Image segmentation method and device for bladder tumor, medium and program product

Abstract

The invention belongs to the field of intelligent medical treatment, and particularly relates to a bladder tumor image segmentation method and device, a medium and a program product. The method comprises the following steps: S101, acquiring an image of a bladder tumor patient; s102, inputting the image into a 2D U-Net for tumor segmentation to obtain a 2D segmentation result; s103, traversing the shape of the tumor based on the 2D segmentation result, inputting the image into 3D U-Net for tumor segmentation to obtain a 3D segmentation result if the shape of the current tumor is regular during traversing, and taking the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, the segmentation result of the current tumor in the 2D segmentation result is obtained, and after traversal is completed, the segmentation results of the multiple tumors are output after merging. Through the hybrid 2D-3D method, the segmentation performance can be optimized by using the complementary advantages of the two methods, and a better segmentation result is obtained.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medicine, and more particularly, to an image segmentation method, device, medium and program product for bladder tumors. Background Art

[0002] Bladder cancer (BCa) is one of the most common malignant tumors of the urinary system, having a significant impact on both incidence and quality of life. According to global cancer statistics, BCa ranks ninth globally in terms of new cases and deaths, with approximately 614,000 new cases and 220,000 deaths in 2022. BCa is histologically classified into non-muscle invasive bladder cancer (NMIBC) and muscle invasive bladder cancer (MIBC), with 75% of new cases being classified as NMIBC. Despite progress in surgical resection, immunotherapy, and chemotherapy, the 5-year survival rate for advanced BCa, especially MIBC, remains low, highlighting the need for better diagnostic and treatment tools. Precise delineation of the tumor boundary and bladder wall layers enables clinicians to assess the size, location, and invasiveness of the tumor, which is crucial for the diagnosis, staging, and treatment planning of BCa. For example, NMIBC and MIBC require different treatment strategies, with the latter typically requiring radical cystectomy. In addition, accurate segmentation supports the monitoring of treatment response and recurrence, which is crucial for improving patient prognosis. However, the hollow structure of the bladder results in dynamic changes in its position, shape, and volume, while complex noise and artifacts in medical images complicate segmentation. Moreover, manual segmentation by radiologists is both time-consuming and subjective, and is susceptible to inter-observer variability, highlighting the need for automated and reliable segmentation methods. Deep learning (DL), with its ability to automatically extract relevant features without manual selection, has become a powerful method for medical image segmentation. Different from traditional machine learning (ML) algorithms, DL can efficiently process large-scale medical images, leveraging substantial computational resources to improve accuracy and robustness. It can analyze complex patterns and is thus particularly suitable for medical imaging tasks such as bladder segmentation.

[0003] In the DL-based segmentation architecture, U-Net is one of the most effective methods for image segmentation in the fully convolutional architecture, laying the foundation for many advanced medical image analysis techniques. Multiple studies have explored the application of U-Net in bladder segmentation. Dolz et al. proposed an extended U-Net model with a gradually increasing dilation rate to enhance the background information for bladder wall and tumor segmentation. The model was trained on T2-weighted MRI images of 60 BCa patients, and the Dice similarity coefficient (DSC) scores for the inner wall, outer wall, and tumor of the bladder were 0.98, 0.84, and 0.69, respectively. Liu et al. proposed an improved U-Net model (PiPNet), which was evaluated on T2-weighted MRI images of 47 BCa patients. The DSC values of PiPNet for the outer wall and tumor of the bladder were 0.89 and 0.95, respectively. Lee et al. proposed RDAU-Net to segment the bladder and lesions in 2136 CT images, with the accuracy rates reaching 96% and 93% respectively, and the training time was shortened by 44%. Jang et al. proposed an improved U-Net model for urogenital system segmentation using non-enhanced CT images. After being trained on 814 CT scans, the DSC value of the model for the bladder organ reached 0.903. These studies have emphasized the effectiveness of U-Net-based models in bladder segmentation. However, they are often limited by a small dataset, potential overfitting, and lack of generality, which may limit their clinical applicability. The original U-Net architecture was introduced as a 2D model and has been widely used in biomedical imaging, especially in medical segmentation tasks. However, 2D U-Net processes individual slices, which may lead to inconsistent volume structures. To address this limitation, Çiçek et al. extended the architecture to 3D U-Net, incorporating the spatial relationships of depth, height, and width, making it more suitable for volume segmentation. Both 2D and 3D U-Net have been widely used in medical imaging, especially for organ and tumor segmentation. However, there has been no systematic study comparing their performance in bladder segmentation. This study aims to systematically evaluate the segmentation performance of 2D and 3D U-Net in the simultaneous segmentation of the bladder wall and tumor. In addition, subgroup analysis is performed based on the clinical characteristics of the tumor to determine the optimal model selection under different clinical conditions, providing a reference for their potential clinical applications. Summary of the Invention

[0004] In view of the above problems, the present invention provides an image segmentation method for bladder tumors, which selects a more suitable segmentation model between 2D U-Net and 3D U-Net according to different clinical scenarios, and determines the optimal model selection under different clinical conditions.

