System and method for early screening of hyperuricemia and gout based on lung CT plain-scan image

Abstract

The invention discloses a system and a method for early screening of hyperuricemia and gout based on lung CT (computed tomography) plain scanning images, belongs to the technical field of image processing, and aims to solve the problem that existing lung CT plain scanning cannot distinguish early bone changes of hyperuricemia and gout. The screening method comprises the following steps: step 1, acquiring a lung CT plain-scan image of a patient, and preprocessing the lung CT plain-scan image; and step 2, inputting the preprocessed lung CT plain scanning image into the BCNN network model, and outputting a hyperuricemia and gout bone early classification result. According to the invention, the BCNN network model is used to re-image the bone of the lung CT plain-scan image, so that early screening of hyperuricemia and gout bone damage can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a system and method for early screening of hyperuricemia and gout based on lung CT plain scan images. Background Art

[0002] Hyperuricemia is mainly caused by excessive production or reduced excretion of uric acid in the body. Gout is a common metabolic disease that causes urate crystals to deposit in joints, soft tissues, and kidneys, causing acute inflammatory reactions.

[0003] The global prevalence of gout is 1%-4%, and the incidence is 0.1%-0.3%. Gout is more common in men than in women, with a ratio of 3:1 to 10:1. The incidence and prevalence of gout increase every decade, with the prevalence increasing to 11%-13% and the incidence increasing to 0.4% in people over 80 years old. Patients often pay attention to it only when they experience pain, and then they are diagnosed and treated in the hospital. If gout can be diagnosed and treated in the early stage, and long-term negative uric acid balance is maintained, the soft tissue swelling, cortical irregularity and intraosseous tophi in the early lesions are reversible and can prevent progression to chronic tophi. Correspondingly, with the progression of hyperuricemia and gout, corresponding characteristic changes in bone occur, such as marginal bone erosion, narrowing of the joint space, and tophi in the soft tissue and bones. Paraarticular erosion with marginal protrusion and narrowing of the joint space in late-stage patients is irreversible. Subsequently, bone calcification, diaphyseal invagination deformity, and finally ankylosis and subluxation occur. In addition to affecting bones, hyperuricemia and gout can also cause complications such as cardiovascular disease, hypertension, and chronic kidney disease (CKD), seriously affecting patients' quality of life and work ability. Therefore, early diagnosis and treatment are crucial to preventing joint degeneration and other adverse complications.

[0004] According to the Chinese Multidisciplinary Expert Consensus on the Diagnosis and Treatment of Hyperuricemia-Related Diseases (2023 Edition), traditional methods for gout diagnosis mainly rely on clinical manifestations, laboratory tests, and imaging examinations. Identification of urate crystals from affected joint fluid is the definitive diagnostic test for diagnosing gout. In practice, this test is only available for a small number of patients diagnosed with gout: 11% in a study of health professionals. However, patients with gout may not meet these criteria when they first present. Serum urate concentrations may be reduced during an acute attack, and normal urate concentrations at this time do not exclude the diagnosis of gout. Dual-energy computed tomography (DECT) has sensitivity (78%-100%) and specificity (89%-100%) in identifying monosodium urate (MSU) deposits. In early gout, false-negative imaging may result if the MSU volume is low or the tophi are very small (less than 2 mm). Although DECT has become an accurate method for detecting MSU deposits and diagnosing gout in various clinical situations, DECT is cumbersome to operate and relatively expensive, making it impractical as a routine screening item. Routine CT examinations are relatively easy to perform, especially lung CT examinations, which are widely used in physical examinations or diagnosis of lung diseases. However, existing conventional CT scans cannot distinguish between hyperuricemia and early bone changes in gout. Summary of the invention

[0005] In response to the above problems, the present invention proposes a system and method for early screening of hyperuricemia and gout based on lung CT plain scan images, and uses a convolutional neural network deep learning model to re-image the bone in the lung CT plain scan images, thereby realizing the significance of early screening for hyperuricemia and gout.

