A medical management system and method based on blockchain

Abstract

The present invention provides a medical management system and method based on blockchain, including a patient registration module, a patient information acquisition module, a patient information storage module, and a patient information query module. The patient registration module is used to receive the registration information sent by the patient, create a login account for the patient and a login password corresponding to the login account, and save the medical information of the patient to the blockchain. In this way, due to the decentralized attribute of the blockchain, each node of the blockchain realizes information transmission and management, achieving information interaction among various hospitals and departments. At the same time, due to the characteristic of leaving a complete record throughout the process of the blockchain, the medical information of the patient is perfectly preserved. Then, a login account for the patient and a login password corresponding to the login account are created through the registration information. Only by using the login account and the login password corresponding to the login account can the medical information of the patient himself be queried, reducing the risk of patient privacy leakage.

Description

Technical Field

[0001] The present invention relates to the medical field, and specifically to a medical management system and method based on blockchain. Background Art

[0002] Currently, with the continuous improvement of the informatization level of medical institutions, many hospitals have established medical management systems to store patients' medical information. The existing practice is to store patients' medical information in a computer for patients to freely query. However, this easily leads to the leakage of patients' privacy, affecting patients and their families. At the same time, since different hospitals and different departments within the same hospital follow different standards and protocols, patients' medical information cannot be well interacted and integrated. And when a patient changes hospitals for treatment, it is very difficult for the current treating doctor to comprehensively judge based on the patient's previous medical history and other medical information, resulting in the doctor being unable to accurately judge the condition and delaying the treatment time. Moreover, once the computer breaks down, patients' medical information will be lost, and the storage stability is low.

[0003] Most hospitals require patients to submit registration information for registration when they are admitted to the hospital in order to correspond patients' medical information with the patients themselves. The registration information generally includes the patient's name, ID number, facial photo, etc. When nurses confirm whether the patient's medical information belongs to the current patient, it is the simplest and most direct to compare the patient himself / herself with the patient's facial photo. However, sometimes the patient's facial photo is very unclear, and nurses cannot make a comparison and confirmation, so they have to use other methods to confirm whether it is the patient himself / herself, resulting in low efficiency.

[0004] Common medical information includes high-resolution spectral images. It is necessary to classify the patient's high-resolution spectral images to achieve the diagnosis of the patient. The current classification method is to directly apply the original hyperspectral remote sensing data for classification. Although this classification method is feasible, the recognition accuracy is relatively low, which easily misdiagnoses patients and causes unnecessary losses and harms to patients. Summary of the Invention

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A medical management system based on blockchain, including a patient registration module, a patient information acquisition module, a patient information storage module, and a patient information query module;

[0006] The patient registration module is used to receive the registration information sent by the patient, create a login account for the patient and a login password corresponding to the login account, and send them to the patient;

[0007] The patient information acquisition module is used to collect the medical information of the patient and the exclusive number of the patient, analyze the medical information of the patient, and upload the analyzed medical information and the exclusive number of the corresponding patient to the patient information storage module;

[0008] The patient information storage module is used to store the registration information of the patient, the analyzed medical information of the patient and the exclusive number of the corresponding patient on the blockchain;

[0009] The patient information query module is used to retrieve the medical information of the patient corresponding to the exclusive number of the patient in the blockchain according to the exclusive number of the patient input by the received patient.

[0010] By saving the medical information of the patient to the blockchain, due to the decentralized property of the blockchain, each node of the blockchain realizes information transmission and management, realizes information interaction among various hospitals and departments, and improves the efficiency of doctor diagnosis and treatment; at the same time, due to the characteristics of the whole process traceability of the blockchain, the medical information of the patient is perfectly preserved and not easily lost; then, a login account for the patient and a login password corresponding to the login account are created through the registration information. Only through the login account and the login password corresponding to the login account can the medical information of the patient himself be queried, reducing the risk of patient privacy leakage and improving the ability to protect the privacy information of the patient.

[0011] Further, the registration information includes the ID number of the patient, the facial image of the patient and the name of the patient to ensure that the medical information of the patient corresponds to the patient himself.