[0005] This application (in the first aspect) discloses an image segmentation method for bladder tumors, including:

[0006] S101: Obtain the images of bladder tumor patients;

[0007] S102: Input the image into a 2D U-Net for tumor segmentation to obtain a 2D segmentation result;

[0008] S103: Traverse the shape of the tumor based on the 2D segmentation result. When traversing, if the current tumor shape is regular, input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result, and take the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, take the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

[0009] The present application (second aspect) discloses an image segmentation method for bladder tumors, and the method includes:

[0010] S301: Obtain the image of a bladder tumor patient;

[0011] S302: Input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result;

[0012] S303: Traverse the shape of the tumor based on the 3D segmentation result. When traversing, if the current tumor shape is regular, take the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, input the image into a 2D U-Net for tumor segmentation to obtain a 2D segmentation result, and take the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

[0013] Further, the method further includes, based on the 2D segmentation result, determining whether the number of tumors is greater than 1. If it is greater than 1, input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result;

[0014] Traverse the shape of the tumor based on the 3D segmentation result. When traversing, if the current tumor shape is regular, take the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, take the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

[0015] Further, the method further includes: simultaneously obtaining the time limit of the segmentation task. If the time limit is lower than the first time threshold, input the image into a 3D U-Net for segmentation to obtain a segmentation result and then output the segmentation result, otherwise input the image into a 2D U-Net or 3D U-Net for tumor segmentation and then traverse the shape of the tumor.

[0016] Further, the method further includes: judging the sphericity of the tumor shape. If the sphericity of the tumor shape meets the requirement of being higher than the sphericity threshold, it is judged as a regular shape, otherwise it is judged as an irregular shape;

[0017] Optionally, the sphericity threshold is constructed based on the sphericity of the training set;

[0018] Optionally, a shape discrimination classifier is used to determine whether the shape of the tumor is regular. The construction method of the shape discrimination classifier is to input the tumor images of the training set and the annotations of whether they are regular into the classifier for training.

[0019] Furthermore, the method further includes: inputting the image into a 2D U-Net for bladder wall segmentation to obtain a bladder wall segmentation result, and performing tumor segmentation based on the image and the bladder wall segmentation result.

[0020] Furthermore, the method further includes: the model architecture of the 2D U-Net consists of 7 downsampling and 7 upsampling stages, and the basic number of features is 32;

[0021] Optionally, the model architecture of the 3D U-Net consists of 5 downsampling and 5 upsampling stages, and the basic number of features is 32;

[0022] Optionally, the test-time augmentation technique is adopted in the training processes of the 2D U-Net and the 3D U-Net;

[0023] Optionally, the image is an enhanced CT image.

[0024] The third aspect of the present application discloses an image segmentation system for bladder tumors, including:

[0025] An acquisition module 201: used to acquire the images of bladder tumor patients;

[0026] A primary segmentation module 202: used to input the image into a 2D U-Net for tumor segmentation to obtain a 2D segmentation result;

[0027] A traversal adjustment output module 203: based on the 2D segmentation result, traverse the shape of the tumor. During the traversal, if the current tumor shape is regular, input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result, and select the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, select the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

[0028] The present application (fourth aspect) discloses an image segmentation system for bladder tumors, including:

[0029] A second acquisition module 301: used to acquire the images of bladder tumor patients;

[0030] A second primary segmentation module 302: used to input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result;

[0031] Second traversal adjustment output module 303: It is used to traverse the shape of the tumor based on the 3D segmentation result. When traversing, if the current tumor shape is regular, the segmentation result of the current tumor in the 3D segmentation result is adopted; if the tumor shape is irregular, the image is input into the 2D U-Net for tumor segmentation to obtain the 2D segmentation result, and the segmentation result of the current tumor in the 2D segmentation result is adopted. After the traversal is completed, the segmentation results of multiple tumors are merged and output.

[0032] The fifth aspect of the present application discloses a computer device, which includes: a memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, it is used to execute the steps of the above method.