[0006] To this end, the technical solution adopted in the present invention is:

[0007] A method for early screening of hyperuricemia and gout based on lung CT plain scan images, the screening method comprising the following steps:

[0008] Step 1: Obtain the patient's lung CT plain scan image and perform preprocessing;

[0009] Step 2: Input the preprocessed lung CT plain scan image into the BCNN network model and output the early classification results of hyperuricemia and gout bone.

[0010] Based on the above technical solution, after the lung CT plain scan image is preprocessed in step 1, the size of the lung CT plain scan image is unified to 128x128x128.

[0011] On the basis of the above technical solution, further, the specific process of step 2 includes the following steps:

[0012] Step 201: input the preprocessed lung CT plain scan image into two branch CNN networks in the BCNN network model to extract feature maps respectively;

[0013] Step 202: The feature maps output by the two branch CNN networks are input into the bilinear pooling layer for feature aggregation;

[0014] Step 203: The features after the aggregation processing in step 202 are input to the fully connected layer. The fully connected layer serves as the output layer, maps the input features to the final classification labels, and outputs the classification results.

[0015] On the basis of the above technical solution, further, in step 201, both branch CNN networks adopt ResNet18_3D network, and separate auxiliary losses are designed for the two branch ResNet18_3D networks respectively. The feature map obtained by the first branch ResNet18_3D network adopts the first cross entropy loss L 1 CE and label smoothing loss L LSL Convergence is achieved, and the loss function is expressed as follows:

[0016] L branch1 =L 1 CE +L LSL (1)

[0017]

[0018] Among them, y i represents the label of sample i, p i represents the probability that sample i is predicted as a positive class, K represents the total number of classes, ε represents the smoothing parameter, and s i Represents the probability distribution of the i-th sample belonging to each category;

[0019] The feature map obtained by the second branch ResNet18_3D network passes through the fully connected layer and uses the second cross entropy loss L 2 CE And loss L1, its loss function is expressed as follows:

[0020] L branch2 =L 2 CE +L1 (4)

[0021] in:

[0022]

[0023] In the above formula, is the predicted value of the i-th sample, and N represents the number of samples.

[0024] On the basis of the above technical solution, further, the features output by the bilinear pooling layer in step 202 are converged using the third cross entropy loss, and its loss function is expressed as follows:

[0025]

[0026] The loss function L of the first branch ResNet18_3D network branch1 , the loss function L of the second branch ResNet18_3D network branch2 And the loss function L of the bilinear pooling layer FCE The loss function of the BCNN network deep learning model is expressed as follows:

[0027] L loss =0.2×(L branch1 +L branch2 )+0.6×L FCE (8)

[0028] in:

[0029] L FCE =L 3 CE (9).

[0030] On the basis of the above technical solution, further, the specific process of step 202 includes:

[0031] (1) Input: The feature maps extracted by the two-branch ResNet18_3D network are classified according to the branch source and recorded as a set f = {f1, f2}. Each feature map contains C channels. The length, width and height of the feature map are recorded as H, W and D respectively. Each batch contains N data.

[0032] (2) Construct a bilinear feature tensor based on the transpose of the feature map f1 extracted by the first branch ResNet18_3D network and the feature map f2 extracted by the second branch ResNet18_3D network Calculate the bilinear feature tensor

[0033] (3) Reshape the bilinear feature tensor to N×C 2 Shape, perform square root operation on the reshaped feature tensor, that is, Where ε = 1e-10;

[0034] (4) Output: b.

[0035] The present invention also provides an early screening system for hyperuricemia and gout based on lung CT plain scan images, the screening system comprising:

[0036] An image acquisition module is used to acquire a lung CT plain scan image of a patient and pre-process the lung CT plain scan image;

[0037] The classification module is used to input the preprocessed lung CT plain scan images into the BCNN network model deployed in the background for classification, and output the early classification results of hyperuricemia and gouty bone.

