Automatic diagnosis method and automatic diagnosis system for excretive bladder urinary tract radiography through deep learning

Abstract

The invention relates to the technical field of medical detection, in particular to an automatic diagnosis method and an automatic diagnosis system for excretive bladder urinary tract radiography through deep learning. According to the method, the bladder, the ureter and the urethra area are automatically positioned and diagnosed based on the VCUG image, VUR automatic grading can be achieved, whether the bladder and the urethra are abnormal or not can be diagnosed, and therefore the clinical application range is wider; besides, automatic diagnosis is carried out by utilizing a deep learning construction model, so that the influence of human factors on results can be effectively reduced, the accuracy is improved, meanwhile, the judgment of doctors, especially primary doctors, can be remarkably assisted, so that the doctors can make correct diagnosis in time, and the problems of low result repeatability among different years and different hospitals and the like can be solved.

Description

Technical Field

[0001] The present invention relates to the field of medical detection technology, and in particular to an automatic diagnosis method and automatic diagnosis system for excretory cystourethrography using deep learning. Background Art

[0002] Voiding cystourethrogram (VCUG) is one of the most common pediatric examinations and is the gold standard for depicting the anatomy of the lower urinary tract and determining its function. Therefore, VCUG is suitable for the investigation of patients with various clinical conditions, including urinary tract infection, antenatal hydroureteral nephropathy, posterior urethral valves, and cloacal abnormalities. During filling, the bladder is evaluated for shape, size, filling defects, and vesicoureteral reflux (VUR). During the micturition phase, in addition to evaluating reflux, the bladder and urethra can also be evaluated for any anatomical abnormalities or voiding dysfunction. VCUG is considered the gold standard for various urological diagnoses, including VUR. Currently, although multiple guidelines for the operation of VCUG have been standardized, there are currently reports of VCUG with mixed results.

[0003] In recent years, deep neural network models have been increasingly used in the field of medical imaging. These models can not only process a large number of medical images efficiently, quickly and automatically, reducing the workload of doctors, but also make image-based diagnosis, evaluation and image-guided intervention or treatment more objective, accurate and intelligent. In the literature “Eroglu Y, Yildirim K, Çinar A, Yildirim M. Diagnosis and grading of vesicoureteralreflux on voidingcystourethrography images in children using a deep hybrid model. Computermethodsandprogramsinbiomedicine2021;210:106369.”, Yesim et al. and in the literature “LiZ, Tan Z,Wang Z, et al. Development and multi-institutional validation of adeep learning model for grading of vesicoureteral reflux onvoidingcystourethrogram: a retrospective multicenter study. EClinicalMedicine2024;69:102466.”, Li Zhanchi et al. proposed a deep learning method based on VCUG images for automatic grading of VUR. However, these current methods can only grade VUR on a single VCUG image. Since VCUG examination is a dynamic process, usually involving multiple images, and in addition to grading VUR, attention should be paid to other abnormalities of the bladder and urethra. Therefore, the coverage of existing deep learning methods in practical clinical applications is still limited.

[0004] The applicant's invention patent application in China (publication number: CN115762753A, publication date: 2023-03-07) discloses a method for automatic classification of vesicoureteral reflux by deep learning, which includes the following steps: collecting VCUG image sets of VUR patients; manually marking the reflux area to complete the collection of the data set; using the integrated deep learning method of the voting method to select the optimal result as the overall prediction label; training the model with the preprocessed training data set to form the final VCUG automatic classification model; inputting new pictures for prediction, first performing the same image data preprocessing to obtain the vesicoureteral reflux classification standard map; according to the VCUG automatic classification model, automatically classifying the obtained vesicoureteral reflux classification standard map. This method also does not involve a method for processing multiple images. Summary of the invention

