Deep learning prediction system for bladder ureter reflux of children

Abstract

A child bladder ureteral reflux deep learning prediction system relates to the field of image processing, and comprises an image acquisition unit used for acquiring a dimercaptosuccinic acid kidney static image of a child patient; the image processing unit is used for carrying out region-of-interest division on the image and carrying out preprocessing; the prediction model is trained to obtain the kidney image processed by the image processing unit and output a classification result, and the prediction model comprises an image block feature embedding module; the multi-layer feature extraction module captures local features through a layered window multi-head self-attention mechanism, and an image block merging layer is inserted between every two basic layers to perform down-sampling on a feature map; and the feature fusion layer is used for fusing the features captured by each layer and outputting a classification result. According to the invention, the local and global features in the medical image can be effectively captured, the effective utilization of the medical image data is realized, and the auxiliary diagnosis effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a deep learning prediction system for vesicoureteral reflux in children. Background Art

[0002] Vesicoureteral reflux (VUR) is one of the common urinary system diseases in childhood. In the diagnosis of VUR, dimercaptosuccinic acid (DMSA) renal static scan is a key auxiliary tool. As a functional imaging examination, DMSA renal static scan can effectively evaluate the infection or scarring of the kidney caused by VUR, helping clinicians understand the extent of kidney damage and its possible causes. For patients with potential VUR, the results of DMSA renal static scan can provide an important basis for whether to further perform voiding cystourethrography (VCUG) examination. When DMSA renal static scan shows significant damage or scarring of the kidney, this may indicate more severe VUR, prompting doctors to recommend voiding cystourethrography to confirm the presence and severity of urine reflux, providing more specific guidance for subsequent treatment decisions.

[0003] However, although the renal static scan of dimercaptosuccinate is of great value in the evaluation and monitoring of renal injury associated with vesicoureteral reflux, it is not the gold standard for vesicoureteral reflux. The renal static scan of dimercaptosuccinate cannot directly show the reflux process of urine, and its diagnostic value is mainly reflected in the indirect indication of possible vesicoureteral reflux. Therefore, its sensitivity and diagnostic value are limited for cases of early vesicoureteral reflux or those that do not cause obvious renal injury. In addition, the renal static scan of dimercaptosuccinate has a low specificity and may mistake renal abnormalities caused by other non-vesicoureteral reflux factors as damage caused by vesicoureteral reflux, which may lead to unnecessary further examinations or treatments.

[0004] Therefore, in the diagnosis and management of vesicoureteral reflux, succimer renal static scanning should be considered as an evaluation tool rather than a single diagnostic method. Combining the patient's medical history, clinical manifestations, and especially the results of voiding cystourethrography (VCUG), the gold standard for vesicoureteral reflux, can more comprehensively and accurately evaluate the presence of vesicoureteral reflux and its impact on the kidneys, thereby formulating the best treatment strategy.

[0005] For example, Chinese patent CN115762753A discloses a method for automatic classification of vesicoureteral reflux using deep learning. This method is based on voiding cystourethrography images and uses an integrated deep learning model to automatically classify the severity of vesicoureteral reflux. The model is trained using a data set annotated by senior doctors, and the prediction results of different deep learning models are integrated using a voting method to form a final classification label. Model preprocessing includes image cropping, standardization, and data enhancement operations. The final model is built based on a residual network (Residual Network, ResNet) architecture, which can handle the classification of unilateral and bilateral reflux at the same time.

[0006] Invasive examinations relying on voiding cystourethrography: This method is based on voiding cystourethrography images to grade reflux. Although it can provide highly accurate grading results, voiding cystourethrography is radioactive and invasive to patients, so it is not suitable for frequent use, especially for children. It is not safe enough. In addition, voiding cystourethrography is widely regarded as the gold standard for judging vesicoureteral reflux, and its diagnostic consistency is high. Even different doctors make consistent judgments on the same voiding cystourethrography image. Therefore, this method introduces deep learning technology to diagnose the grade of vesicoureteral reflux based on voiding cystourethrography images, but the room for improving the practicality of diagnosis is very limited, and therefore its auxiliary value for actual clinical work is insufficient. Summary of the invention

[0007] In view of the deficiencies in the prior art, the object of the present invention is to provide a deep learning prediction system for vesicoureteral reflux in children.