[0012] Further, the facial image of the patient is processed to improve the resolution as follows:

[0013] Extract the facial image features of a low-resolution patient from the source network that shares the patient's facial image. The input of the source network that shares the patient's facial image is the facial image of the low-resolution patient. Use the convolutional layer and the feature extraction function of the facial image of the low-resolution patient to extract the facial image feature map of the low-resolution patient from the input facial image of the low-resolution patient; the source network that shares the patient's facial image includes a deep facial image feature extraction module of the patient connected in sequence; the connected-in-sequence block includes at least one mobile adaptive ratio residual small module of the patient's facial image, and the mobile adaptive ratio residual small module of the patient's facial image is composed of pointwise convolution of the patient's facial image, a parameter normalization layer of the patient's facial image, a maximum suppression activation function of the patient's facial image, and depth convolution of the patient's facial image; first, use the pointwise convolution of the patient's facial image to perform a convolution operation on the upper-layer output facial image feature map of the previous layer to obtain the first pointwise convolution facial image feature map of the patient, then perform parameter normalization on the first pointwise convolution facial image feature map of the patient, decompose the obtained ratio vector into two parts: a direction vector and a vector norm. This way of ratio decomposition fixes the distance norm of the ratio vector and achieves the effect of regularization, obtaining a regularized facial image feature map of the patient. Then, use the maximum suppression activation function of the patient's facial image to process the obtained regularized facial image feature map of the patient to obtain the first activated facial image feature map of the patient; use the depth convolution of the patient's facial image to perform a convolution operation on the first activated facial image feature map of the patient to obtain the depth facial image feature map of the patient; after obtaining the depth facial image feature map of the patient, perform parameter normalization and maximum suppression activation function processing of the patient's facial image again to obtain the second activated facial image feature map of the patient; use the pointwise convolution to perform a convolution operation on the second activated facial image feature map of the patient to obtain the second pointwise convolution facial image feature map of the patient; after obtaining the second pointwise convolution facial image feature map of the patient, perform parameter normalization processing; multiply the obtained facial image feature map of the patient by a scale function coefficient corresponding to an adaptive ratio and add it to the upper-layer output facial image feature map of the patient multiplied by a scale function coefficient corresponding to an adaptive ratio to obtain the output facial image feature map of the patient; the scale function coefficient parameter will be updated automatically during learning; connect the mobile adaptive ratio residual small modules of the patient's facial image in sequence to obtain the connected-in-sequence block; the source network that shares the patient's facial image includes an adaptive ratio sharing source module of the patient's facial image, and use this module to perform non-linear transformation processing. Among them, the adaptive ratio sharing source module of the patient's facial image is created by cascading and connecting the obtained connected-in-sequence blocks in sequence. There is an adaptive ratio scale function coefficient between the facial image feature map of the low-resolution patient and the output facial image feature map of each connected-in-sequence block;The facial image feature map of the low-resolution patient is multiplied by the scale function coefficients of the corresponding adaptive ratio, and then added to the facial image feature map of the patient obtained by multiplying the output facial image feature map of the successive connection block by the scale function coefficients of the corresponding adaptive ratio to obtain the facial image feature map of the patient, which is the input facial image feature map of the next successive connection block; the facial image of the patient is recreated using the shared source network; the low-resolution facial image of the patient input to the model is processed using a convolutional layer and then upsampled using a sub-pixel convolutional layer; the extracted low-resolution facial image feature map of the patient is input to the facial image adaptive ratio shared source module of the patient, and the output of the facial image adaptive ratio shared source module of the patient is operated on using a convolutional layer and then upsampled using a sub-pixel convolutional layer; the results of the two upsamplings are superimposed to obtain the recreated super-resolution facial image of the patient;

[0014] Furthermore, the cost function of the sparse regular operator is calculated for the recreated super-resolution facial image of the patient and the standard high-resolution facial image of the patient, and the parameters of the model are iterated using a supervised learning algorithm; then, the parameters of the model are continuously iterated using different pairs of high- and low-resolution facial images of the patient to obtain a well-trained model. Finally, the facial image of the patient that needs to be recreated with super-resolution is input to the well-trained model for operation, and the enlarged recreated facial image of the patient can be obtained.