[0038] Based on the above technical solution, the BCNN network model includes two branch ResNet18_3D networks, a bilinear pooling layer and at least one fully connected layer from front to back. The process of using the BCNN network model to classify the input preprocessed lung CT plain scan image includes:

[0039] The preprocessed lung CT plain scan images were input into two branch ResNet18_3D networks for feature extraction, and then the features output by the two branch ResNet18_3D networks were input into the bilinear pooling layer for feature fusion. The fused features were output to the fully connected layer for classification, and the early classification results of hyperuricemia and gouty bone were output.

[0040] On the basis of the above technical solution, further, the specific process of the bilinear pooling layer for feature fusion includes:

[0041] (1) Input: The feature maps extracted by the two-branch ResNet18_3D network are classified according to the branch source, recorded as a set f = {f1, f2}, each feature map contains C channels, the length, width and height of the feature map are recorded as H, W and D respectively, and each batch contains N data;

[0042] (2) Construct a bilinear feature tensor based on the transpose of the feature map f1 extracted by the first branch ResNet18_3D network and the feature map extracted by the second branch ResNet18_3D network Calculate the bilinear feature tensor

[0043] (3) Reshape the bilinear feature tensor to N×C 2 Shape, perform square root operation on the reshaped feature tensor, that is, Where ε = 1e-10;

[0044] (4) Output: b.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] 1. The present invention first discovered that the use of the BCNN network model to re-image conventional CT bone in the lungs can achieve early screening of hyperuricemia and gout bone damage.

[0047] 2. This application demonstrates the development and value of the BCNN network model to increase the early diagnosis of patients with hyperuricemia and gout based only on conventional CT images of the lungs, and the early diagnosis method of the present invention can currently achieve a classification effect of 67.53%, which is of positive significance for preventing joint degeneration and other adverse complications. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is the BCNN network structure diagram;

[0049] Figure 2 It is a CT image resized to 128x128x128 size;

[0050] Figure 3 is the CT image after grayscale processing;

[0051] Figure 4 It is the log file of the training process. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0053] This example discusses the use of a convolutional neural network (CNN) deep learning model to re-image conventional CT bone in the lungs, thereby providing early screening for hyperuricemia and gout.

[0054] 1. Research subjects

[0055] 1. Diagnostic criteria for hyperuricemia and gout The inclusion criteria for patients with hyperuricemia are a fasting blood uric acid level >420 μmol / L on two different days under a normal purine diet to diagnose HUA. The inclusion criteria for patients with gout are the 2015 American College of Rheumatology (ACR) / European League Against Rheumatism (EULAR) gout scoring standards.

[0056] 2. Healthy person standard

[0057] According to the above criteria, this research group selected 307 patients from the Third People's Hospital of Bazhou City, Hebei Province from January 2022 to December 2023, including 126 healthy people, 114 patients with simple hyperuricemia, and 67 patients with gout. The basic information, blood biochemical results, and lung CT plain scan images of the selected patients were collected, as shown in Table 1. This study was approved by the Ethics Committee of the General Hospital of the People's Liberation Army.

[0058] Table 1 Basic information and blood biochemical results of selected patients

[0059]

[0060]

[0061] *Chi-square test was used for gender, and Kruskal Wallis H test was used for other items.

[0062] 2. Research Methods

[0063] 1. Image processing

[0064] This experiment is based on the bilinear pooling BCNN network for classification. The screening method of this embodiment is to unify the patient's CT chest film (lung CT plain scan) to a size of 128x128x128 after preprocessing, and then send it to the two branches of the BCNN network for feature extraction. The feature maps output by the two branches are not only sent to a classifier respectively, but also sent to the bilinear pooling layer for feature aggregation. The features after aggregation are sent to a third classifier for prediction. The results of each classifier are used to calculate different loss functions to constrain network training, so that features with small structural differences can be obtained to obtain better classification.