[0008] To achieve the above object, the present invention provides the following technical solutions: A deep learning prediction system for vesicoureteral reflux in children, comprising: An image acquisition unit, used for acquiring a static image of the succimer kidney of the child; An image processing unit divides the acquired static image of the kidney of dithiosuccinic acid into regions of interest, performs background correction, and performs quantitative processing to obtain a processed kidney image with risks; The prediction model can be trained to obtain the kidney image processed by the image processing unit and output the classification result. The prediction model includes: The image block feature embedding module divides the input image into several non-overlapping image blocks through convolution operations, and maps each image block into an embedding vector of fixed dimension through the convolution layer; Several layers of feature extraction modules capture local features through a layered window multi-head self-attention mechanism, and insert an image block merging layer between each basic layer to downsample the feature map; The feature fusion layer fuses the features captured by each layer and outputs the classification result.

[0009] The image processing unit preprocesses the dithiosuccinic acid kidney static image based on the following steps: S1, background correction of a static image of a succinimidyl kidney; S2, kidney contours were delineated on the background-corrected image, and the image was cropped and scaled; S3, select the kidney image on the severely damaged side; S4, data enhancement and normalization.

[0010] In S1, the background subtraction after integration was used for correction. The upper left area of ​​the left kidney and the upper right area of ​​the right kidney were selected as the background area, the region of interest was automatically divided, and the grayscale value of the background area was calculated and subtracted.

[0011] In S2, The automatic threshold selection technology based on the image histogram is used to outline the contours of the two kidneys and calculate the center lines of the two kidneys. The center lines are kept unchanged and the kidney images are processed symmetrically to obtain the original and symmetrical kidney annulus contours: Use a rectangular frame to include the outlines of both kidneys and crop the image to an image with an aspect ratio of 2:1; Scale the cropped image to a fixed size.

[0012] Each layer of the feature extraction module is provided with a sliding window transformer module.

[0013] The sliding window transformer module uses a 7×7 window size to divide the feature map into several non-overlapping windows, and each window contains several feature points. Multi-head self-attention is performed in each window to calculate the correlation between feature points and capture local features; In adjacent sliding window transformer modules, the window position will move by half the window size to achieve cross-window information interaction; The features processed by the attention mechanism are input into a multi-layer perceptron consisting of two fully connected layers and an activation function for nonlinear transformation; In each sliding window transformer module, residual connections and layer normalization are used.

[0014] The feature extraction module consists of 4 stages: In stage 1, the input feature dimension is 96, including 2 sliding window transformer modules and 3 attention heads; In stage 2, the input feature dimension is 192, including 2 sliding window transformer modules and 6 attention heads; In stage 3, the input feature dimension is 384, including 6 sliding window transformer modules and 12 attention heads; In stage 4, the input feature dimension is 768, it contains 2 sliding window transformer modules and the number of attention heads is 24.

[0015] The beneficial effects of the present invention are as follows: based on the analysis of the static image of the kidney of dithiothiosuccinic acid, there is no need to use the traditional invasive examination of urinary cystourethrography, which fundamentally reduces the physical harm to patients, especially children, and reduces the discomfort and medical risks during the examination. At the same time, it does not generate additional radiation risks and is more suitable for long-term health monitoring of children. Traditional methods rely on manual judgment, which has certain subjectivity and inconsistency. Our method realizes automated analysis and prediction based on the network model of the multi-head attention mechanism, provides quantifiable and objective diagnostic results, and improves the accuracy and consistency of diagnosis. In our model, through the hierarchical window multi-head self-attention mechanism, key features can be extracted in the local window, and the interaction of cross-window information can be realized. This multi-layer nested structure helps to capture the deep features in the static image of the kidney of dithiothiosuccinic acid, so that the model can autonomously learn to distinguish the image features of vesicoureteral reflux and effectively deal with the complex structures in the image. The model can automatically learn the importance of features without a large number of preprocessing operations to achieve accurate classification effects. Therefore, our model can quantify the subjective judgment of doctors on static images of S-dimer kidneys, solving the problems of lack of quantitative standards, reliance on subjective judgment, and discomfort and radiation exposure to pediatric patients. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 It is a design flow chart of the present invention.

[0016] Figure 2 Schematic diagram of the multi-head attention mechanism model of the present invention.

[0017] Figure 3 Schematic diagram of image segmentation of the present invention.

[0018] Figure 4 These are images of vesicoureteral reflux. a is an image with vesicoureteral reflux, and b is an image without vesicoureteral reflux and with abnormal image display.

[0019] Figure 5 Schematic diagram of a model of a sliding window converter module used in an embodiment of the present invention.

[0020] Figure 6 It is a schematic diagram of the test results of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in 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.