[0005] In order to address the deficiencies of the prior art and expand the scope of application of VCUG intelligent grading in clinical practice, the present invention provides a deep learning automatic diagnosis method for voiding cystourethrography, which automatically locates and diagnoses the bladder, ureter and urethra areas based on VCUG images, and can efficiently and accurately realize the automated diagnosis of voiding cystourethrography.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions: A deep learning automatic diagnosis method for voiding cystourethrography, the method comprising the following steps: Step 1: Data collection and image screening: Images of children undergoing voiding cystourethrography were collected, and images with poor image quality were excluded; Step 2: Image annotation and dataset construction: The bladder, urethra, and areas with vesicoureteral reflux were segmented and classified from the selected VCUG photos. If there was any disagreement between the two, the senior doctor would make the final classification to complete the collection of the data set. Step 3: Deep learning model construction: The automatic diagnosis model of VCUG is constructed using deep learning methods. The model mainly consists of two main parts: 1) Segmentation and classification based on a single VCUG image: Use a unified transformer-based architecture to handle classification and segmentation tasks simultaneously; input a single VCUG image and output a lesion mask and diagnosis; 2) Build a VCUG automatic segmentation and diagnosis model based on multiple VCUG images of a single person: extract high-dimensional feature representations of each VCUG image based on the output of the first part of the model, and superimpose these disease-related representations into a matrix; then design a multi-task attention layer to dynamically learn the contribution of each high-dimensional representation, and finally predict whether the patient has bladder or urethral abnormalities and determine the grade of ureteral reflux; Step 4: Testing of deep learning model: Multiple VCUG images of a new single patient are input, and image preprocessing is first performed. Then, the standardized VCUG images are input into the constructed automatic diagnosis model; the model automatically completes the segmentation of the bladder, urethra and ureteral areas, and diagnoses lesions in the segmented areas.

[0007] Preferably, the image data collection includes: collecting a sufficient number of images of children undergoing VCUG examinations; the labels for training the convolutional neural network model include: a senior doctor segments the bladder, urethra and ureteral areas in a single VCUG image and makes a diagnosis, and the diagnosis results include whether the bladder and urethra are abnormal, and the level of vesicoureteral reflux; for multiple VCUG images of a single person, a final diagnosis is made after a comprehensive judgment based on the results of each VCUG image, and the diagnosis results include whether the bladder and urethra are abnormal, and the level of vesicoureteral reflux.

[0008] Preferably, the process of processing the input VCUG images when constructing the automatic diagnosis model based on excretory cystourethrography includes: using the network model of automatic diagnosis of excretory cystourethrography, firstly segmenting the bladder, urethra and ureter for each image of the obtained VCUG, and then classifying the segmented areas; then, for multiple VCUG images of a single person, the model integrates the segmentation and diagnosis results of each VCUG photo to calculate the contribution, and finally makes a final diagnosis, the diagnosis results including whether the bladder and urethra are abnormal, and the grade of vesicoureteral reflux.

[0009] Preferably, the preprocessing of each VCUG image data input includes: cropping to remove edge annotations, filling the edges of the image with black to make the image shape a square, and then uniformly scaling the image to a size of 224×224 pixels, and finally normalizing the pixel value of the image to the range of 0~1, and standardizing the image according to the channel, that is, affine transformation, perspective transformation, irregular deformation, and adding random noise; preferably, each image is randomly rotated 0~15°.

[0010] Preferably, the automatic diagnosis method model of the excretory cystourethrography comprises: 1) Collect VCUG images with segmentation and diagnosis results as a dataset of training samples; 2) Use data augmentation technology to collect training sample data sets, transform images to increase the number and variation of images, thereby improving the generalization ability of the network model and the robustness of image segmentation capabilities; 3) Feature extraction network of deep learning; 4) Model for automated diagnosis of voiding cystourethrography.

[0011] Preferably, the deep learning models involved in excretory cystourethrography are all based on convolutional neural networks for feature extraction, with the processed VCUG image as input, and feature extraction through convolutional layers, pooling layers, and activation function layers, ultimately obtaining a feature map with complex semantic information; the automatic diagnosis model for excretory cystourethrography includes: expanding the feature map of the VCUG image to obtain the feature vector corresponding to the image, performing feature fusion through a fully connected layer, and then calculating the probability of the image being classified into each level through a softmax function, and finally comparing the size relationship between the probabilities to make a level judgment.