[0065] 3. Backbone network: For the backbone part of the network, we chose to use ResNet. Resnet18 is a commonly used deep learning model proposed by He Kaiming and others from Microsoft Research in 2015. ResNet is modified based on the VGG19 network and adds residual units through a shortcut mechanism, thereby achieving a deeper network while avoiding the gradient degradation problem. The changes in ResNet are mainly reflected in the direct use of convolutions with a stride of 2 for downsampling, and replacing the fully connected layer with a global average pooling layer. An important design principle of ResNet is that when the feature size is reduced by half, the number of features is doubled to maintain the complexity of the network layer. ResNet networks have network structures of different depths, such as the ResNet network structures of different depths in Table 2. The depth of the network is crucial to the performance of the model. When the number of network layers is increased, the network can extract more complex feature patterns, so increasing the depth of the model can achieve better results. However, when the network depth increases to a certain extent, the network accuracy becomes saturated or even decreases. Therefore, a deeper network structure does not necessarily mean a better training effect. Therefore, in this experiment, we use the Resnet18 network structure.

[0066] Table 2 ResNet network structures of different depths

[0067]

[0068] Since the data processed in this experiment is three-dimensional lung CT data, the original ResNet needs to be modified into a three-dimensional ResNet, denoted as ResNet_3D. Table 3 shows the network structure of the ResNet_3D network with layer depths of 18 and 34 respectively. The initial layer is usually a single three-dimensional convolutional layer, followed by batch normalization and ReLU activation functions. Next is a sequence of four residual blocks, and the number of residual blocks and the depth of the feature map in each sequence will vary. The end of the network contains a global pooling layer that compresses the multi-dimensional feature map into a one-dimensional feature vector, followed by one or more fully connected layers for the final classification or regression task. In this experiment, the data features between normal people and patients with hyperuricemia and gout are slightly different. In order to avoid overfitting, dropout is added after each convolution.

[0069] Table 3 ResNet_3D network structure

[0070]

[0071] BCNN: Figure 1 The BCNN network structure is shown in the figure. The output images are sent to two convolutional neural networks for feature extraction. In this experiment, the convolutional neural network adopts the ResNet18_3D structure.

[0072] During the training process, in order to increase the network's sensitivity to slight differences in images, two branch CNN networks with the same structure are designed. Each branch CNN network is connected to a fully connected layer. The fully connected layer serves as the output layer, and the features extracted by the two branch CNN networks are directly mapped to the classification labels. The classification results output by the two fully connected layers are constrained for training with different loss functions. By jointly converging the losses of the two branches, the slight differences between the output features of the two branch CNN networks are reduced, the network can recognize slight features, and the accuracy of the final classification is increased.

[0073] Initializing the two branch CNNs with the same parameters can improve efficiency during training. The features extracted by the two branch networks of BCNN are input into the bilinear pooling layer. In this process, symmetric activation may lead to suboptimal solutions, so separate auxiliary losses are designed in the two branch CNNs to break the symmetry between the two feature extractors used in the bilinear pooling layer.

[0074] The bilinear pooling layer first multiplies the two features at the same position to obtain a matrix, adds the features of all positions in the matrix to obtain a new matrix, reshapes the matrix, and then performs square root and normalization operations. The final features can be used for classification.

[0075] The response algorithm process of the bilinear pooling layer is as follows:

[0076] (1) Input: The feature maps extracted by the two-branch ResNet18_3D network are classified according to the branch source and recorded as a set f = {f1, f2}. Each feature map contains C channels. The length, width and height of the feature map are recorded as H, W and D respectively. Each batch contains N data.

[0077] (2) Construct a bilinear feature tensor based on the transpose of the feature map f1 extracted by the first branch ResNet18_3D network and the feature map extracted by the second branch ResNet18_3D network Calculate the bilinear feature tensor

[0078] (3) Reshape the bilinear feature tensor to N×C 2 Shape, perform square root operation on the reshaped feature tensor, that is, Where ε = 1e-10;

[0079] (4) Output: b.