[0022] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0023] like Figure 1 As shown, the present invention provides a deep learning prediction system for vesicoureteral reflux in children, which includes: An image acquisition unit, used for acquiring a static image of the succimer renal tissue of the child, wherein the static image of the succimer renal tissue can be input by a doctor; An image processing unit divides the acquired static image of the kidney of dithiosuccinic acid into regions of interest, performs background correction, and performs quantitative processing to obtain a processed kidney image with risks; The prediction model can be trained to obtain the kidney image processed by the image processing unit and output the classification result. The prediction model includes: The image block feature embedding module divides the input image into several non-overlapping image blocks through convolution operations, and maps each image block into an embedding vector of fixed dimension through the convolution layer; Several layers of feature extraction modules capture local features through a layered window multi-head self-attention mechanism, and insert an image block merging layer between each basic layer to downsample the feature map; The feature fusion layer fuses the features captured by each layer and outputs the classification result.

[0024] The specific design flow chart is as follows: Figure 1 shown.

[0025] The prediction model needs to be trained, and the data set is constructed as follows: A1. Patient screening criteria: Children with acute urinary tract infection (UTI) who meet the "Guidelines for the Diagnosis and Treatment of Urinary Tract Infection in Children (2021 Edition)" were selected to ensure that the clinical data were complete. These children had static renal static examinations with dithiothiocyanate to verify abnormalities and underwent voiding cystourethrography during hospitalization.

[0026] A2. Clinical data collection: Systematically collect the imaging data and report results of patients' succimer renal static images and voiding cystourethrography.

[0027] The image processing unit needs to pre-process the dithiothreitol kidney static image according to the following method: B1. Background correction: In order to reduce the error caused by the concentration of dithiothiocarbamate in normal tissue, the background subtraction method after integration is used for correction. By selecting the upper left area of ​​the left kidney and the right area of ​​the right kidney as the background area, the region of interest (ROI) is automatically divided by programming, and the gray value of the background area is calculated and subtracted to reduce the influence of background noise and improve image quality.

[0028] B2. Image cropping and scaling: Use an automatic threshold segmentation algorithm to outline the kidneys, keep the center lines of the kidneys, and process the kidney images symmetrically. Use a rectangular box to contain the kidney outline and keep appropriate pixel margins to prevent the region of interest from being cropped. Crop the image to a fixed aspect ratio and scale it to a uniform size to ensure the consistency of the input data.

[0029] B3. Selection of unilateral kidney images: Quantitative analysis of the regions of interest of both kidneys was performed, and the grayscale values ​​after background correction were calculated. In the vesicoureteral reflux group, the kidney images of the severely damaged side (high-grade vesicoureteral reflux) were selected. The validity of the selection criteria was verified by statistical tests to ensure that the selected images could more accurately reflect the actual reflux characteristics.

[0030] B4. Data enhancement and normalization: To improve the generalization ability of the model, data enhancement techniques such as random flipping, rotation, cropping, and color jittering are used to increase the diversity of data. Images are standardized to meet the input requirements of the model.

[0031] The prediction model is a network model based on a multi-head attention mechanism, such as Figure 2 As shown on the left, the method for medical image data processing can efficiently extract and classify medical image data. This method uses a hierarchical structure and a multi-head self-attention mechanism. Figure 2 On the right, multi-scale features are captured and important features are adaptively learned, thereby improving the performance and robustness of the model.

[0032] Model architecture design: A1. Feature embedding module: The input medical image is divided into image blocks, and each image block is converted into an embedding vector of fixed dimension through a convolutional layer. This process achieves downsampling, reduces the amount of computation, and obtains a feature map with spatial and channel information.

[0033] A2. Attention mechanism module: The core layer of the present invention is composed of multiple basic units, each of which integrates a feature interaction component and an information integration unit based on a multi-head attention mechanism. These components achieve effective capture of local information and feature communication between different regions by defining feature interactions within a specific region.

[0034] A3. Feature fusion layer: An image block merging layer is added between the basic layers. Through the pooling operation, the feature map is downsampled, the information of adjacent image blocks is fused, the number of channels is increased and the integrity of the information is maintained.

[0035] B.Model features: B1. Multi-head attention mechanism: The multi-head attention mechanism contained in each basic unit allows the model to focus on different parts of the image at the same time, so that it can process information of multiple scales in parallel. This mechanism not only enhances the understanding of local features, but also captures different types of dependencies through multiple "heads", increasing the model's ability to learn complex patterns.

[0036] B2. Hierarchical structure: Through hierarchical design, the model can capture multi-scale features and adaptively learn important features, thereby improving the performance and robustness of the model.