[0012] Furthermore, the present invention also discloses the application of the method in designing automatic diagnosis software for voiding cystourethrography.

[0013] Furthermore, the present invention also discloses an automatic diagnosis system using the method, the system comprising: An image acquisition module is used to acquire images of children undergoing voiding cystourethrography and screen the images to exclude those with poor image quality; An image annotation module is used to segment and classify the bladder, urethra and ureter regions of the selected VCUG images and generate an annotation data set; A deep learning model module, which is used to build an automatic diagnosis model based on deep learning. The model includes a segmentation and classification module based on a single VCUG image, and a comprehensive diagnosis module based on multiple VCUG images of a single person. The model dynamically learns the contribution of each image through a multi-task attention layer and outputs a diagnosis result, including whether the bladder and urethra are abnormal and the grade of vesicoureteral reflux; The data preprocessing module is used to preprocess the input VCUG image, including cropping, padding, scaling and normalization operations; The automatic diagnosis module is used to receive multiple VCUG images of patients, complete image segmentation and diagnosis, and generate segmentation and lesion diagnosis results of the bladder, urethra and ureter areas by analyzing the high-dimensional features of multiple images.

[0014] Preferably, the deep learning model module is initialized by the ViT-base model pre-trained by the ImageNet dataset, and the masked self-supervised learning technology is used to enhance the system's feature extraction capability for medical images and accelerate the model training process.

[0015] Preferably, the system also includes a feature fusion module for fusing the feature vectors of each image through a fully connected layer, and calculating the probability of the image being classified into each level through a softmax function, and finally determining the diagnostic level based on the probability.

[0016] The present invention adopts the above-mentioned technical solution, which can not only realize automatic VUR classification, but also diagnose whether the bladder and urethra are abnormal, thus expanding the scope of clinical application; in addition, the use of deep learning to build a model for automatic diagnosis can effectively reduce the impact of human factors on the results and improve accuracy. At the same time, it can significantly assist doctors in their judgment, especially junior doctors, so that they can make correct diagnoses in a timely manner, which is conducive to solving the problem of low repeatability of results between different years of service and different hospitals. It is specifically reflected in the following technical effects: 1. Automation and efficient diagnosis The present invention's deep learning-based automatic diagnosis system for voiding cystourethrography (VCUG) achieves automatic segmentation and classification of multiple VCUG images by using deep learning models and Transformer architecture. This enables the system to quickly process large amounts of image data, effectively reducing the workload of doctors and improving diagnostic efficiency. Compared with traditional manual labeling and judgment, the system can process multiple images in a short period of time and automatically generate diagnostic results, greatly improving the efficiency of clinical diagnosis.

[0017] 2. Improved diagnostic accuracy and consistency The deep learning model of the present invention can comprehensively analyze the high-dimensional features in multiple VCUG images through pre-training and multi-task attention mechanisms, and dynamically learn the contribution of each image in the diagnosis. This not only improves the diagnostic accuracy of bladder, urethral abnormalities and vesicoureteral reflux (VUR) grades, but also effectively reduces diagnostic inconsistencies caused by human factors, especially in the operation of junior doctors or less experienced doctors, and can ensure the reliability of diagnostic results.

[0018] 3. Comprehensive multi-task diagnosis The present invention designs a multi-task model that can simultaneously segment and diagnose the bladder, urethra, and ureter, and combines the correlation information between multiple tasks to further improve the comprehensiveness of the diagnosis. This ability to perform multi-task segmentation and classification simultaneously enables the system to not only detect a single disease, but also evaluate the abnormal conditions of multiple related organs, providing more comprehensive diagnostic information and broadening the application scope of the system.