[0015] Through the newly created deep facial image feature extraction module of the patient: the successive connection block, the non-linear transformation module: the facial image adaptive ratio shared source module of the patient, and the recreation module, more rapid and efficient recreation of the facial image of the patient can be achieved. At the same time, by introducing depthwise separable convolution and skip long and short skip connections, the present invention reduces the number of parameters and speeds up the calculation speed without overly affecting the final recreation effect, thereby enabling more rapid recreation of the super-resolution facial image of the patient. Moreover, the shared source network of the facial image of the patient in the present invention adopts an adaptive weight ratio, allowing more data to be extracted without increasing the parameters.

[0016] Furthermore, the patient registration module is specifically configured to receive the registration information sent by the patient, create a login account for the patient and a login password corresponding to the login account, store the login account and login password of the patient in the form of a mapping table, and send them to the patient.

[0017] Further, the patient information acquisition module analyzes the medical information of the patient, including classifying the high-resolution spectral image of the patient in the medical information, specifically as follows: First, an overall model is created. The overall model includes a backbone model, a self-attention mechanism unit, and a context encoder unit. The backbone model obtains hierarchical high-resolution spectral image features of the patient through three dilated convolutional layers and an average pooling layer; Subsequently, the hierarchical high-resolution spectral image features obtained by the backbone model are used as the input of the self-attention mechanism unit for self-attention iteration to create a spatial contrast relationship between pixels and obtain self-attention high-resolution spectral image features of the patient; The self-attention high-resolution spectral image features of the patient are then used as the input of the context encoder unit to iterate the global context high-resolution spectral image features of the patient; Specifically: Initialize and assign the parameters in the overall model to make it suitable for the Gaussian distribution; Perform a normalization operation on the original high-resolution spectral image of the patient to obtain a standardized high-resolution spectral image of the patient. Input the standardized high-resolution spectral image of the patient into the first dilated convolutional layer of the backbone model for operation to obtain the high-resolution spectral image features of the patient in the first dilated convolutional layer. Then, input the obtained high-resolution spectral image features of the patient in the first dilated convolutional layer into the second dilated convolutional layer for operation to obtain the high-resolution spectral image features of the patient in the second dilated convolutional layer. Next, input the obtained high-resolution spectral image features of the patient in the second dilated convolutional layer into the first pooling layer for average pooling operation to obtain the high-resolution spectral image features of the patient in the first pooling layer. Finally, input the obtained high-resolution spectral image features of the patient in the first pooling layer into the third dilated convolutional layer for operation to obtain the high-resolution spectral image features of the patient in the third dilated convolutional layer; Input the obtained high-resolution spectral image features of the patient in the third dilated convolutional layer into the self-attention mechanism unit to iterate the self-attention high-resolution spectral image features of the patient. Specifically: To reduce the computational burden in the iteration process of the self-attention high-resolution spectral image features of the patient, use an average pooling layer to perform an average pooling operation on the input high-resolution spectral image features of the patient in the third dilated convolutional layer to halve its spatial size and obtain the high-resolution spectral image features of the patient in the second pooling layer; Input the obtained high-resolution spectral image features of the patient in the second pooling layer into three convolutional layers respectively to obtain the corresponding first convolutional high-resolution spectral feature map, second convolutional high-resolution spectral feature map, and third convolutional high-resolution spectral feature map; Adjust the first convolutional high-resolution spectral feature map, second convolutional high-resolution spectral feature map, and third convolutional high-resolution spectral feature map to a preset size, and perform a product calculation on the first convolutional high-resolution spectral feature map and the second convolutional high-resolution spectral feature map to obtain a spatial attention high-resolution spectral image of the patient;Multiply the obtained high-resolution spectral image of the spatial attention patient by the high-resolution spectral feature map of the third convolution patient to obtain a new first high-resolution spectral feature map of the patient. Subsequently, adjust the new first high-resolution spectral feature map of the patient to a preset size; perform double bilinear interpolation upsampling calculation on the adjusted new first high-resolution spectral feature map of the patient to obtain a new second high-resolution spectral feature map of the patient. Then, use a convolutional layer to perform non-linear mapping calculation on the new second high-resolution spectral feature map of the patient to obtain the high-resolution spectral image feature of the self-attention patient; input the high-resolution spectral image feature of the self-attention patient iterated by the self-attention mechanism unit into the context encoder unit to iterate the high-resolution spectral image feature of the context patient. Specifically: use a convolutional layer to perform dimensionality reduction calculation on the input high-resolution spectral image feature of the self-attention patient to obtain a dimensionality-reduced high-resolution spectral feature map, and adjust the obtained dimensionality-reduced high-resolution spectral feature map to a preset size; use the global statistical information data in the adjusted dimensionality-reduced high-resolution spectral feature map to iterate the encoding table of the visual center, and calculate the standardized residual between the dimensionality-reduced high-resolution spectral feature map and the encoding table; perform batch normalization calculation on the calculated standardized residual to obtain a global context vector; use a fully connected layer to increase the dimensionality of the obtained global context vector to a preset dimension; perform calculation on the globally context vector after dimensionality increase through dot multiplication on the channel dimension to obtain the high-resolution spectral image feature of the context patient; perform double bilinear interpolation upsampling calculation on the calculated high-resolution spectral image feature of the context patient, and integrate the results with the high-resolution spectral image features of the first dilated convolutional layer patient and the second dilated convolutional layer patient in a concatenated manner to obtain the integrated high-resolution spectral image feature of the patient; input the obtained integrated high-resolution spectral image feature of the patient into a convolutional layer, use the exponential normalization function to obtain the probability map of the high-resolution spectral image of the patient predicted by the model, and calculate the cross-entropy classification cost function between the predicted probability map of the high-resolution spectral image of the patient and the true label; use the gradient descent optimization algorithm to iterate the cross-entropy classification cost function; repeat learning the overall model until the overall model is robust to obtain a well-learned overall model; input the target high-resolution spectral image of the patient to be recognized into the well-learned overall model to complete the classification of the high-resolution spectral image of the high-resolution spectral patient.;