[0037] C. Model output: C1. Global feature fusion: Finally, the abstract features of each layer are fused into the global average pooling layer to extract global information.

[0038] C2. Classification results: The classification results are output through the fully connected layer to achieve accurate classification of medical images.

[0039] Through the above method, the present invention can effectively capture local and global features in medical images, realize the effective use of medical imaging data, optimize the training process, improve the accuracy of disease diagnosis, reduce the need for invasive examinations, and has high clinical application value.

[0040] Network model training based on multi-head attention mechanism The present invention provides a neural network model training process based on a multi-head self-attention mechanism to improve the recognition accuracy of children with vesicoureteral reflux. The method makes full use of the imaging features of the severely damaged side of the kidney through an optimized data set partitioning strategy and model training process, thus improving the generalization ability and prediction performance of the model.

[0041] A. Dataset partitioning strategy: The data set is divided into training set, validation set and test set. The training set and validation set use the 10-fold cross-validation method to further improve the generalization ability of the model and avoid overfitting.

[0042] B. Model training strategy: B1. Training data selection: During the model training process, the image of the severely damaged side of the kidney is selected as the input of the training data. This allows the model to focus on learning the key features related to vesicoureteral reflux and enhance the ability to identify pathological changes.

[0043] B2. Model testing phase: During the model testing phase, the imaging data of the patient's bilateral kidneys are input, and the model determines whether vesicoureteral reflux has occurred by comprehensively analyzing the features of the left and right sides.

[0044] C. Model optimization and evaluation: C1. Optimization algorithm: Use appropriate optimization algorithm and loss function for model training to improve the convergence speed and accuracy of the model.

[0045] C2. Performance evaluation: By evaluating the performance on the validation set and independent test set, we calculate the model's accuracy and other indicators to ensure that the model has good generalization ability and stability.

[0046] Through the above method, the present invention effectively utilizes the characteristics of medical imaging data, optimizes the training process of the model, improves the recognition accuracy of children with vesicoureteral reflux, reduces the dependence on invasive examinations, and has important clinical application value. Specific embodiment: Based on the sliding window transformer (swin_transfomer) neural network.

[0048] In this embodiment, Dataset selection criteria A1. Patient screening: 346 children with acute urinary tract infection (UTI) admitted to the Department of Pediatric Nephrology, the Second Affiliated Hospital of Wenzhou Medical University from January 2019 to January 2023 were selected. Patients must meet the domestic diagnostic criteria of the "Guidelines for the Diagnosis and Treatment of Urinary Tract Infection in Children (2021 Edition)" and have complete clinical data. All patients had static renal static examinations verified abnormalities with dithiothiocyanate and underwent voiding cystourethrography during hospitalization.

[0049] A2. Clinical data collection: Collect imaging data and report results of patients undergoing dithiothreitol succinate renal static scanning and urinary cystourethrography.

[0050] Data preprocessing B1. Succimer kidney static image data acquisition: Using uniformly purchased instruments, in accordance with standardized acquisition techniques and protocols, the static imaging data of the renal succimer was collected 1 hour after the patient was injected with radioactive drugs. All static renal succimer imaging data were uploaded to the PACS (Picture Archiving and Communication System, PACS) of the Second Affiliated Hospital of Wenzhou Medical University. Images were acquired at the same window width and window level to ensure the consistency of image parameters.

[0051] B2: Background correction method: Background correction principle: In order to reduce the error caused by the concentration of dithiothiosuccinic acid in normal tissues, the background subtraction method after integration is used for correction.

[0052] Background area selection: Traditional methods usually use the area below the kidney for subtraction, but related studies show that the intensity of the upper left area of ​​the left kidney and the area directly to the right of the right kidney is close to the intensity of the empty kidney area. Since the metabolic laws of dimercaptosuccinic acid and diethylenetriamine pentaacetic acid (DTPA) in the body are similar, it is more accurate to select these areas as background areas.

[0053] Specific steps: Use the program and automatic threshold segmentation algorithm to divide the region of interest, including the kidney area and the corresponding background area. Calculate the grayscale intensity of each region of interest. For each kidney region of interest, subtract the grayscale value of the corresponding background area to obtain the corrected image.

[0054] B3. Image cropping and scaling (such as Figure 3 shown): Kidney contouring: Use the automatic threshold selection technique based on the image histogram (Otsu's) to outline the contours of both kidneys. Calculate the center lines of both kidneys, keep the center lines unchanged, and perform symmetric processing on the kidney images to obtain the original and symmetric kidney contours.