[0033] The sixth aspect of the present application discloses a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0034] The seventh aspect of the present application discloses a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0035] The present application has the following beneficial effects:

[0036] (1) The hybrid 2D-3D method of the present application can utilize the complementary advantages of the two to optimize the segmentation performance and obtain better segmentation results;

[0037] (2) Through research, the present application determines how to select a segmentation model in different scenarios to achieve the best effect, so that a more suitable model can be selected from the widely used models during clinical application, thereby achieving an improvement in the overall effect in bladder cancer tumor segmentation. Description of the Drawings

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0039] Figure 1 It is a schematic flowchart of the method provided in the first aspect of the embodiments of the present invention;

[0040] Figure 2 It is a schematic diagram of the program product provided in the third aspect of the embodiments of the present invention;

[0041] Figure 3 It is a schematic diagram of the computer device provided in the embodiments of the present invention;

[0042] Figure 4 It is a schematic diagram of the architecture of an exemplary computing device provided by an embodiment of the present invention;

[0043] Figure 5 It is a schematic diagram of a storage medium provided by an embodiment of the present invention;

[0044] Figure 6 It is a schematic diagram of the process of segmenting tumors by a 3D U-net and a 2D U-net applying the test-time augmentation technique provided by an embodiment of the present invention;

[0045] Figure 7 It is a comparison of the bladder tumor segmentation performance of a 2D U-Net and a 3D U-Net in internal (A–C) and external (D–F) test queues provided by an embodiment of the present invention. The box plots show the Dice similarity coefficient (DSC) (A, D), average surface distance (ASD) (B, E), and 95% Hausdorff distance (95% HD) (C, F). A higher DSC indicates better segmentation accuracy, while lower ASD and 95% HD indicate higher boundary accuracy. In both queues, the 2D U-Net generally obtained a higher DSC but showed greater fluctuations in ASD and 95% HD;

[0046] Figure 8 It is the correlation between the Dice similarity coefficient (DSC) of a 2D U-Net (blue) and a 3D U-Net (orange) and radiomic shape features provided by an embodiment of the present invention; among them, the scatter plot shows the relationship between DSC and tumor shape features, including elongation rate (A), flatness (B), minimum axis length (C), surface area to volume ratio (D), maximum two-dimensional diameter (E), mesh volume (F), sphericity (G), and voxel volume (H). The 2D U-Net showed stronger correlations in terms of elongation rate, flatness, and sphericity, while the 3D U-Net was more affected by irregular morphologies. The surface area to volume ratio was negatively correlated with DSC, indicating more challenging segmentation in highly irregular tumors;

[0047] Figure 9 It is a flow chart of the enrollment and screening of research subjects provided by an embodiment of the present invention;

[0048] Figure 10 It is a schematic diagram of the process of 2D-3D selection provided by an embodiment of the present invention;

[0049] Figure 11 It is a schematic diagram of the process of 2D-3D selection provided by an embodiment of the present invention. Detailed implementation manners

[0050] To enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0051] In some processes described in the specification, claims and the above-mentioned drawings of the present invention, a plurality of operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as S101, S102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0053] Figure 1 It is a schematic flow chart of an image segmentation method for evaluating bladder tumors provided by an embodiment of the present invention. Specifically, the method includes the following steps:

[0054] S101: Obtain the images of bladder tumor patients;

[0055] S102: Input the images into a 2D U-Net for tumor segmentation to obtain a 2D segmentation result;

[0056] S103: Traverse the shape of the tumor based on the 2D segmentation result. When traversing, if the current tumor shape is regular, input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result, and take the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, take the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

[0057] The method of this application is obtained based on the following research.

[0058] The segmentation of bladder cancer (BCa) is crucial for accurate diagnosis, staging, and treatment planning. The study aimed to systematically compare the performance of 2D U-Net and 3D U-Net in bladder wall and tumor segmentation, with a focus on different tumor subgroups and radiomics features. Methods: Enhanced CT scans of 435 BCa patients from two centers were retrospectively studied and divided into a training group (n = 325), an internal test group (n = 72), and an external test group (n = 38). Both models were trained using the nnU-Net framework and evaluated using the Dice similarity coefficient (DSC), average surface distance (ASD), and 95th percentile Hausdorff distance (95% HD). Subgroup analysis was performed based on tumor characteristics, including size, multiplicity, muscle invasion, and shape-based radiological features. Results: For bladder wall segmentation, 3D U-Net performed slightly better than 2D U-Net in the internal test cohort but showed lower generalization in the external cohort. In tumor segmentation, 2D U-Net consistently obtained higher DSC scores, especially for small tumors (< 0.001). For large tumors (> 3 cm), both models performed well, but 2D U-Net maintained higher accuracy (P = 0.028). 3D U-Net showed slightly better performance in multi-tumor segmentation, with lower ASD and significantly faster inference time (P < 0.001). Radiomic feature analysis showed that 2D U-Net performed better on irregular tumors, while 3D U-Net was more effective for spherical lesions. Conclusion: 2D U-Net demonstrated excellent segmentation accuracy, especially for small, single, and irregular tumors, making it the preferred choice for early BCa and NMIBC. 3D U-Net has a faster inference time and volume continuity, which may be beneficial for high tumor burden cases or intraoperative applications. These findings emphasize the importance of selecting a segmentation model based on tumor characteristics, and future studies should explore hybrid architectures to optimize performance.