[0019] 4. Improved generalization and robustness of the model By using data enhancement techniques (such as affine transformation, perspective transformation, and adding random noise) in data processing, the present invention effectively enhances the generalization ability and robustness of the model. When processing actual clinical data, the system can better adapt to differences in image quality, angle, and deformation, ensuring that stable and high-quality diagnostic results can still be provided in different scenarios.

[0020] 5. Significant clinical application value By applying deep learning to the automatic diagnosis of voiding cystourethrography, the present invention significantly reduces the diagnostic differences between doctors of different years of experience and effectively solves the problem of poor repeatability caused by the strong subjectivity of manual judgment. At the same time, the system can significantly assist junior doctors in making correct diagnoses, improve the level of medical services, and enable more hospitals and doctors to benefit from it, especially in areas where resources are relatively scarce.

[0021] 6. Reduce human intervention and reduce misdiagnosis rate The present invention automatically completes the segmentation and diagnosis of the bladder, urethra and ureter through a deep learning model, effectively reducing the risk of misdiagnosis caused by human intervention. At the same time, through the standardized diagnosis process of the system, it can improve the consistency of diagnosis, reduce the probability of misdiagnosis and missed diagnosis, and provide patients with more reliable diagnostic services. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The present invention is an embodiment of a reflux diagnosis model for a patient's bladder, urethra and ureter based on VCUG images.

[0023] Figure 2 A schematic diagram of a model for automated diagnosis based on VCUG images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] A method for automatic diagnosis of voiding cystourethrography comprises the following steps: Step 1: Collect image sets of children undergoing VCUG examination; Step 2: For all the VCUG photos obtained, two senior pediatric urologists segmented and diagnosed the bladder, urethra, and ureteral regions of each VCUG image, and made a final diagnosis based on multiple VCUG images of a single person. If there was a disagreement between the two, the senior doctor made the final classification to complete the collection of the data set; Step 3: Use the Transformer network to build a voiding cystourethrography model. The model architecture mainly consists of two parts: segmentation and classification based on a single image, and diagnosis of VCUG results for a single person based on the results of a single image. The model based on a single VCUG image uses a unified transformer-based architecture to handle classification and segmentation tasks simultaneously. It inputs an image, outputs segmented regions (bladder, urethra, ureter) and performs a diagnosis: including bladder abnormalities, urethral abnormalities, and grading of vesicoureteral reflux. The model is initialized using ViT-base weights pre-trained on ImageNet with masked self-supervision, which stabilizes training, accelerates convergence, and enhances the representation ability of image-level models.

[0026] Diagnosis based on multiple VCUG images of a single person: Use the weights of the image-level model from the single image level as its initial weights. We freeze these weights to extract high-dimensional feature representations for each image and stack these disease-related representations into a matrix. Then design a multi-task attention layer to dynamically learn the contribution of each high-dimensional representation, and finally predict whether the patient has bladder or urethral abnormalities and determine the grade of ureteral reflux.

[0027] Step 4: Input new single-person multiple VCUG images for prediction, first perform the same image data preprocessing to obtain a standard image; according to the constructed automatic diagnosis model based on voiding cystourethrography, first perform segmentation and diagnosis of the bladder, urethra and ureteral regions of a single VCUG image based on a single photo model; then integrate the results of each VCUG and perform a final diagnosis of multiple VCUG images of a single person; Image data collection: a certain number of patient images undergoing VCUG examinations are collected, and senior doctors segment and diagnose the bladder, ureter and urethra areas. The automatic diagnosis model of voiding cystourethrography constructed by deep learning model is named VCUG-DAM. Using the collected samples as training set, the model is responsible for completing the segmentation and diagnosis of the lesion area (bladder, urethra and ureter) of a single VCUG image; and making the final diagnosis based on the integration of each result of multiple VCUGs of a single person.

