 ± 16.32	41.82 ± 15.70	0.489
Gender	277/327	196/239	81/88	0.452
Disease course	12.00 [6.00–21.00]	12.00 [5.00–20.00]	12.00 [5.00–21.00]	0.025
Maximal bladder capacity	300.50 [241.75–367.25]	307.00 [245.50–369.50]	300.00 [244.00–365.50]	0.028
Post-void residual volume	0.00 [0.00–234.25]	0.00 [0.00–226.50]	0.00 [0.00–236.25]	0.203
Maximum detrusor pressure	10.00 [5.00–20.00]	9.00 [4.00–18.00]	10.00 [4.00–19.00]	<0.001
Compliance	30.70 [15.64–65.40]	36.00 [19.09–78.13]	32.22 [16.06–71.63]	<0.001
Outcome components
VUR (±)	499/105	435/0	64/105	<0.001
Upper urinary tract dilation (±)	491/113	435/0	56/113	<0.001
Impaired renal function (±)	586/18	435/0	151/18	<0.001

UUTD, upper urinary tract dysfunction; VUR, vesicoureteral reflux.

Traditional bladder morphology grading method

First, four independent individuals classified 604 bladder images into four, three, and two categories, referring to previous studies^[2–5]. The results showed that the reliability was better for the binary classification, whereas in the three- and four-category classifications, reliability was relatively poor except for the consistent assessment of Grade 0 (Supplemental Digital Content Table 1, available at: http://links.lww.com/JS9/F36). This indicates that researchers have a relatively consistent understanding of what constitutes a smooth bladder, but there is considerable difficulty in reaching a consensus on what defines a non-smooth bladder, especially in grading its severity. Therefore, it is necessary to establish a quantifiable and interpretable bladder morphology parameter for clinical research.

Feature extraction, selection, and signature development based on manually segmented bladder

To achieve quantification of bladder morphology, we extracted morphology features of the bladder. After manually outlining the ROIs, a total of 36 features across three categories were extracted, including 9 shape features, 24 LoG histogram features, and 3 fractal-based features. 28 (77.78%) of the morphology features exhibited significant differences between Grade 0 and Grade 1, with Shape_Perimeter_to_Surface_ratio, LoG_Mean_1, and LoG_Skewness_2.5 being the most notable (Fig. 2A–C, Supplemental Digital Content Table 2, available at: http://links.lww.com/JS9/F37). In the Grade 1 group, a reduction in elongation and minor axis length was observed, indicating a shift toward a more slender and deformed bladder morphology. The decrease in sphericity, accompanied by a significantly increased perimeter-to-surface area ratio, reflects greater boundary irregularity and morphological complexity. Regarding the LoG histogram features, the Grade 1 group showed more negative mean values, higher standard deviations, lower absolute skewness, and reduced kurtosis. These findings suggest stronger and more heterogeneous edge responses, indicative of increased textural complexity and reduced boundary smoothness. Among the fractal features, only lacunarity exhibited a marginally significant difference, indicating irregular edge or texture distribution, which may reflect the presence of voids, cracks, or structural disruptions. We further investigated the correlations between these morphological features and urodynamic parameters (Supplemental Digital Content Figure 1, available at: http://links.lww.com/JS9/F44). These analyses indicate that features reflecting increased bladder irregularity and decreased surface smoothness, such as a higher perimeter-to-surface area ratio, lower sphericity, shorter minor axis length, more negative mean values, reduced absolute skewness and kurtosis, and elevated lacunarity, are associated with reduced MBC, supporting their potential functional significance. Since researchers have not reached a consensus on other grading methods, differences in features under additional classification criteria were not explored.

Figure 2.