[0055] Image cropping: Use a rectangular frame to include the outline of both kidneys, and keep at least 10 pixels on the top, bottom, left, and right sides of the rectangular frame to prevent the area of ​​interest from being cropped. Crop the image to an image with an aspect ratio of 2:1.

[0056] Image scaling: Scale the cropped images to a fixed size, 224 × 448 pixels, to unify the image size.

[0057] B4. Selection of unilateral kidney images: Selection of unilateral kidney images: The python program was used to quantitatively analyze the bilateral regions of interest of the static dimercaptosuccinic acid angiography images, and the grayscale values ​​obtained after background subtraction were calculated. Python was also used to compare the grayscale values ​​of the bilateral kidneys in the vesicoureteral reflux group (176 cases) after background subtraction. It was found that among the 102 cases with different grades of vesicoureteral reflux on the left and right sides, the grayscale values ​​of 95 cases with high-grade VUR were smaller than those on the low-grade side, and the grayscale values ​​of the other 7 cases were not much different on both sides. The McNemar test was performed on these 102 patients to compare the differences in this paired binary data. According to our grayscale value results, the p value of the McNemar test was 3.3*10-12, which was much smaller than the significance level (p<0.05), indicating that the grayscale value of the kidney on the high-grade vesicoureteral reflux side was smaller than that on the low-grade vesicoureteral reflux side, which was statistically significant. This indicates that in the vesicoureteral reflux group, the grayscale value of the kidney on the side with high-grade vesicoureteral reflux is lower, which may be related to inflammatory response or other pathological changes. In clinical practice, we use the high-grade side as the standard for formulating treatment plans. Since the purpose of the study is to identify whether there is vesicoureteral reflux from static dimercaptosuccinic acid development in children with acute urinary tract infection, in terms of kidney image selection, we selected the image on the side with high-grade vesicoureteral reflux after background subtraction in the vesicoureteral reflux group.

[0058] Vesicoureteral reflux group (176 patients): Among the 102 patients with different grades of vesicoureteral reflux on the left and right sides, the kidney image on the side with the most severe damage (high-grade vesicoureteral reflux) was selected. The reasons for selecting unilateral kidney images (such as Figure 4 As shown in the figure): Some cases of vesicoureteral reflux may appear healthy on the image, but in fact vesicoureteral reflux has occurred. This is mainly due to the severe damage to the other kidney, which causes ureteral reflux to occur in the damaged weak kidney. Figure 4As shown in (a), there is no reflux in the left and right kidneys, and the images of both kidneys are bright, complete and without any missing parts. This is the correspondence between normal images and reflux levels. Figure 4 (b) shows that both the left and right kidneys have high-grade reflux, and the left kidney image has obvious image loss, but the right kidney image does not show it. Therefore, choosing the side with more severe damage can better reflect the actual reflux characteristics.

[0059] B5. Data enhancement and normalization: Data enhancement: In order to increase the diversity of data and prevent the model from overfitting, the image is subjected to a variety of data enhancement processing: Resize images: Resize images to 1.2 times their original size to increase scale variation.

[0060] Random Horizontal Flip: Randomly flip the image horizontally to simulate the change of the position of the left and right kidneys.

[0061] Random rotation: Randomly rotate the image within the range of ±15 degrees to enhance the model’s robustness to angle changes.

[0062] Random Cropping and Scaling: Randomly crop parts of the image and scale them to the target size to increase the diversity of the image.

[0063] Color dithering: Randomly adjust the brightness and contrast of an image to simulate changes in imaging conditions.

[0064] Normalization: Convert image data into tensor form. Use predefined mean and standard deviation to standardize the image to meet the input requirements of the model.

[0065] B6: Random seed setting: To ensure the repeatability of the experimental results, a random seed is set to ensure the consistency of data partitioning and model initialization.

[0066] The structure of the prediction model constructed in this embodiment is as follows Figure 5 As shown in the figure, after four feature extraction and downsampling operations, each feature extraction module captures local features through a layered window multi-head self-attention mechanism. Each layer achieves cross-window feature interaction by shifting windows, allowing the model to effectively integrate global and local information. In addition, an image block merging layer is added between each layer to downsample the feature map, and finally the abstract features of each layer are fused into the global average pooling layer, and the classification results are output through the fully connected layer.