[0028] Furthermore, the image data collection of the present invention includes: Collect VCUG photos of a sufficient number of patients and exclude photos with poor image quality (e.g., blurry, with large annotation overlap, and without kidney and ureteral areas). Senior doctors segment and diagnose the bladder, ureter, and urethra areas on the collected VCUG photos; The image datasets collected above will be determined as training samples for the subsequent construction of an automatic diagnosis model based on voiding cystourethrography.

[0029] Furthermore, the process of processing the input single-person multiple VCUG images when building the model includes: Using an automatic diagnosis model based on voiding cystourethrography: First, the image-based model uses a unified transformer-based architecture to handle classification and segmentation tasks simultaneously. It inputs an image, outputs segmented regions (bladder, urethra, ureter) and performs a diagnosis: including bladder abnormalities, urethral abnormalities, and vesicoureteral reflux classification. The model is initialized using masked self-supervision ViT-base weights pre-trained on ImageNet, which stabilizes training, accelerates convergence, and enhances the representation capabilities of image-level models.

[0030] The patient-based model then uses the weights from the image-level model as its initial weights. We freeze these weights to extract high-dimensional feature representations for each image and stack these disease-related representations into a matrix. A multi-task attention layer is then designed to dynamically learn the contribution of each high-dimensional representation, ultimately predicting whether the patient has bladder or urethral abnormalities and determining the grade of ureteral reflux.

[0031] Furthermore, the construction and processing of each module in the automatic segmentation and diagnosis task of multiple VCUG images of a single child is as follows: 1. Preprocessing of input VCUG images Preprocessing the experimental data is a key step to improve data quality. It can standardize the image and reduce the noise and interference in the image through some simple operations. In this study, we first cropped all the images to remove the edge annotations. Then the edges of the image were filled with black to make the image shape a square with equal length and width, and then the image was uniformly scaled to 224×224 pixels to provide image inputs of consistent size to the machine learning model. The reason for padding the image is that the width and length of the ureter in the VCUG image are important for the grade classification of the VCUG image. By first filling it into a square and then scaling it, the shape of the ureter in the image can be effectively avoided. Distortion. Finally, the pixel values ​​of the image are normalized to the range of 0~1, and the image is standardized according to the channel, that is, the mean is subtracted and then divided by the variance, which can speed up the convergence of the model.

[0032] Data augmentation is a common method to reduce overfitting and improve classification performance of small data sets. During the training process, by performing random data augmentation operations on each image, the machine learning model can learn more generalized features and improve the robustness of the model. In this study, we perform affine transformation, perspective transformation, and irregular deformation on each image; preferably, each image is randomly rotated 0~15°.

[0033] 2. Segmentation and diagnosis model based on a single VCUG image We propose a unified transformer-based architecture that performs multi-label classification and semantic segmentation simultaneously, aiming to improve overall performance through the complementarity of these tasks. To stabilize training and improve feature representation, we initialize the model using ViT-Base pre-trained on the ImageNet dataset using masked self-supervised learning. This pre-trained initialization accelerates model convergence and improves performance on both classification and segmentation tasks.

[0034] Our model first encodes the input multi-label image into classification tokens and patch tokens. After being processed by the Transformer encoder, the classification tokens are input into a fully connected layer to perform a multi-label classification task. This classification task is divided into four subtasks: bladder abnormality, urethral abnormality, left ureteral reflux grade, and right ureteral reflux grade. To predict the specific label for each subtask, we design a fully connected layer to map the classification head into a 16-dimensional feature vector, where different parts of the vector correspond to specific subtask categories: T[0:2] is used to predict bladder abnormality, T[2:4] predicts urethral abnormality, T[4:10] predicts left ureteral reflux grade, and T[10:16] predicts right ureteral reflux grade.

[0035] While training the classification task, we simultaneously designed a masked Transformer module to perform semantic segmentation, specifically targeting the bladder, ureters, and urethra. To further enhance the interactivity between the classification and segmentation tasks, we added a self-attention module to facilitate communication between the two. The semantic information from the classification task helps the segmentation module to better locate the region of interest, while the segmentation task in turn provides semantic features that aid classification. This design promotes information sharing between tasks and generates more expressive high-dimensional feature representations.