Feature selection and signature development based on manually segmented bladder. (A) Comparison of Shape_Perimeter_to_Surface_ratio, LoG_Mean_1, and LoG_Skewness_2.5 between Grade 0 and Grade 1. (B) The process of constructing a morphology score using the least absolute shrinkage and selection operator (LASSO) model. (C) Receiver operating characteristic (ROC) curves of single morphology and clinical features. (D) ROC curves of morphology features based on different models in the training and test cohorts. (E) ROC curves, decision curves, calibration curves, and nomograms of clinical, morphology, and combined models based on support vector machine (SVM) in the training and test cohorts.

We further developed a UUTD prediction model based on morphological features (Fig. 2D–F). A total of six morphology features were selected to construct the morphology score using an LASSO LR model. Among these features, Shape_Perimeter_to_Surface_ratio contributed the most to the diagnosis of UUTD. The morphology score is defined as follows: Morphology score = 0.2798 − 0.0356 * Shape Elongation + 0.2419 * Shape Perimeter to Surface ratio + 0.1678 * Shape Pixel Surface − 0.0604 * LoG Std_1 + 0.0992 * LoG Std 3.5 − 0.0597 * LoG Kurtosis 1.

We evaluated the predictive accuracy of using a single morphology or clinical feature for UUTD (Fig. 2G–H). The results showed that the highest ROC was achieved by Shape_Perimeter_to_Surface_ratio, with an AUC of 0.722 (95% CI: 0.674–0.769), and compliance with an AUC of 0.633 (95% CI: 0.583−0.684), respectively. For comparison, we also assessed the performance of established trabeculation grading systems proposed by Cho^[2] and Selby^[4]. The AUCs for their four-, three-, and two-tier grading schemes were 0.681 (95% CI: 0.639–0.723), 0.695 (95% CI: 0.651–0.739), and 0.700 (95% CI: 0.656–0.745), respectively, which were comparable to single morphology or clinical feature (Supplemental Digital Content Figure 2, available at: http://links.lww.com/JS9/F44). To further improve prediction accuracy, we have built multiple machine learning models for UUTD. Supplemental Digital Content Table 3, available at: http://links.lww.com/JS9/F38, provides an overview of all the models used to predict UUTD. Among several classifiers evaluated, the SVM model was selected as the final model due to its superior generalization performance, as evidenced by consistent AUC values in the training (0.860, 95% CI: 0.820–0.899) and testing (0.830, 95% CI: 0.744–0.915) cohorts (Fig. 2I–J). In contrast, tree-based models such as XGBoost, while achieving a higher training AUC (0.902, 95% CI: 0.872–0.933), exhibited a notable performance drop on the test set (AUC = 0.764, 95% CI: 0.673–0.855), suggesting potential overfitting (Supplemental Digital Content Table 3, available at: http://links.lww.com/JS9/F38).

We further explored the integration of morphology features and clinical features (Fig. 2K–Q, Supplemental Digital Content Table 4, available at: http://links.lww.com/JS9/F39). The combined model (AUC = 0.823, 95% CI: 0.736–0.911) demonstrated superior predictive performance compared with the clinical model alone (AUC = 0.650, 95% CI: 0.541–0.759, P < 0.001 by DeLong test). However, there was no significant difference compared with the morphology signature (AUC = 0.830, 95% CI: 0.744–0.915, P = 0.262 by DeLong test). These findings suggest that the morphology signature possesses strong independent predictive capability, whereas the clinical model alone has relatively poor performance. Incorporating clinical information did not further enhance the overall predictive performance of the model.

Feature extraction, selection, and signature development based on automatically segmented bladder

However, the above approach requires manual delineation of the bladder ROI. To achieve E2E risk identification, we explored the signature development based on automatic bladder segmentation using multiple deep learning algorithms. All models demonstrated high segmentation accuracy, with Dice scores exceeding 96% and mIoU values exceeding 94% (Supplemental Digital Content Table 5, available at: http://links.lww.com/JS9/F40). Among them, fcn_resnet101 achieved the highest performance, yielding a Dice score of 0.9848 and an mIoU of 0.9705. Based on these results, fcn_resnet101 was selected as the segmentation model for subsequent analyses. Representative examples of the automated segmentation results are presented in Figure 3A.