[0059] The details are as follows:

[0060] In an approved two - center retrospective study, the ethics committee waived the requirement for written informed consent due to its retrospective design (number I - 22PJ887). All procedures complied with the Declaration of Helsinki (2013 version). The dataset included patients pathologically diagnosed with BCa between October 2021 and August 2023. All patients received enhanced abdomino - pelvic CT scans within two weeks before surgery. The exclusion criteria were as follows: (1) Poor image quality, including severe motion artifacts, high noise levels, or insufficient bladder filling, which hindered assessment; (2) Only thick - slice enhanced CT images were provided; (3) No macroscopically visible tumors in the CT scan. A total of 435 patients were included, with 325 in the training set, 72 in the internal test set, and 38 in the external test set. The flow chart is as shown in Figure 9 Figure [1]. The imaging protocol CT images were acquired using scanners from different institutions. The dataset included portal CT images with slice thicknesses of 0.625 and 1 mm. Patients in our hospital were scanned using a GE Discovery CT (GE Healthcare) or a Siemens Somatom Definition Flash CT (Siemens Healthineers), with scanning parameters: tube voltage = 120 kVp, rotation time = 0.5 s, collimation = 128×0.6 mm, pitch = 0.9 mm, slice thickness = 1 mm, and contrast agent injection protocol of 100 mL of iopamidol 370 mg / mL at an injection rate of 4 - 4.5 mL / s, and images were acquired in the arterial phase (30 s), portal phase (50 s), and delayed phase (240 s). The external examination center used a Siemens SOMATOM Perspective 64 - slice scanner with a detector combination of 0.625 mm×64, tube voltage of 120 kV, tube current of 300 mAs, pitch of 0.984 mm, and enhanced scanning parameters consistent with our hospital. Ground truth creation A radiologist with four years of experience in urogenital imaging manually outlined the tumors and bladder wall using 3DSlicer (https: / / www.slicer.org). A senior radiologist with 14 years of experience then reviewed and refined the outlined results. The final segmentation results served as the benchmark for model training. Model architecture and training process Both the 2D U - Net and 3D U - Net models were trained using the self - configured nnU - Net framework. The training methods were standardized among the models, with the main differences being the input patch size, batch size, and model architecture.

[0061] 2D U - Net: The input images were resampled and cropped to 512×512 pixels, and then z - score intensity normalization was performed. The model architecture consisted of 7 downsampling and 7 upsampling stages, with a basic number of features of 32. The batch size used during training was 12.

[0062] 3D U-Net: The input images were resampled and cropped to 128×128×128 with a batch size of 2. The architecture had 5 downsampling and 5 upsampling stages with a base number of features of 32. Both models were trained using 5-fold cross-validation for 1000 epochs with an initial learning rate of 0.01. Test-time augmentation techniques were applied during inference to enhance prediction robustness ( Figure 6 ).

[0063] Model comparison based on tumor characteristics and prediction time used DSC, average surface distance (ASD), and 95th percentile Hausdorff distance (95%HD) to evaluate the segmentation performance of 2D and 3D UNet models. Two experienced board-certified radiologists reviewed the clinical, imaging, and pathology records. In addition to visually identifiable categorical parameters, we also extracted radiomics features from manually segmented tumors as the reference standard. Clinical data included gender and age. Imaging parameters included tumor location (high-risk: trigone and ureteral orifice), number (single vs. multiple), and the maximum tumor diameter for single tumors. Tumors were classified as small tumors (≤3 cm) and other tumors to evaluate the effect of size on segmentation performance. Pathology data included the presence of muscle invasion (MIBC or NMIBC).

[0064] Radiomics features: Thirteen morphological radiomics features were extracted using Python for additional model comparison. The inference time of both models was measured. Our aim was to evaluate which model achieved superior segmentation performance in various clinical scenarios to determine the method most suitable for different diagnostic and treatment settings. Statistical analysis was performed using SPSS software (version 26.0, IBM Corp., Armonk, NY, USA) and R software (version 4.4.2, R Foundation for Statistical Computing, Vienna, Austria). Quantitative variables were expressed as mean ± standard deviation (SD), while categorical variables were expressed as frequency and percentage to describe patient demographics and variable distribution.