[0018] By creating an overall model, which includes a backbone model, a self-attention mechanism unit, and a context encoder unit, the resistance of the model to adversarial samples can be effectively improved. Compared with the existing classification of high-resolution spectral images of patients based on deep convolutional neural networks, through the self-attention mechanism unit and the context encoder unit of the present invention, the spatial contrast relationship between pixels in the high-resolution spectral image of the patient is constructed, and the global context features of the high-resolution spectral image of the patient are obtained. On the high-resolution spectral data contaminated by adversarial attacks, excellent recognition accuracy can still be maintained.

[0019] Preferably, when storing the patient's registration information, the analyzed medical information of the patient, and the exclusive number of the corresponding patient on the blockchain, a timestamp and a hash value corresponding to this storage are also generated.

[0020] A medical management method based on blockchain includes the following steps:

[0021] Receive the registration information sent by the patient through the patient registration module, create a login account for the patient and a login password corresponding to the login account according to the registration information, and send them to the patient; collect the medical information of the patient and the exclusive number of the patient through the patient information acquisition module, analyze the medical information of the patient, and upload the analyzed medical information and the corresponding exclusive number of the patient to the patient information storage module; store the patient's registration information, the analyzed medical information of the patient, and the corresponding exclusive number of the patient on the blockchain through the patient information storage module; retrieve the medical information of the patient corresponding to the exclusive number of the patient in the blockchain according to the exclusive number of the patient input by the patient received through the patient information query module.

[0022] The present invention provides a medical management system and method based on blockchain. It has the following beneficial effects:

[0023] 1. By saving the medical information of the patient in the blockchain, due to the decentralized property of the blockchain, information transmission and management are realized among the nodes of the blockchain, and information interaction among various hospitals and departments is achieved, improving the efficiency of doctors' diagnosis and treatment; at the same time, due to the characteristics of the whole-process traceability of the blockchain, the medical information of the patient is perfectly saved and not easily lost; then a login account for the patient and a login password corresponding to the login account are created through the registration information, and only through the login account and the login password corresponding to the login account can the medical information of the patient himself be queried, reducing the risk of patient privacy leakage and improving the ability to protect the privacy information of the patient.