[0067] A. Model architecture design: A1. Image block feature embedding module (Patch Embedding): The preprocessed medical image (size 224×448 pixels) is used as the input of the model. A 4×4 convolution operation with a stride of 4 is used to divide the input image into several non-overlapping image blocks. Each image block is mapped to an embedding vector of fixed dimension through the convolution layer to achieve downsampling and reduce the amount of calculation. A normalization layer is added to the embedded feature data to stabilize the feature data and improve the stability of model training.

[0068] A2. Sliding Window Transformer Module: Use a 7×7 window size to divide the feature map into several non-overlapping windows, each containing 49 feature points. Window Multi-Head Self-Attention (W-MSA): Perform a multi-head self-attention mechanism in each window to calculate the correlation between feature points and capture local features. Shifted Window Multi-Head Self-Attention (SW-MSA): In adjacent sliding window transformer modules, the window position will move half the window size (i.e., move left and up) to achieve cross-window information interaction and capture a wider range of features. Multi-layer Perceptron: The features processed by the attention mechanism are input into a multi-layer perceptron (MLP) containing two fully connected layers and an activation function for nonlinear transformation to enhance the model's expressiveness. Residual Connection and Layer Normalization: In each module, residual connection and layer normalization are used to prevent gradient disappearance and promote the stability of model training.

[0069] A3. Patch Merging Layer: Inserted between each basic layer for downsampling operations. Merge adjacent 2×2 image blocks to reduce the size of the feature map (height and width are halved), while increasing the number of channels (usually doubled) to maintain feature integrity. Perform linear transformation and normalization on the merged features to adjust feature dimensions and reduce information loss.

[0070] B. Model structure design: B1. Hierarchical design: The model consists of four stages, each with different input and output feature dimensions, specifically: Stage 1: The input feature dimension is 96, including 2 sliding window transformer modules and 3 attention heads.

[0071] Stage 2: The input feature dimension is 192, including 2 sliding window transformer modules and 6 attention heads.

[0072] Stage 3: The input feature dimension is 384, including 6 sliding window transformer modules and 12 attention heads.

[0073] Stage 4: The input feature dimension is 768, including 2 sliding window transformer modules and 24 attention heads.

[0074] B2. Feature extraction and downsampling: In each stage, feature extraction is performed through the base layer module, and downsampling is performed through the image patch merging layer between stages.

[0075] C. Model output layer: C1. Global average pooling layer: The feature map of the last layer is subjected to global average pooling to obtain a feature vector of fixed length.

[0076] C2. Fully connected layer: The feature vector is input into the fully connected layer, and the final classification result is output to determine whether the child has vesicoureteral reflux.

[0077] D. Model parameter settings: D1. Embedding dimension: The initial embedding dimension is set to 96, and the feature dimension is gradually doubled as the stage progresses.

[0078] D2. Window size: The window size is set to 7, which not only ensures the capture of local features but also controls the computational complexity.

[0079] D3. Activation function: Use Gaussian Error Linear Unit (GELU) activation function in multi-layer perceptron to improve nonlinear expression ability.

[0080] D4. Normalization layer: Use the normalization layer (Layer Normalization, LayerNorm) to normalize the features and stabilize the training process.

[0081] The training process of the prediction model is as follows: A. Dataset division: A1. Initial division: The medical imaging dataset of 346 children with acute urinary tract infection was divided into training set and validation set in a ratio of 9:1. Training set and test set (90%): used for model training and cross-validation. Independent validation set (10%): reserved as test data for final model performance evaluation, not involved in training and cross-validation.

[0082] A2. Ten-fold cross validation: In 90% of the data in the training set, the ten-fold cross validation method is used: the data is divided into 10 equal parts, one of which is selected as the test set each time (accounting for 9% of the total data), and the remaining nine parts are used as the training set (accounting for 81% of the total data), and this cycle is repeated ten times.

[0083] This method ensures that each data sample is used for validation once, making full use of the data and improving the generalization ability of the model.

[0084] B. Model training: B1. Training data preparation: During the training process, for each patient, images of the side of the kidney with severe damage were selected as input data for the model.

[0085] B2. Model training process: Parameter setting: Use the pre-built sliding window transformer model, set the learning rate to 0.0001, the batch size to 8, the optimizer to select the optimizer with weight decay (Adam with Weight Decay Regularization, AdamW), and the weight decay to 0.01.

[0086] Training process: In each cross-validation fold, the training set is used for model training, and the test set is used to evaluate the performance of the model on unseen data. During the training process, the cross entropy loss function is used as the loss indicator. The early stopping mechanism is adopted, and the patience number is set to 20. When the performance of the model on the validation set no longer improves, the training is stopped in advance to prevent overfitting. In each fold of training, the model parameters with the best performance on the validation set are saved for subsequent evaluation.