[0080] The set of features b after the square root of the output bilinear feature tensor is input into the fully connected layer. Each feature has a certain weight for the final classification. By training the neural network, the weight assigned to each feature can be automatically adjusted (this feature is possessed by the BCNN network). The set of features b is multiplied by their respective weights and the sum is obtained to obtain the final probability, which represents the possibility of belonging to the classification label.

[0081] Loss: In the experiment, separate auxiliary losses were designed for the two CNN branches to break the symmetry between the two feature extractors used in the bilinear pooling layer. The two independent CNNs were initialized with the same pre-trained parameters. The features obtained by the first CNN were subjected to the cross entropy loss and label smoothing loss after passing through the fully connected layer (the fully connected layer is only used during training). The feature map obtained by the second CNN uses the cross entropy loss and L1 loss. The feature maps obtained by the two CNNs are output after passing through the bilinear pooling layer and the cross entropy loss is used. The loss of the first CNN branch and the loss of the second CNN branch are added together as the loss function of the network.

[0082] The first cross entropy loss L 1 CE , the second cross entropy loss L 2 CE And the third cross entropy loss L 3 CE The calculation method of is the same, both are calculated by the following formula:

[0083]

[0084] The first branch CNN network loss function is expressed as follows:

[0085] L branch1 =L 1 CE +L LSL

[0086] The L1 loss is expressed as follows:

[0087]

[0088] L LSL The loss is expressed as follows:

[0089]

[0090] The second branch CNN loss function is expressed as follows:

[0091] L branch2 =L 2 CE +L1

[0092] The loss function of the BCNN network is as follows:

[0093] L loss =0.2×(L branch1 +L branch2 )+0.6×L FCE

[0094] L FCE =L 3 CE

[0095] Where: y i represents the label of sample i, p i represents the probability that sample i is predicted to be a positive class, is the predicted value of the i-th sample, ε represents the smoothing parameter, s i represents the probability distribution of the i-th sample belonging to each category, K represents the total number of categories, and N represents the number of samples.

[0096] 3. Experimental process

[0097] Results Tables 5 and 6 show the amount of data divided into two categories and three categories. In the three-category dataset, gout, normal, and hyperuricemia are set as labels 0, 1, and 2, respectively. In the two-category dataset, it is divided into two categories: normal and abnormal, with labels 1 and 0, respectively. Among them, abnormal is the combination of gout and hyperuricemia. The data of gout, hyperuricemia patients and normal people are 67, 126, and 114 cases, respectively. The amount of data is unbalanced, but in the learning process of the network, it is necessary to learn the three types of data equally as much as possible, so in the three categories, the training set is 51 cases, and the remaining data is the test set. For the two-category division, the data of gout and hyperuricemia are combined.

[0098] 3.1 Experimental Details

[0099] In this experiment, the open source PyTorch framework was used for model development and training. All models were trained on NVIDIA RTXA6000 GPU to make full use of its powerful computing power and memory resources to speed up training and improve model performance. To optimize the model, the Adam optimizer was used. The Adam optimizer combines the advantages of Momentum and RMSProp, and dynamically adjusts the learning rate of each parameter by calculating the first-order momentum and second-order momentum of the gradient. This method enables the Adam optimizer to show good convergence and stability when dealing with different gradient amplitudes, effectively accelerating the training process of deep learning models. Before training begins, pre-trained weights are used to initialize the network. Pre-trained weights usually come from models pre-trained on large-scale datasets. These pre-trained models have learned rich feature representations and can serve as good initialization points, thereby accelerating convergence and improving the performance and training efficiency of the model on new tasks. In this way, we can use existing knowledge to train our model more efficiently. Specifically, the pre-trained ResNet18_3D model is loaded. In this way, the model not only retains the common features learned, but can also be fine-tuned for new datasets to further optimize performance.