[0024] 2. Through the newly created deep patient facial image feature extraction module: sequentially connecting blocks, non-linear transformation modules: the patient facial image adaptive ratio sharing source module and the re-creation module can achieve faster and more efficient re-creation of the patient's facial image. At the same time, by introducing depthwise separable convolution and skip long and short skip connections, the present invention reduces the number of parameters and speeds up the calculation speed without overly affecting the final re-creation effect, thereby enabling faster super-resolution re-creation of the patient's facial image. Moreover, the patient facial image sharing source network in the present invention adopts an adaptive weight ratio, allowing more data to be extracted without increasing parameters.

[0025] 3. By creating an overall model, which includes a backbone model, a self-attention mechanism unit, and a context encoder unit, it can effectively improve the model's resistance to adversarial samples. Compared with the existing method of directly applying the original hyperspectral remote sensing data for classification, the present invention constructs the spatial contrast relationship between pixels in the patient's high-resolution spectral image through the self-attention mechanism unit and the context encoder unit, obtains the global context patient high-resolution spectral image features, and can still maintain excellent recognition accuracy on the high-resolution spectral data contaminated by adversarial attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0028] Please refer to Figure 1 , a blockchain-based medical management system, including a patient registration module, a patient information acquisition module, a patient information storage module, and a patient information query module;

[0029] The patient registration module is used to receive the registration information sent by the patient, create a login account for the patient and a login password corresponding to the login account, and send them to the patient;

[0030] The patient information acquisition module is used to collect the medical information of the patient and the exclusive number of the patient, analyze the medical information of the patient, and upload the analyzed medical information and the corresponding exclusive number of the patient to the patient information storage module;

[0031] The patient information storage module is used to store the patient's registration information, the analyzed medical information of the patient, and the exclusive number of the corresponding patient on the blockchain;

[0032] The patient information query module is used to retrieve the medical information of the patient corresponding to the exclusive number of the patient in the blockchain according to the exclusive number of the patient input by the received patient.

[0033] By saving the patient's medical information to the blockchain, due to the decentralized property of the blockchain, each node of the blockchain realizes information transmission and management, realizes information interaction among various hospitals and departments, and improves the efficiency of doctors' diagnosis and treatment; at the same time, due to the characteristics of the blockchain's full traceability, the patient's medical information is perfectly preserved and not easily lost; then, a login account for the patient and a login password corresponding to the login account are created through the registration information. Only through the login account and the login password corresponding to the login account can the medical information of the patient himself be queried, reducing the risk of patient privacy leakage and improving the ability to protect the patient's privacy information.

[0034] The registration information includes the patient's ID number, the patient's facial image, and the patient's name, ensuring that the patient's medical information corresponds to the patient himself.

[0035] Perform high-resolution processing on the patient's facial image as follows:

[0036] Extract the low-level and low-resolution facial image features of the patient using the facial image sharing source network of the patient. The input of the facial image sharing source network of the patient is the low-resolution facial image of the patient. Use the convolutional layer and the feature extraction function of the low-resolution facial image of the patient to extract the low-resolution facial image feature map from the input low-resolution facial image of the patient; the facial image sharing source network of the patient includes a deep facial image feature extraction module of the patient connected in sequence; the connected-in-sequence block includes at least one moving adaptive ratio residual small module of the facial image of the patient, and the moving adaptive ratio residual small module of the facial image of the patient is composed of pointwise convolution of the facial image of the patient, the parameter normalization layer of the facial image of the patient, the maximum suppression activation function of the facial image of the patient, and the depth convolution of the facial image of the patient; first, use the pointwise convolution of the facial image of the patient to perform a convolution operation on the upper-layer output facial image feature map of the previous layer to obtain the first pointwise convolution facial image feature map of the patient, and then perform parameter normalization on the first pointwise convolution facial image feature map of the patient. Decompose the obtained ratio vector into two parts: the direction vector and the vector norm. This way of ratio decomposition fixes the distance norm of the ratio vector and achieves the regularization effect to obtain the regularized facial image feature map of the patient. Then, use the maximum suppression activation function of the facial image of the patient to process the obtained regularized facial image feature map of the patient to obtain the first activated facial image feature map of the patient; use the depth convolution of the facial image of the patient to perform a convolution operation on the first activated facial image feature map of the patient to obtain the depth facial image feature map of the patient; after obtaining the depth facial image feature map of the patient, perform parameter normalization and the maximum suppression activation function processing of the facial image of the patient to obtain the second activated facial image feature map of the patient; use the pointwise convolution to perform a convolution operation on the second activated facial image feature map of the patient to obtain the second pointwise convolution facial image feature map of the patient; after obtaining the second pointwise convolution facial image feature map of the patient, perform parameter normalization processing; multiply the obtained facial image feature map of the patient by a corresponding scale function coefficient of the adaptive ratio and add it to the upper-layer output facial image feature map of the patient multiplied by a corresponding scale function coefficient of the adaptive ratio to obtain the output facial image feature map of the patient; the scale function coefficient parameter will be updated automatically during learning; connect the moving adaptive ratio residual small modules of the facial image of the patient in sequence to obtain the connected-in-sequence block; the facial image sharing source network of the patient includes an adaptive ratio sharing source module of the facial image of the patient, and use this module to perform non-linear transformation processing. Among them, the adaptive ratio sharing source module of the facial image of the patient is created by cascading and connecting the obtained connected-in-sequence blocks in sequence, and there is an adaptive ratio scale function coefficient between the low-resolution facial image feature map of the patient and the output facial image feature map of each connected-in-sequence block;The facial image feature map of the low-resolution patient is multiplied by the scale function coefficients of the corresponding adaptive ratio, and then added to the facial image feature map of the patient obtained by multiplying the output of the sequential connection block by the scale function coefficients of the corresponding adaptive ratio to obtain the facial image feature map of the patient, which is the input facial image feature map of the next sequential connection block; the facial image of the patient shares the source network for re-creation; the low-resolution facial image of the patient input to the model is processed using a convolutional layer and then upsampled using a sub-pixel convolutional layer; the extracted low-resolution facial image feature map of the patient is input to the facial image adaptive ratio sharing source module of the patient, and the output of the facial image adaptive ratio sharing source module of the patient is operated using a convolutional layer and then upsampled using a sub-pixel convolutional layer; the results of the two upsamplings are superimposed to obtain the re-created super-resolution facial image of the patient.

[0037] The cost function of the sparse regular operator is calculated for the re-created super-resolution facial image of the patient and the standard high-resolution facial image of the patient, and the parameters of the model are iterated using a supervised learning algorithm; then, the parameters of the model are continuously iterated using different pairs of high- and low-resolution facial images of the patient to obtain a well-trained model. Finally, the facial image of the patient that needs to be re-created with super-resolution is input to the well-trained model for operation, and the enlarged re-created facial image of the patient can be obtained.

[0038] Through the newly created deep facial image feature extraction module of the patient: the sequential connection block, the non-linear transformation module: the facial image adaptive ratio sharing source module of the patient, and the re-creation module, more rapid and efficient re-creation of the facial image of the patient can be achieved. At the same time, by introducing depthwise separable convolution and skip long-short skip connections, the present invention reduces the number of parameters and speeds up the calculation speed without overly affecting the final re-creation effect, thereby enabling more rapid re-creation of the super-resolution facial image of the patient. Moreover, the facial image sharing source network in the present invention adopts an adaptive weight ratio, allowing more data to be extracted without increasing the parameters.

[0039] The patient registration module is specifically configured to receive the registration information sent by the patient, create a login account for the patient and a login password corresponding to the login account, store the login account and login password of the patient in the form of a mapping table, and send them to the patient.