Figure 3.

Representative examples (A) and correlation (B) of the manually and automatically segmented bladder segmentation.

We explored the correlation of morphology features obtained from manual and automatic segmentation. As shown in Figure 3B, the two methods exhibited strong consistency, with all feature correlation coefficients exceeding 90%. Similarly, we analyzed the differences between Grade 0 and Grade 1 (Fig. 4A–C, Supplemental Digital Content Table 6, available at: http://links.lww.com/JS9/F41) and examined the association between urodynamic parameters (Supplemental Digital Content Figure 3, available at: http://links.lww.com/JS9/F44). The results aligned with those from manual segmentation, and the significant features identified in manual segmentation all remained notable in automatic segmentation.

Figure 4.

Feature selection and signature development based on automatically segmented bladder. (A) Comparison of Shape_Perimeter_to_Surface_ratio, LoG_Mean_1, and LoG_Skewness_2.5 between Grade 0 and Grade 1. (B) The process of constructing a morphology score using the least absolute shrinkage and selection operator (LASSO) model. (C) Receiver operating characteristic (ROC) curves of single morphology and clinical features. (D) ROC curves of morphological features based on different models in the training and test cohorts. (E) ROC curves, decision curves, calibration curves, and nomograms of clinical, morphology, and combined models based on support vector machine (SVM) in the training and test cohorts.

Similarly, we developed the prediction model based on morphology features (Fig. 4D–F). A total of eight morphology features were selected. Consistent with the results of manually segmented bladders, the Shape_Perimeter_to_Surface ratio contributed the most to UUTD diagnosis, achieving the highest AUC (AUC = 0.660, 95% CI: 0.608–0.712) among all individual morphology features (Fig. 4G-H). The morphology score is defined as follows: Morphology score = 0.2798–0.0426 * Shape Elongation + 0.2051 * Shape Perimeter to Surface ratio + 0.1839 * Shape Pixel Surface – 0.1360 * LoG Std 1 + 0.1920 * LoG Std 3.5–0.0677 * LoG Kurtosis 1 − 0.0785 * LoG Kurtosis 1.5 + 0.0452 * LoG Kurtosis 3.5.

Among all machine learning models, the SVM model also demonstrated the highest performance, with an AUC of 0.843 (95% CI: 0.801–0.885) in the training cohort and 0.815 (95% CI: 0.729–0.900) in the test cohort, which was significantly higher than the performance of the using single features (Fig. 4I–J, Supplemental Digital Content Table 7, available at: http://links.lww.com/JS9/F42). Similar to the results of manually segmented bladders, the morphology features (AUC 0.815, 95% CI: 0.729–0.900) outperformed the clinical features significantly (AUC 0.566, 95% CI: 0.449–0.683, P < 0.001 by DeLong test), and incorporating clinical information did not further improve the overall predictive performance (AUC 0.812, 95% CI: 0.725–0.898, P = 0.334 by DeLong test, Figure 4K–Q, Supplemental Digital Content Table 8, available at: http://links.lww.com/JS9/F43).

Discussion

In this study, we aim to quantify bladder morphology and explore its association with the risk of UUTD. The morphology of the neurogenic bladder was first described in detail by Friedenberg and Ney in 1960, including presentations such as the large atonic bladder, hourglass deformity, pine tree bladder, and severely contracted bladder^[17]. Subsequently, studies have associated bladder deformities observed in cystographic images with UUTD^[18]. However, most descriptions of bladder morphology remain qualitative and subjective, lacking clear boundaries between different grades, particularly at the higher levels. For example, Sebly et al^[4] proposed a subjective evaluation method that categorizes the bladder into three groups: smooth bladder, bladder with shallow cellules/diverticula/saccules, and bladder with deep cellules/diverticula/saccules. Cho et al^[2] proposed a similar four-tier classification system, including bladder with no trabeculation, mild trabeculation, moderate trabeculation, and severe trabeculation. Therefore, there remains an urgent need to develop objective measures to define bladder morphology in neurogenic bladder, and to establish practical correlations between morphological changes and upper urinary tract deterioration.