[0100] 3.2 Evaluation Metrics

[0101] In our work, we distinguish between normal and abnormal conditions (high uric acid, gout) from CT images. There are many measurement criteria in the binary classification task of medical images, among which accuracy (ACC) is the key indicator for evaluating classification tasks. In addition, there are other indicators such as precision, recall, specificity, and score. In our work, we mainly use accuracy to measure the ability of our network.

[0102] The accuracy is the ratio of the number of correctly predicted samples to the total number of samples, and the formula is:

[0103]

[0104] The precision reflects the proportion of true positive samples among the samples predicted as positive by the model, and the formula is:

[0105]

[0106] The recall rate reflects the ability of the model to predict positive samples, and the formula is:

[0107]

[0108] Specificity reflects the model's ability to predict negative samples, and the formula is:

[0109]

[0110] The F1 score is defined as the harmonic mean of precision and recall, and the formula is:

[0111]

[0112] Among them: TP is the number of abnormal samples correctly predicted as abnormal.

[0113] TN is the number of normal samples correctly predicted as normal.

[0114] FP is the number of normal samples that are incorrectly predicted as non-normal.

[0115] FN is the number of non-normal samples that are incorrectly predicted as normal.

[0116] As shown in Table 4, the confusion matrix intuitively compares the predicted values ​​with the true values, where the values ​​on the main diagonal are TP and TN, and the values ​​on the non-main diagonal are FP and FN. The confusion matrix can clearly show the classifier's predictions for each category, and intuitively see which categories are easily confused. By analyzing the confusion matrix, we can better understand where the defects of the model are, providing a basis for further optimizing the model.

[0117] Table 4 Confusion matrix

[0118]

[0119] 3.3 Data Preprocessing

[0120] Table 5 and Table 6 show the data volume divided into two categories and three categories. In the three-category data set, gout, normal, and high uric acid are set as labels 0, 1, and 2 respectively. In the two-category data set, it is divided into two categories: normal and abnormal, with labels 1 and 0 respectively. Among them, abnormal is the collection of gout and high uric acid. The data of gout, high uric acid patients and normal people are 67, 126, and 114 cases respectively. The amount of data is unbalanced, but in the learning process of the network, it is necessary to learn the three types of data equally as much as possible, so in the three categories, the training set is 51 cases, and the remaining data is the test set. For the two-category division, the data of gout and high uric acid are merged. After investigation, it was found that gout patients and high uric acid patients are related to tophi and urate crystals. Tophi and urate crystals are easily found in joint synovium, cartilage, bone and soft tissue around joints, so the image is processed in grayscale. In this experiment, the bone CT value less than 200 is regarded as air. Figure 3 The image after grayscale processing is shown. Since the original images are of different sizes, they need to be resized to 256x256x256. Due to the large amount of data in the post-processing process, the video memory is overrun. Figure 2 The image is resized to 128x128x128 and then normalized.

[0121] Table 5 Binary data classification

[0122] Tag 0 Tag 1 Training set 51 cases of gout + 51 cases of hyperuricemia 51 cases normal Test Set 16 cases of gout + 63 cases of hyperuricemia 75 cases normal

[0123] Table 6 Three-category data division

[0124] Tag 0 Tag 1 Tag 2 Training set 51 cases of gout 51 cases normal 51 cases of high uric acid Test Set 16 cases of gout 75 cases normal 63 cases of high uric acid

[0125] 3.4 Experimental Results

[0126] First, Table 7 shows the influence of different layer depths of ResNet network on the experimental results. The experimental results show that ResNet10 may not be able to learn the characteristics of the data well. ResNet50 is oversaturated due to too many layers and does not bring better results. Then Table 8 shows that when testing the effect of ResNet18 network, it is easy to overfit, that is, it can learn good results during training, but it cannot get the corresponding effect on the test set. Therefore, dropout is added after each convolution layer of the ResNet18 network structure to prevent overfitting. The ResNet18 network with dropout solves the overfitting phenomenon in terms of effect, but the test results are low. This may be that the network has not been able to fully learn the characteristics of the image. Therefore, the BCNN network structure is adopted. Table 9 adopts the BCNN network, which does bring certain effects on the test set. Finally, considering that the difference between gout and hyperuricemia patients on CT images may be small, gout and hyperuricemia patients are combined into one category as abnormal patients. Table 10 can achieve a classification effect of 67.53% on the test set.