[0036] Furthermore, we constructed a VCUG automatic segmentation and diagnosis model based on multiple VCUG images of a single person as follows: After achieving segmentation and diagnosis based on a single VCUG image, we extended the model to instance-level classification, aiming to automatically identify the contribution of each image in the patient image collection and generate the final classification. The process is as follows: First, we load the pre-trained weights from the image-level encoder and freeze these parameters, denoted as T_"frozen". We set the maximum number of images for each patient to 10. For patients with less than 10 images, zero-value image padding is used to ensure that the input size remains consistent. These images are then input into the frozen classifier T_"frozen" in sequence to extract high-dimensional feature representations. The features of all images are superimposed into a matrix, denoted as H∈R^(10×D), where D represents the feature dimension.

[0037] To perform instance-level classification, we introduce a set of K learnable class embeddings, “cls” = [cls_1, …, cls_K] ∈ R^(K×D), where K = 3, corresponding to the tasks of bladder abnormality, urethral abnormality, and ureteral reflux grading. Each class embedding, together with the high-dimensional feature matrix H of the 10 images, is processed through a decoder. Finally, we use three independent classification heads, where each head maps a class embedding to a probability distribution over possible disease classes. This structure enables the model to automatically learn the contribution of each image to each disease classification task.

[0038] The training results of the model in a single center are shown in Tables 1 and 2.

[0039] Table 1. Diagnostic performance of the model on a single photo on the internal validation dataset Table 2 Model diagnostic performance for a single person on the internal validation dataset The above is a description of the embodiments of the present invention. Through the above description of the disclosed embodiments, professionals and technicians in the field can implement or use the present invention. Various modifications to these embodiments will be apparent to professionals and technicians in the field. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in this article, but will conform to the widest range consistent with the principles and novelties disclosed herein.

Claims

1. A deep learning automatic diagnosis method for voiding cystourethrography, characterized in that: The method comprises the following steps: Step 1: Data collection and image screening: Images of children undergoing voiding cystourethrography were collected, and images with poor image quality were excluded; Step 2: Image annotation and dataset construction: The bladder, urethra, and areas with vesicoureteral reflux were segmented and classified from the selected VCUG photos. If there was any disagreement between the two, the senior doctor would make the final classification to complete the collection of the data set. Step 3: Deep learning model construction: The automatic diagnosis model of VCUG is constructed using deep learning methods. The model mainly consists of two main parts: 1) Segmentation and classification based on a single VCUG image: Use a unified transformer-based architecture to handle classification and segmentation tasks simultaneously; input a single VCUG image and output a lesion mask and diagnosis; 2) Build a VCUG automatic segmentation and diagnosis model based on multiple VCUG images of a single person: extract high-dimensional feature representations of each VCUG image based on the output of the first part of the model, and superimpose these disease-related representations into a matrix; then design a multi-task attention layer to dynamically learn the contribution of each high-dimensional representation, and finally predict whether the patient has bladder or urethral abnormalities and determine the grade of ureteral reflux; Step 4: Testing of deep learning model: Multiple VCUG images of a new single patient are input, and image preprocessing is first performed. Then, the standardized VCUG images are input into the constructed automatic diagnosis model; the model automatically completes the segmentation of the bladder, urethra and ureteral areas, and diagnoses lesions in the segmented areas.

2. The automatic diagnosis method according to claim 1, characterized in that: The image data collection includes: collecting a sufficient number of images of children undergoing VCUG examinations; the labels for training the convolutional neural network model include: senior doctors segment the bladder, urethra and ureteral areas in a single VCUG image and make a diagnosis, and the diagnosis results include whether the bladder and urethra are abnormal, and the level of vesicoureteral reflux; for multiple VCUG images of a single person, a final diagnosis is made after a comprehensive judgment based on the results of each VCUG image, and the diagnosis results include whether the bladder and urethra are abnormal and the level of vesicoureteral reflux.