Some studies have attempted to quantify bladder deformation. Given that the neurogenic bladder often becomes vertically elongated, the height-to-width ratio (HWR) has been proposed as a metric to define morphological abnormalities^[19,20]. However, changes in HWR associated with the typical high-pressure bladder may take time to manifest, potentially delaying diagnosis. Wang et al^[21] introduced the concept of bladder dispersion, which quantifies the area difference derived from a Butterworth filter to assess the smoothness of the bladder. However, the aforementioned methods focus on single indicators of bladder deformation and still rely on manually delineating or measuring bladder parameters, which introduce potential bias. Moreover, no studies to date have provided clinically applicable risk prediction models. Therefore, developing a fully automated segmentation and risk identification algorithm is crucial, as it enables the early identification of UUTD and provides valuable guidance for clinical decision-making.

In this study, we extracted the shape features of the bladder, which represent fundamental bladder geometric measurements, including perimeter, surface area, diameters in different directions, and sphericity. We found that several indicators showed significant differences across different bladder grades. These include Elongation, which reflects the relationship between the two largest principal components of the ROI shape, where lower values suggest a more elongated bladder; Sphericity, which measures how closely the shape resembles a perfect sphere, with lower values indicating greater deviation from a spherical form. The clinical implications revealed by these indicators are similar to those of HWR^[19,20] or the formula proposed by Chang et al^[5], which reflects the degree of bladder elongation or its deviation from a spherical shape, thereby indicating the extent of bladder deformation. The Perimeter-to-Surface Ratio, also a statistically significant indicator, quantifies shape complexity, with higher values generally reflecting a more irregular bladder morphology. However, while these individual indicators demonstrated statistically significant differences among bladder grades, their standalone predictive performance for UUTD remained suboptimal. This suggests that bladder deformation is a multifactorial phenomenon that may not be fully captured by any single geometric descriptor.

In this study, we incorporated additional quantitative indicators to capture a broader range of bladder characteristics, aiming to enhance the sensitivity and specificity of morphological assessment. The LoG method is a highly effective approach for edge feature extraction and can be used to quantitatively assess the degree of bladder trabeculation. It operates by first applying Gaussian filtering to suppress image noise, followed by the application of the Laplacian operator to compute the second-order derivative, thereby enhancing regions with abrupt intensity changes. Edge locations are subsequently identified through zero-crossing detection, allowing for precise delineation of structural irregularities. In this study, we observed that higher-grade bladders showed lower Mean values, which reflects a more substantial decrease in pixel intensities after LoG filtering, suggesting greater local structural heterogeneity. Additionally, the higher Std values reflect more pronounced pixel intensity variations following LoG filtering, implying enhanced edge contrast and increased structural complexity within the images. This method is designed to quantitatively assess bladder morphology features and shares conceptual similarities with the work of Wang et al ^[21], which characterized bladder features by calculating the area difference derived from a Butterworth filter. In addition, we investigated fractal features that characterize the complexity and structural variability of the bladder contour across multiple spatial scales. Our findings indicate that higher-grade bladders exhibit increased Lacunarity, with a P-value approaching significance, suggesting greater structural heterogeneity and a higher likelihood of irregular or trabeculated regions. However, no significant differences were observed in Fractal Dimension or Dissimilarity, indicating that these parameters may not effectively distinguish between different bladder grades.