[0065] (1) For within-group comparison of 2D U-Net and 3D U-Net, the non-parametric paired test, Wilcoxon signed-rank test, was used to compare the segmentation performance (DSC, ASD, and 95%HD) of each model (2D U-Net and 3D U-Net) under different tumor characteristics. The comparison evaluated whether there were significant differences in the segmentation accuracy measured by DSC between subgroups within the same model (e.g., single tumors vs. multiple tumors or different tumor sizes). A P value < 0.05 was considered statistically significant.

[0066] (2) Comparison between the 2DU-Net and 3DU-Net models within each tumor subgroup also used the Wilcoxon signed-rank test to compare the segmentation performance of the 2DU-Net and 3DU-Net models for each tumor subgroup. The evaluation metrics included DSC, ASD, and 95% HD to assess the segmentation accuracy and boundary precision. A P-value < 0.05 was considered statistically significant, indicating a significant difference in the segmentation performance of the two models for specific tumor features.

[0067] (3) Linear regression analysis was performed to evaluate the relationship between the DSC of the 2D U-Net and 3DU-Net models and the morphological radiomics features for the association between the segmentation performance and radiomics features. The regression lines for each model were plotted to visually show the trends. The P-value and confidence interval (CI) were reported to assess the statistical significance, and P < 0.05 was considered statistically significant.

[0068] Results:

[0069] Patient characteristics and overall segmentation performance A total of 435 patients were included in this study, including 325 in the training cohort, 72 in the internal test cohort, and 38 in the external test cohort. The average age of the patients in the training cohort was 64 ± 12 years, in the internal test cohort was 66 ± 11 years, and in the external test cohort was 65 ± 9 years (Table 1). The segmentation performance of the 2D U-Net and 3D U-Net for the bladder wall and tumors was evaluated. For bladder wall segmentation, the 3D U-Net performed slightly better in the internal test set, while the 2D U-Net performed better than the 3D U-Net in the external test set. However, for tumor segmentation, the 2D U-Net obtained higher Dice scores in both the internal (0.877 ± 0.073 vs. 0.828 ± 0.146) and external (0.87 ± 0.149 vs. 0.81 ± 0.157) test sets. In addition, the 2D U-Net had lower error values for ASD and 95% HD, especially in tumor segmentation, showing better boundary delineation and robustness. (Table 1, Figure 7 )

[0070] Table 1. Patient demographics and segmentation performance of the 2D U-Net and 3D U-Net models

[0071]

[0072] Table note: DSC, Dice similarity coefficient; ASD, average surface distance; 95% HD, 95th percentile Hausdorff distance

[0073] Within-model comparison of segmentation performance based on tumor subtypes:

[0074] (1) Tumor location: The segmentation performance of high-risk tumor sites (trigone and ureteral orifice) was compared with that of other sites. No statistically significant differences were observed in either model (P = 0.888 for 2D U-Net and P = 0.12 for 3D U-Net). This indicates that tumor location has little effect on the segmentation accuracy of both models (Table 2-4).

[0075] (2) Tumor multiplicity: For 2D U-Net, the Dice score for single tumors was 0.885 ± 0.080, higher than that for multiple tumors (0.844 ± 0.157), although the difference was not statistically significant. Similarly, for 3D U-Net, the Dice score for single tumors (0.833 ± 0.152) was higher than that for multiple tumors (0.796 ± 0.139), but the difference was barely significant. That is, when using the same segmentation technique for both single-tumor and multi-tumor segmentation, the performance of the single-tumor segmentation task is better (Table 3).

[0076] (3) Muscle invasion status: The segmentation accuracy between MIBC and NMIBC was analyzed. Both models performed slightly better on NMIBC tumors, but no significant differences were found, indicating that tumor invasiveness has little impact on the segmentation performance.

[0077] (4) Tumor size: For tumor size, there were significant differences in the segmentation performance between the two models. For small tumors (< 3 cm), the segmentation performance of both models improved. In 2D U-Net, the Dice score for tumors > 3 cm (0.915 ± 0.041) was significantly higher than that for other tumors (0.861 ± 0.116, P = 0.034). Similarly, in 3D U-Net, the Dice score for larger tumors (0.883 ± 0.061) was significantly higher than that for other tumors (0.804 ± 0.165, P = 0.023).

[0078] Table 2. Comparison of within-group DSC (Dice similarity coefficient) of 2D U-Net and 3D U-Net respectively.