[0087] C. Model testing: C1. Test data preparation: During the model testing phase, for each patient, the image data of both kidneys are input.

[0088] C2. Model prediction: Unilateral prediction: The model predicts the images of the left and right kidneys respectively and obtains the unilateral classification results. Bilateral comprehensive judgment: If the prediction results of both the left and right sides are 0, the comprehensive judgment is 0 (no vesicoureteral reflux); otherwise, it is judged as 1 (vesicoureteral reflux exists).

[0089] C3. Performance evaluation: Calculate the model's loss, unilateral accuracy, bilateral accuracy and other indicators on the validation set and independent test set. Calculate the average result of the ten-fold cross validation to evaluate the model's stability and generalization ability.

[0090] D. Model optimization and adjustment: D1. Learning rate adjustment: Use the stagnation-based learning rate reduction learning rate scheduler (Reduce LearningRate on Plateau, ReduceLROnPlateau) to automatically adjust the learning rate according to the validation set loss to improve model performance.

[0091] D2. Model parameter adjustment: According to the training and validation results, adjust the model's hyperparameters, such as learning rate, batch size, training rounds, etc., to optimize model performance.

[0092] Through the above specific implementation steps, the training and testing of the sliding window transformer model were completed, and the effectiveness and reliability of the model in identifying children with vesicoureteral reflux were verified, providing a powerful auxiliary tool for clinical diagnosis.

[0093] Training Results After completing the training and validation of the model, we evaluated the performance of the model, such as Figure 6 The specific results are as follows: A. 10-fold cross validation training set performance: A1. Training set accuracy: During the 10-fold cross validation process, the average accuracy of the model on the training set reached 0.8947. This shows that the model has a good performance in learning the characteristics of the training data and can effectively fit the training data.

[0094] B. 10-fold cross validation set performance: B1. Unilateral validation set accuracy: The validation set accuracy of the model on unilateral kidney images is 0.8621. This means that the model has a high accuracy in identifying whether vesicoureteral reflux exists in a unilateral kidney.

[0095] B2. Bilateral validation set accuracy: When considering both kidney images, the validation set accuracy is 0.7931. Although it is lower than the unilateral prediction, the model can still accurately determine whether the patient has vesicoureteral reflux.

[0096] C. Test set performance: C1. Accuracy of unilateral test set: On the independent test set, the accuracy of the model in unilateral kidney images reached 0.8182. This shows that the model still has good prediction ability on unseen data, reflecting the generalization performance of the model.

[0097] C2. Bilateral test set accuracy: When using bilateral kidney images for prediction, the test set accuracy is 0.7879. Although the accuracy is reduced, the model still has reference value in practical applications.

[0098] Results analysis: The sliding window transformer model achieved high accuracy on the training set, validation set, and test set, especially in the prediction of unilateral kidney images, showing the model's effective learning and recognition ability for vesicoureteral reflux-related features. The accuracy of unilateral prediction is higher than that of bilateral prediction. This may be because the model mainly focuses on the features of the severely damaged side during training, while in bilateral prediction, the features of the healthy side may interfere with the model's judgment. The accuracy of the test set is slightly lower than that of the training set and validation set, but it still remains at a high level, indicating that the model has good generalization ability and can adapt to unseen case data.

[0099] In addition, a variety of different deep learning models were compared, and the experimental results are shown in Table 1. Although other models also showed the possibility of judging vesicoureteral reflux in children, their performance was generally not as good as the sliding window transformer model. These results further confirm the unique advantages of the sliding window transformer model.

[0100] At the same time, experiments were conducted on different parameter configurations of the sliding window transformer model to optimize its performance and find the optimal parameter configuration. The experiments covered a variety of configuration versions, including micro, small, basic, and large. The configurations of different parameters differed in the embedding dimension, number of attention heads, and layer depth. According to the experimental results, as shown in Table 2, the model under the basic configuration performed best, surpassing other configuration versions.

[0101] The above experimental results show that the deep learning method based on the sliding window transformer model has good performance in identifying whether children with acute urinary tract infection have vesicoureteral reflux. The stable performance of the model on multiple data sets verifies its effectiveness and reliability, provides a valuable auxiliary tool for clinical diagnosis, reduces dependence on invasive examinations, and has important clinical application value.

[0102] At the same time, the network architecture based on the multi-head self-attention mechanism of the present invention has been experimentally proven to be replaced by multiple deep learning models such as residual network 34, residual network 50, Maximum Vision Transformer (MaxVit), Google network, densely connected convolutional network (DesNet), etc., but the prediction accuracy of these deep learning models after training is lower than that of our network architecture, as shown in the following table.