However, we found that these individual indicators were insufficient for robust UUTD prediction, as they could not fully capture the complex characteristics of the bladder, with the highest AUC reaching only 0.722, followed by 0.662 for the next best indicator. Therefore, we integrated multiple morphology indicators to develop a machine learning-based predictive model, which achieved an AUC of 0.860 in the training cohort and 0.830 in the test cohort. This integrated approach demonstrated significantly enhanced predictive performance compared to using individual indicators alone. In the predictive model, features associated with shape irregularity, such as the Perimeter-to-Surface Ratio, consistently ranked among the most important. Additionally, features derived from LoG-filtered images demonstrated high significance, highlighting the value of integrating multi-scale information.

Interestingly, when clinical features were incorporated into the model, no further improvement in predictive performance was observed. Although we observed that patients with UUTD had a longer disease course, smaller MBC, higher MDP, and poorer compliance, these clinical parameters alone demonstrated poor predictive performance for UUTD, with AUC values ranging from 0.514 to 0.633. The nomogram further illustrates that clinical parameters contribute only a small proportion of the overall score. Previous studies have shown that there is still no consensus regarding which clinical features can predict UUTD, and their predictive performance remains inconsistent^[22–24]. In clinical practice, although MDP >40 cmH₂O is often used as an empirical predictor of UUTD^[25,26], it lacks robust validation. In our study, MDP alone achieved only a modest AUC of 0.594, further highlighting its limitations as a standalone predictor. In contrast, bladder morphological parameters emerged as strong independent predictors of UUTD, distinct from clinical parameters, demonstrating superior predictive performance. This suggests that bladder morphology may partially reflect underlying pathophysiological changes and capture structural alterations resulting from cumulative and multifactorial processes that are not directly measurable by clinical parameters alone. It may represent an additional dimension that enriches the understanding of bladder dysfunction. However, this result should be interpreted with caution. The lack of improvement when combining clinical and morphological features may be due to the high variability and complexity of neurogenic bladder patients. Our study also included only a limited set of clinical parameters, which might not represent the full clinical picture. Additionally, the small sample size and retrospective design could limit the ability to detect subtle effects. Overfitting or less effective feature integration may also reduce the benefit of combining data types. Future studies with more detailed clinical data and larger samples are needed to better evaluate and improve the model.

Previous research has largely relied on manual expertise, including the quantitative indicators proposed by Wang et al, which also depend on manual delineation of the bladder. This approach is time-consuming, labor-intensive, and prone to variability both between observers and within the same observer over time. Therefore, we aimed to develop a fully E2E automated segmentation and risk identification algorithm to overcome these limitations. Utilizing the Fcn-ResNet101 architecture, we successfully achieved automatic and precise bladder ROI segmentation from fluoroscopic images. The bladder morphology parameters extracted from both automated and manual segmentations exhibited a strong correlation, indicating that the automated method effectively preserved essential bladder edge features critical for morphological assessment. Subsequently, we constructed a UUTD prediction model, which demonstrated comparable accuracy to the manual segmentation-based approach (AUC = 0.843 in the training cohort and 0.815 in the test cohort). This fully automated pipeline allows a single bladder fluoroscopic image to be directly input into the system without any manual intervention, enabling rapid and objective UUTD risk prediction. The E2E workflow significantly reduces analysis time and potential inter-operator variability, making it highly suitable for routine clinical application. Moreover, it facilitates continuous monitoring and early detection of UUTD in patients with neurogenic bladder, potentially guiding timely intervention and improving patient outcomes.