[0079]

[0080] Table note: DSC, Dice similarity coefficient; MIBC, muscle-invasive bladder cancer; NMIBC, non-muscle-invasive bladder cancer

[0081] Table 3. Comparison of within-group average surface distance of 2D U-Net and 3D U-Net respectively

[0082]

[0083] Table 4. Comparison of the within-group 95% Hausdorff distance of 2D U-Net and 3D U-Net respectively

[0084]

[0085] These findings indicate that both models achieved better segmentation performance for larger tumors, while smaller tumors presented greater challenges, especially for 3D U-Net. Therefore, in smaller tumors, using 2D UNet for segmentation will yield relatively better results.

[0086] Comparison of segmentation performance models among different tumor subtypes:

[0087] For a single tumor, 2D U-Net performed significantly better than 3D U-Net in terms of Dice (P < 0.001), ASD (P = 0.002), and 95% HD (P = 0.034).

[0088] Similarly, for NMIBC tumors, 2D U-Net showed significantly higher Dice scores (0.879 ± 0.090 vs. 0.817 ± 0.162, P < 0.001), and significantly lower ASD (P = 0.006).

[0089] For tumor size, for small tumors (< 3 cm), 2D U-Net was also superior to 3D U-Net in terms of DSC (P = 0.028), while ASD (P = 0.008) and 95% HD (P = 0.029) were significantly lower, confirming higher segmentation accuracy for larger tumors.

[0090] For multiple tumors, there was a significant difference in Dice scores between 2D U-Net (0.796 ± 0.139) and 3D U-Net (0.844 ± 0.157) (P = 0.049), and the ASD of 3D U-Net was lower (although not significantly), but this indicates that 3D U-Net has potential advantages in dealing with multiple lesions.

[0091] For high-risk tumor locations and MIBC, there was no significant difference between the two models, indicating comparable performance in these subgroups (Table 5). Comparison based on radiomorphological features Eight morphological features showed significance ( Figure 8 ) Figure 8The relationship between tumor shape-based radiomics features and segmentation performance (DSC) of 2DU-Net and 3DU-Net was illustrated. Both models showed better segmentation accuracy for tumors with higher elongation, flatness, minimum axis length, and sphericity, and 2DU-Net showed stronger correlations in most cases. The surface area-to-volume ratio was negatively correlated with the Dice score (P < 0.001 for both models), and 3DUNet was more affected ( Figure 8 D).

[0092] Table 5. Inter-model comparison of three metrics between 2D U-Net and 3D U-Net in different tumor subgroups

[0093]

[0094] 2DU-Net showed significant positive correlations with the maximum 2D diameter, grid volume, and voxel volume, while 3DU-Net had no significant associations. 3DUNet relied more on sphericity (linear regression: P < 0.001), performed best on spherical tumors, but had difficulties in dealing with irregular morphologies. These results indicate that tumor shape significantly affects segmentation accuracy, 2DU-Net shows stronger adaptability to different morphologies, while 3D U-Net performs better on more regular tumor shapes.

[0095] Prediction time: On the internal test set, the prediction time of the 2DU-Net model was 630.818 ± 181.072 seconds, while that of the 3DU-Net model was 595.505 ± 179.131 seconds. There was a significant difference between the two models (p < 0.05). Both models achieved high segmentation accuracy, but 2D U-Net was still superior to 3D U-Net. This indicates that 3D U-net may be applicable to scenarios with higher requirements for segmentation speed, such as intraoperative imaging or real-time treatment planning for advanced MIBC cases.

[0096] Determine the first time threshold based on the performance on the test set if 3D U-Net segmentation is used in intraoperative imaging or real-time treatment;

[0097] In some embodiments, the first time threshold is 10 minutes.

[0098] For image segmentation of NMIBC, 2D U-Net is the preferred choice.

[0099] For a single tumor, 2DU-Net was significantly superior to 3DU-Net in terms of DSC, ASD, and 95%HD, consolidating its advantage in accurately depicting tumor boundaries.

[0100] Conversely, 3DU-Net shows slightly better performance and faster inference time in multi-tumor segmentation, which may make it the preferred choice for intraoperative imaging or real-time treatment planning in advanced MIBC cases, as speed and handling complex geometries are prioritized in these cases.