[0127] Table 7 Effects of different structural depths of Resnet on three classifications

[0128]

[0129] Table 8 Effect of adding dropout on three-classification training

[0130]

[0131] Table 9 Effect of modifying the dual-branch network for three-classification training

[0132]

[0133] Table 10 Effect of adding data threshold processing on classification

[0134]

[0135] result:

[0136] 1. In the study population, according to statistics, there were no significant differences in the nine items of age, white blood cell count, RBC, HGB hemoglobin concentration, HCT hematocrit, ALP alkaline phosphatase, TBIL total bilirubin, DBIL direct bilirubin, and IDBIL indirect bilirubin among the three groups. There were significant differences in the eight items of gender, TP total protein, ALB albumin, TBA total bile acid, BUN urea, UA uric acid, CREA creatinine, and β2 microglobulin.

[0137] 2. In this study, first, Table 7 shows the influence of different layer depths of ResNet network on the experimental effect. The experimental results show that ResNet10 may not be able to learn the characteristics of the data well. ResNet50 is oversaturated due to too many layers and does not bring better results. Then Table 8 shows that when testing the effect of ResNet18 network, it is easy to overfit, that is, it can learn good results during training, but it cannot get the corresponding effect on the test set. Therefore, dropout is added after each convolution layer of the ResNet18 network structure to prevent overfitting. The ResNet18 network with dropout solves the overfitting phenomenon in terms of effect, but the test results are low. This may be that the network has not been able to fully learn the characteristics of the image. Therefore, the BCNN network structure is adopted. Table 9 uses the BCNN network, which does bring certain effects on the test set. Finally, considering that the difference between gout and hyperuricemia patients on CT images may be small, gout and hyperuricemia patients are combined into one category as abnormal patients. Table 10 can achieve a classification effect of 67.53% on the test set. Figure 4 The logs during the training process are shown, and the training loss has converged to a certain extent.

[0138] In this study, the development and value of deep learning models were demonstrated to enhance the early diagnosis of patients with hyperuricemia and gout based solely on lung CT plain scan images. The model can currently achieve a classification effect of 67.53%, which has positive significance for preventing joint degeneration and other adverse complications.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the protection scope of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A method for early screening of hyperuricemia and gout based on lung CT plain scan images, characterized in that: The screening method includes the following steps: Step 1: Obtain the patient's lung CT plain scan image and perform preprocessing; Step 2: Input the preprocessed lung CT plain scan image into the BCNN network model and output the early classification results of hyperuricemia and gout bone.

2. The method for early screening of hyperuricemia and gout based on lung CT plain scan images according to claim 1, characterized in that: After the lung CT plain scan images are preprocessed in step 1, the size of the lung CT plain scan images is unified to 128x128x128.

3. The method for early screening of hyperuricemia and gout based on lung CT plain scan images according to claim 2, characterized in that: The specific process of step 2 includes the following steps: Step 201: input the preprocessed lung CT plain scan image into two branch CNN networks in the BCNN network model to extract feature maps respectively; Step 202: The feature maps output by the two branch CNN networks are input into the bilinear pooling layer for feature aggregation; Step 203: The features after the aggregation processing in step 202 are input to the fully connected layer. The fully connected layer serves as the output layer, maps the input features to the final classification labels, and outputs the classification results.