[0103]

[0104] The network proposed in this invention uses the base parameter amount, in addition to the optional tiny, small, and large parameter amounts. The differences between the tiny, small, base, and large versions are mainly concentrated in the embedding dimension, the number of attention heads, and the layer depth. These differences affect the model's expressiveness, computing requirements, and applicable scenarios. After experimental comparison, the experimental results of the base parameter amount are the best, as shown in the following table.

[0105]

[0106] The hyperparameters of the model used in the present invention, such as window size, embedding dimension, number of attention heads, layer depth, learning rate, optimizer, etc., can be set to appropriate parameters according to the specific situation. In addition, the multi-head self-attention mechanism can also be replaced with other attention mechanisms (such as global self-attention) to adapt to different task requirements. At the same time, the downsampling method and activation function of the image block merging layer can also be replaced or adjusted according to the task requirements to optimize the model performance.

[0107] The embodiments should not be regarded as limiting the present invention, but any improvements made based on the spirit of the present invention should be within the protection scope of the present invention.

Claims

1. A deep learning prediction system for vesicoureteral reflux in children, characterized by: It includes: An image acquisition unit, used for acquiring a static image of the succimer kidney of the child; An image processing unit divides the acquired static image of the kidney of dithiosuccinic acid into regions of interest, performs background correction, and performs quantitative processing to obtain a processed kidney image with risks; The prediction model can be trained to obtain the kidney image processed by the image processing unit and output the classification result. The prediction model include: The image block feature embedding module divides the input image into several non-overlapping image blocks through convolution operations, and maps each image block into an embedding vector of fixed dimension through the convolution layer; Several layers of feature extraction modules capture local features through a layered window multi-head self-attention mechanism, and insert an image block merging layer between each basic layer to downsample the feature map; The feature fusion layer fuses the features captured by each layer and outputs the classification result.

2. A deep learning prediction system for vesicoureteral reflux in children according to claim 1, characterized in that: The image processing unit preprocesses the dithiosuccinic acid kidney static image based on the following steps: S1, background correction of a static image of a succinimidyl kidney; S2, kidney contours were delineated on the background-corrected image, and the image was cropped and scaled; S3, select the kidney image on the severely damaged side; S4, data enhancement and normalization.

3. A deep learning prediction system for vesicoureteral reflux in children according to claim 2, characterized in that: In S1, the background subtraction after integration was used for correction. The upper left area of ​​the left kidney and the upper right area of ​​the right kidney were selected as the background area, the region of interest was automatically divided, and the grayscale value of the background area was calculated and subtracted.

4. The deep learning prediction system for vesicoureteral reflux in children according to claim 2, characterized in that: In S2, The automatic threshold selection technology based on the image histogram is used to outline the contours of the two kidneys and calculate the center lines of the two kidneys. The center lines are kept unchanged and the kidney images are processed symmetrically to obtain the original and symmetrical kidney annulus contours: Use a rectangular frame to include the outlines of both kidneys and crop the image to an image with an aspect ratio of 2:1; Scale the cropped image to a fixed size.

5. The deep learning prediction system for vesicoureteral reflux in children according to claim 1, characterized in that: Each layer of the feature extraction module is provided with a sliding window transformer module.

6. A deep learning prediction system for vesicoureteral reflux in children according to claim 5, characterized in that: The sliding window transformer module uses a 7×7 window size to divide the feature map into several non-overlapping windows, and each window contains several feature points. Multi-head self-attention is performed in each window to calculate the correlation between feature points and capture local features; In adjacent sliding window transformer modules, the window position will move by half the window size to achieve cross-window information interaction; The features processed by the attention mechanism are input into a multi-layer perceptron consisting of two fully connected layers and an activation function for nonlinear transformation; In each sliding window transformer module, residual connections and layer normalization are used.

7. A deep learning prediction system for vesicoureteral reflux in children according to claim 5 or 6, characterized in that: The feature extraction module consists of 4 stages: In stage 1, the input feature dimension is 96, including 2 sliding window transformer modules and 3 attention heads; In stage 2, the input feature dimension is 192, including 2 sliding window transformer modules and 6 attention heads; In stage 3, the input feature dimension is 384, including 6 sliding window transformer modules and 12 attention heads; In stage 4, the input feature dimension is 768, it contains 2 sliding window transformer modules and the number of attention heads is 24.

Images