[0101] In summary, the excellent performance of 2D U-Net in small and single-tumor segmentation suggests its potential as a more reliable tool for early BCa assessment, while the efficiency of 3D U-Net in handling multiple lesions may support its use in cases with high tumor burden. Radiomics feature analysis provides additional insights into the segmentation performance of both models. 2D U-Net shows stronger correlations with shape-based features such as elongation, flatness, and minimum axis length, indicating its effectiveness in segmenting elongated or irregularly shaped tumors. In contrast, 3D U-Net shows a higher dependence on sphericity, which may explain its decreased performance in non-spherical tumors. The negative correlations between Dice of both models and surface area-to-volume ratio indicate that highly irregular tumors pose segmentation challenges, but 3DU-Net is more affected, possibly because it relies on volume continuity rather than fine spatial details. These results suggest that 2D U-Net is more adaptable to diverse morphologies, while 3D U-Net is most suitable for spherical tumors. In conclusion, our findings suggest that different segmentation models may be more preferable depending on tumor characteristics and clinical needs. The exceptional accuracy of 2DU-Net in small tumors (<1 cm), NMIBC, and irregularly shaped lesions indicates its potential for early BCa monitoring and precise tumor delineation. In contrast, the comparable performance and faster inference time of 3DU-Net in multiple tumors suggest that it may be beneficial for cases with high tumor burden or real-time applications such as intraoperative guidance. Additionally, tumor morphology affects segmentation accuracy, with 2D U-Net performing well in irregular tumors and 3DU-Net being more suitable for round lesions. Given their complementary advantages, a hybrid 2D-3D approach can optimize segmentation performance by leveraging the detailed feature extraction of 2D models and the volume continuity of 3D models. Future studies should explore hybrid architectures and integrate multimodal imaging (MRI, PET-CT) and deep learning-based classification models to further enhance clinical decision-making in BCa.

[0102] In summary, this study comprehensively compares the applications of 2DU-Net and 3DU-Net in BCa segmentation, highlighting the exceptional accuracy of 2DU-Net in small, single, and irregular tumors, while highlighting the potential advantages of 3DU-Net in segmenting multiple tumors and its significantly shorter inference time. These findings provide valuable insights into the optimal application of DL-based segmentation models in different clinical scenarios and emphasize the need to tailor model selection according to specific tumor characteristics.

[0103] In some embodiments, the process of the image segmentation method for bladder tumors is as Figure 10 shown.

[0104] In some embodiments, the proposed method of the present invention is used to segment the images in the portal phase.

[0105] In some embodiments, the images in the portal phase are segmented using the Figure 11 method shown.

[0106] Figure 3 FIG. is a schematic diagram of a computer device provided by an embodiment of the present invention, as Figure 3 shown. The device 2000 may include: one or more processors 2010, and one or more memories 2020; wherein, computer-readable code is stored in the memory, and when the computer-readable code is run by the one or more processors, the above-mentioned method can be executed.

[0107] The processor in this embodiment may be an integrated circuit chip with signal processing capabilities. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The various methods, operations and logic block diagrams disclosed in the embodiments of the present disclosure can be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc., and may be of the X86 architecture or the ARM architecture.

[0108] Generally speaking, the various example embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that can be executed by a controller, a microprocessor or other computing devices. When the aspects of the embodiments of the present disclosure are illustrated or described as block diagrams, flowcharts or using some other graphical representation, it will be understood that the blocks, devices, systems, technologies or methods described herein may be implemented as non-limiting examples in hardware, software, firmware, dedicated circuits or logic, general hardware or controllers or other computing devices, or some combination thereof.

[0109] For example, the method or device according to the embodiments of the present disclosure may also be implemented by means of the Figure 4 architecture of the computing device 3000 shown. As Figure 4As shown, the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input / output component 3060, a hard disk 3070, etc. Storage devices in the computing device 3000, such as the ROM 3030 or the hard disk 3070, may store various data or files used for the processing and / or communication of the methods provided by the present disclosure, as well as program instructions executed by the CPU. The computing device 3000 may also include a user interface 3080. Of course, Figure 4 the architecture shown is only exemplary, and when implementing different devices, one or more components in the computing device shown may be omitted according to actual needs. Figure 4

[0110] An embodiment of the present invention also provides a computer-readable storage medium, such as Figure 5 As shown, it is a schematic diagram of the storage medium 4000 provided by an embodiment of the present invention. Computer-readable instructions 4010 are stored on the computer storage medium 4020. When the computer-readable instructions 4010 are run by a processor, the methods according to the embodiments of the present disclosure described with reference to the above figures can be executed. The computer-readable storage medium in the embodiments of the present disclosure may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM). It should be noted that the memories for the methods described herein are intended to include, but are not limited to, these and any other suitable types of memories. It should be noted that the memories for the methods described herein are intended to include, but are not limited to, these and any other suitable types of memories.