4. The method for early screening of hyperuricemia and gout based on lung CT plain scan images according to claim 3, characterized in that: In step 201, both branch CNN networks use ResNet18_3D networks. Separate auxiliary losses are designed for the two branch ResNet18_3D networks. The feature map obtained by the first branch ResNet18_3D network uses the first cross entropy loss L 1 CE and label smoothing loss L LSL Convergence is achieved, and the loss function is expressed as follows: L branch1 =L 1 CE +L LSL (1) Among them, y i represents the label of sample i, p i represents the probability that sample i is predicted as a positive class, K represents the total number of categories, ε represents the smoothing parameter, and s i Represents the probability distribution of the i-th sample belonging to each category; The feature map obtained by the second branch ResNet18_3D network adopts the second cross entropy loss L 2 CE And the loss L1 converges, and its loss function is expressed as follows: L branch2 =L 2 CE +L1 (4) in: In the above formula, is the predicted value of the i-th sample, and N represents the number of samples.

5. The method for early screening of hyperuricemia and gout based on lung CT plain scan images according to claim 4, characterized in that: The features output by the bilinear pooling layer in step 202 are processed using the third cross entropy loss L 3 CE Convergence is achieved, and the loss function is expressed as follows: The loss function L of the first branch ResNet18_3D network branch1 , the loss function L of the second branch ResNet18_3D network branch2 And the loss function L of the bilinear pooling layer FCE The loss function of the BCNN network deep learning model is expressed as follows: L loss =0.2×(L branch1 +L branch2 )+0.6×L FCE (8) in: L FCE =L 3 CE (9) 6. The method for early screening of hyperuricemia and gout based on lung CT plain scan images according to claim 5, characterized in that: The specific process of step 202 includes: (1) Input: The feature maps extracted by the two-branch ResNet18_3D network are classified according to the branch source, recorded as a set f = {f1, f2}, each feature map contains C channels, the length, width and height of the feature map are recorded as H, W and D respectively, and each batch contains N data; (2) Construct a bilinear feature tensor based on the transpose of the feature map f1 extracted by the first branch ResNet18_3D network and the feature map extracted by the second branch ResNet18_3D network Calculate the bilinear feature tensor (3) Reshape the bilinear feature tensor to N×C 2 Shape, perform square root operation on the reshaped feature tensor, that is, Where ε = 1e-10; (4) Output: b.

7. An early screening system for hyperuricemia and gout based on lung CT plain scan images, characterized in that: The screening system includes: An image acquisition module is used to acquire a lung CT plain scan image of a patient and pre-process the lung CT plain scan image; The classification module is used to input the preprocessed lung CT plain scan images into the BCNN network model deployed in the background for classification, and output the early classification results of hyperuricemia and gouty bone.

8. The early screening system for hyperuricemia and gout based on lung CT plain scan images according to claim 7, characterized in that: The BCNN network model includes two branch ResNet18_3D networks, a bilinear pooling layer and at least one fully connected layer from front to back. The process of using the BCNN network model to classify the input preprocessed lung CT plain scan image includes: The preprocessed lung CT plain scan images were input into two branch ResNet18_3D networks for feature map extraction. Then, the feature maps output by the two branch ResNet18_3D networks were input into the bilinear pooling layer for feature fusion. The fused features were output to the fully connected layer for classification, and the early classification results of hyperuricemia and gouty bone were output.

9. The early screening system for hyperuricemia and gout based on lung CT plain scan images according to claim 7, characterized in that: The specific process of feature fusion in the bilinear pooling layer includes: (1) Input: The feature maps extracted by the two-branch ResNet18_3D network are classified according to the branch source and recorded as a set f = {f1, f2}. Each feature map contains C channels. The length, width and height of the feature map are recorded as H, W and D respectively. Each batch contains N data. (2) Construct a bilinear feature tensor based on the transpose of the feature map f1 extracted by the first branch ResNet18_3D network and the feature map f2 extracted by the second branch ResNet18_3D network Calculate the bilinear feature tensor (3) Reshape the bilinear feature tensor to N×C 2 Shape, perform square root operation on the reshaped feature tensor, that is, Where ε = 1e-10; (4) Output: b.

Images