3. The automatic diagnosis method according to claim 1, characterized in that: The process of processing the input VCUG images when constructing the automatic diagnosis model based on excretory cystourethrography includes: using the network model of automatic diagnosis of excretory cystourethrography, firstly segmenting the bladder, urethra and ureter for each image of the obtained VCUG, and then classifying the segmented areas; then, for multiple VCUG images of a single person, the model integrates the segmentation and diagnosis results of each VCUG photo to calculate the contribution, and finally makes a final diagnosis, and the diagnosis results include whether the bladder and urethra are abnormal, and the reflux level of the vesicoureter.

4. The automatic diagnosis method according to claim 1, characterized in that: The preprocessing of each VCUG image data input includes: cropping to remove edge annotations, filling the edges of the image with black to make the image shape square, and then scaling the image to a uniform size of 224×224 pixels. Finally, the pixel value of the image is normalized to the range of 0~1, and the image is standardized according to the channel, that is, affine transformation, perspective transformation, irregular deformation, and adding random noise; preferably, each image is randomly rotated 0~15°.

5. The automatic diagnosis method according to claim 1, characterized in that: The automatic diagnosis method model of the excretory cystourethrography includes: 1) Collect VCUG images with segmentation and diagnosis results as a dataset of training samples; 2) Use data augmentation technology to collect training sample data sets, transform images to increase the number and variation of images, thereby improving the generalization ability of the network model and the robustness of image segmentation capabilities; 3) Feature extraction network of deep learning; 4) Model for automated diagnosis of voiding cystourethrography.

6. The automatic diagnosis method according to claim 5, characterized in that: The deep learning models involved in voiding cystourethrography are all based on convolutional neural networks for feature extraction. They take the processed VCUG images as input, and after feature extraction through convolutional layers, pooling layers, and activation function layers, they finally obtain feature maps with complex semantic information. The automatic diagnosis model for voiding cystourethrography includes: expanding the feature map of the VCUG image to obtain the feature vector corresponding to the image, fusing the features through the fully connected layer, and then calculating the probability of classifying the image into each level through the softmax function, and finally comparing the size relationship between the probabilities to make a level judgment.

7. Use of the method according to any one of claims 1 to 6 in designing automatic diagnostic software for voiding cystourethrography.

8. An automatic diagnosis system using the method according to any one of claims 1 to 6, characterized in that: The system comprises: An image acquisition module is used to acquire images of children undergoing voiding cystourethrography and screen the images to exclude those with poor image quality; An image annotation module is used to segment and classify the bladder, urethra and ureter regions of the selected VCUG images and generate an annotation data set; A deep learning model module, which is used to build an automatic diagnosis model based on deep learning. The model includes a segmentation and classification module based on a single VCUG image, and a comprehensive diagnosis module based on multiple VCUG images of a single person. The model dynamically learns the contribution of each image through a multi-task attention layer and outputs a diagnosis result, including whether the bladder and urethra are abnormal and the grade of vesicoureteral reflux; The data preprocessing module is used to preprocess the input VCUG image, including cropping, padding, scaling and normalization operations; The automatic diagnosis module is used to receive multiple VCUG images of patients, complete image segmentation and diagnosis, and generate segmentation and lesion diagnosis results of the bladder, urethra and ureter areas by analyzing the high-dimensional features of multiple images.

9. The automatic diagnosis system according to claim 8, characterized in that: The deep learning model module is initialized by the ViT-base model pre-trained by the ImageNet dataset, and uses mask self-supervised learning technology to improve the system's feature extraction capability for medical images and accelerate the model training process.

10. The automatic diagnosis system according to claim 8, characterized in that: The system also includes a feature fusion module, which is used to fuse the feature vectors of each image through a fully connected layer, and calculate the probability of the image being classified into each level through a softmax function, and finally make a diagnosis level judgment based on the probability.

Images