[0111] An embodiment of the present disclosure also provides a computer program product or a computer program, which when executed by a processor implements the steps of the above method, such as Figure 2 shown, the computer program product or the computer program includes:

[0112] Acquisition module 201: used to acquire images of bladder tumor patients;

[0113] Initial segmentation module 202: used to input the image into 2D U-Net for tumor segmentation to obtain 2D segmentation results;

[0114] Traversal adjustment output module 203: traverse the shape of the tumor based on the 2D segmentation results. During traversal, if the current tumor shape is regular, input the image into 3D U-Net for tumor segmentation to obtain 3D segmentation results, and take the segmentation results of the current tumor in the 3D segmentation results; if the tumor shape is irregular, take the segmentation results of the current tumor in the 2D segmentation results. After traversal, merge and output the segmentation results of multiple tumors.

[0115] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0116] Generally speaking, various example embodiments of the present disclosure can be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Some aspects can be implemented in hardware, while other aspects can be implemented in firmware or software that can be executed by a controller, a microprocessor, or other computing devices. When aspects of the embodiments of the present disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, devices, systems, technologies, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuits or logic, general hardware or controllers or other computing devices, or some combination thereof.

[0117] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0118] In several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in electrical, mechanical, or other forms.

[0119] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0120] In addition, in each embodiment of the present invention, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0121] The exemplary embodiments of the present disclosure described in detail above are merely illustrative and not restrictive. Those skilled in the art should understand that various modifications and combinations can be made to these embodiments or their features without departing from the principles and spirit of the present disclosure, and such modifications should fall within the scope of the present disclosure.

Claims

1. An image segmentation method for bladder tumors, characterized in that, The method includes: S101: Obtain the image of a bladder tumor patient; S102: Input the image into a 2D U-Net for tumor segmentation to obtain a 2D segmentation result; S103: Traverse the shape of the tumor based on the 2D segmentation result. When traversing, if the current tumor shape is regular, input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result, and take the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, take the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

2. The image segmentation method for bladder tumors according to claim 1, wherein, The method further includes, based on the 2D segmentation result, determining whether the number of tumors is greater than 1. If it is greater than 1, input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result; Traverse the shape of the tumor based on the 3D segmentation result. When traversing, if the current tumor shape is regular, take the segmentation result of the current tumor in the 3D segmentation result; If the tumor shape is irregular, take the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

3. An image segmentation method for bladder tumors, characterized in that, The method includes: S301: Obtain the image of a bladder tumor patient; S302: Input the image into a 3D U-Net for tumor segmentation to obtain a 3D segmentation result; S303: Traverse the shape of the tumor based on the 3D segmentation result. When traversing, if the current tumor shape is regular, take the segmentation result of the current tumor in the 3D segmentation result; if the tumor shape is irregular, input the image into a 2D U-Net for tumor segmentation to obtain a 2D segmentation result, and take the segmentation result of the current tumor in the 2D segmentation result. After the traversal is completed, merge and output the segmentation results of multiple tumors.

4. The method for image segmentation of bladder tumors according to claim 2 or 3, characterized in that, The method further includes: Simultaneously obtain the time limit of the segmentation task. If the time limit is lower than the first time threshold, input the image into a 3D U-Net for segmentation, and output the segmentation result after obtaining the segmentation result. Otherwise, input the image into a 2D U-Net or 3D U-Net for tumor segmentation and then traverse the shape of the tumor.

5. The image segmentation method for bladder tumors according to claim 1 or 3, characterized in that, The method further includes: Judging the sphericity of the tumor shape. If the sphericity of the tumor shape meets a threshold higher than the sphericity threshold, it is judged as regular in shape, otherwise it is judged as irregular in shape; Optionally, the sphericity threshold is constructed based on the sphericity of the training set; Optionally, a shape discrimination classifier is used to judge whether the tumor shape is regular. The construction method of the shape discrimination classifier is to input the tumor images of the training set and the annotation of whether it is regular into the classifier for training.

6. The image segmentation method of bladder tumors according to claim 1 or 3, characterized in that, The method further includes: Input the image into a 2D U-Net for bladder wall segmentation to obtain a bladder wall segmentation result, and perform tumor segmentation based on the image and the bladder wall segmentation result.

7. The method for image segmentation of bladder tumors according to claim 1 or 3, characterized in that The method further includes: The model architecture of the 2D U-Net consists of 7 downsampling and 7 upsampling stages, and the basic number of features is 32; Optionally, the model architecture of the 3D U-Net consists of 5 downsampling and 5 upsampling stages, and the basic number of features is 32; Optionally, the test-time augmentation technique is adopted in the training process of the 2D U-Net and 3D U-Net; Optionally, the image is an enhanced CT image.

8. A computer device, characterized in that, The device includes: a memory and a processor; the memory is used for storing a computer program; the processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1-7 are implemented.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, the steps of the method according to any one of claims 1-7 are implemented.

Images