Our study is the first to analyze bladder morphology characteristics for UUTD risk assessment and develop an automated segmentation and risk identification algorithm. However, our study also has some limitations. First, our model showed strong performance in internal validation; however, its generalizability is limited by the single-center, retrospective study design. The use of data from a single institution may introduce biases related to imaging protocols, patient demographics, and comorbidities, which could reduce predictive accuracy in other clinical settings. Moreover, the model was trained and tested exclusively on neurogenic bladder patients without stratification by underlying etiology, disease severity, treatment history, or age, which may affect bladder morphology and function. This limits the applicability of the model to other bladder disorders (e.g., bladder outlet obstruction, posterior urethral valves) and neglects intra-group heterogeneity. Future multicenter prospective studies incorporating diverse populations and detailed subgroup analyses are needed to externally validate and refine the model, enabling disease- and patient-specific adaptations. Until such validation is completed, our results should be interpreted as preliminary but promising evidence for fully automatic UUTD risk stratification. Second, a key limitation of this study is the use of a composite UUTD outcome encompassing VUR, upper tract dilation, and renal impairment. Although this definition aligns with established clinical criteria^[16,27], the heterogeneity among these pathologies may reduce the interpretability of the model. Moreover, a substantial proportion of patients exhibited overlapping features (e.g. both VUR and dilation), which limited the feasibility of conducting meaningful stratified performance analyses. Future studies with larger and more phenotypically stratified cohorts are needed to develop subtype-specific models and enhance clinical applicability. Third, from a translational perspective, future work should focus on ensuring the model’s compatibility with clinical imaging systems, including Digital Imaging and Communications in Medicine standard support and seamless Picture Archiving and Communication System integration. Furthermore, optimizing inference speed for real-time or near-real-time use is important to enable practical deployment in busy clinical environments and support timely decision-making. Fourth, although VUDS is considered the optimum procedure for assessing neuro-urological disorders, they are not routinely performed in all patients with neurogenic bladder in current clinical practice. However, in high-risk populations, particularly those requiring evaluation of lower urinary tract morphology and potential VUR, VUDS is typically recommended^[28]. The proposed pipeline is intended to augment clinical risk stratification in this high-risk subset, especially in cases where fluoroscopic bladder images are already acquired as part of routine care. Rather than replacing standard urodynamic evaluation, our model is envisioned as a complementary decision-support tool that offers objective, image-based risk assessment to inform further diagnostic steps or early intervention.

Conclusion

The bladder morphology features are robust predictors of UUTD. This study successfully established a fully automated, E2E pipeline for bladder segmentation and UUTD risk prediction based on morphology features. The proposed signature outperformed traditional clinical parameters, offering a quantitative and objective tool for early risk assessment. However, the single-institution retrospective design may limit the generalizability of our findings. Therefore, large-scale, multicenter prospective studies are essential to further validate the clinical applicability and robustness of the proposed model.

Ethical approval

The study has been approved by the ethics committee of our institution and complies with the Declaration of Helsinki.

Consent

This study was a retrospective study, and the data were processed to remove relevant private information of patients, so the ethics committee agreed to exempt informed consent.

Sources of funding

This work was supported by the National Key Research and Development Program of China (2023YFC2605301), National Natural Science Foundation of China (82470810, 82500940), Beijing Natural Science Foundation (7254460), Beijing Postdoctoral Research Foundation (2025-ZZ-194), and Fundamental Research Funds for Central Public Welfare Research Institutes (2023CZ-1, CRSI2025ZD-1-1, CRSI2025QA-2). The funders had no involvement in the study design, data collection and analysis, writing, or submission.

Author contributions

L.L. designed and oversaw the study. Z.Z., T.H., J.W., T.X., P.Z., X.L., G.L., J.J., and Y.L. collected the clinical data. Z.Z. extracted the margin features, developed the model, and wrote the manuscript. L.L. made major contributions to manuscript editing. All authors have directly accessed and verified the underlying data in the manuscript. All authors confirmed that they had full access to all the data and accept responsibility to submit for publication.

Conflicts of interest disclosure

Not applicable.

Research registration unique identifying number (UIN)

This study was publicly registered in the Chinese Clinical Trial Registry (ChiCTR2400083598, https://www.chictr.org.cn/showproj.html?proj=207736).

Guarantor

Limin Liao.

Provenance and peer review

Not commissioned, externally peer-reviewed.

Data availability statement

The data are available from the authors on reasonable request.