[0040] The patient information acquisition module analyzes the medical information of the patient, including classifying the patient's high-resolution spectral image in the medical information, specifically as follows: First, create an overall model. The overall model includes a backbone model, a self-attention mechanism unit, and a context encoder unit. The backbone model obtains hierarchical patient high-resolution spectral image features through three dilated convolutional layers and an average pooling layer; Subsequently, the hierarchical patient high-resolution spectral image features obtained by the backbone model are used as the input of the self-attention mechanism unit for self-attention iteration to create a spatial contrast relationship between pixels, resulting in self-attention patient high-resolution spectral image features; The self-attention patient high-resolution spectral image features are then used as the input of the context encoder unit to iterate the global context patient high-resolution spectral image features; Specifically: Initialize and assign the parameters in the overall model to make it suitable for the Gaussian distribution; Perform a normalization operation on the original patient high-resolution spectral image to obtain a patient standard high-resolution spectral image. Input the patient standard high-resolution spectral image into the first dilated convolutional layer of the backbone model for operation to obtain the patient high-resolution spectral image features of the first dilated convolutional layer. Then input the obtained patient high-resolution spectral image features of the first dilated convolutional layer into the second dilated convolutional layer for operation to obtain the patient high-resolution spectral image features of the second dilated convolutional layer. Then input the obtained patient high-resolution spectral image features of the second dilated convolutional layer into the first pooling layer for average pooling operation to obtain the patient high-resolution spectral image features of the first pooling layer. Finally, input the obtained patient high-resolution spectral image features of the first pooling layer into the third dilated convolutional layer for operation to obtain the patient high-resolution spectral image features of the third dilated convolutional layer; Input the obtained patient high-resolution spectral image features of the third dilated convolutional layer into the self-attention mechanism unit to iterate the self-attention patient high-resolution spectral image features. Specifically: To reduce the computational burden during the iteration of the self-attention patient high-resolution spectral image features, use an average pooling layer to perform an average pooling operation on the input patient high-resolution spectral image features of the third dilated convolutional layer to halve its spatial size and obtain the patient high-resolution spectral image features of the second pooling layer; Input the obtained patient high-resolution spectral image features of the second pooling layer into three convolutional layers respectively to obtain the corresponding first convolutional patient high-resolution spectral feature map, second convolutional patient high-resolution spectral feature map, and third convolutional patient high-resolution spectral feature map; Adjust the first convolutional patient high-resolution spectral feature map, second convolutional patient high-resolution spectral feature map, and third convolutional patient high-resolution spectral feature map to a preset size, and perform a product calculation on the first convolutional patient high-resolution spectral feature map and the second convolutional patient high-resolution spectral feature map to obtain a spatial attention patient high-resolution spectral image;Multiply the obtained high-resolution spectral image of the spatial attention patient by the third convolutional high-resolution spectral feature map of the patient to obtain a new first high-resolution spectral feature map of the patient. Subsequently, adjust the new first high-resolution spectral feature map of the patient to a preset size; perform double bilinear interpolation upsampling calculation on the new first high-resolution spectral feature map of the patient with the adjusted size to obtain a new second high-resolution spectral feature map of the patient, and then use a convolutional layer to perform non-linear mapping calculation on the new second high-resolution spectral feature map of the patient to obtain the high-resolution spectral image features of the self-attention patient; input the high-resolution spectral image features of the self-attention patient iterated by the self-attention mechanism unit into the context encoder unit to iterate the high-resolution spectral image features of the context patient. Specifically: use a convolutional layer to perform dimensionality reduction calculation on the input high-resolution spectral image features of the self-attention patient to obtain a dimensionality-reduced high-resolution spectral feature map of the patient, and adjust the obtained dimensionality-reduced high-resolution spectral feature map of the patient to a preset size; use the global statistical information data in the dimensionality-reduced high-resolution spectral feature map of the patient with the adjusted size to iterate the coding table of the visual center, and calculate the normalized residual between the dimensionality-reduced high-resolution spectral feature map of the patient and the coding table; perform batch normalization calculation on the calculated normalized residual to obtain a global context vector; use a fully connected layer to increase the dimensionality of the obtained global context vector to a preset dimension; perform calculation on the global context vector after dimensionality increase through dot multiplication on the channel dimension to obtain the high-resolution spectral image features of the context patient; perform double bilinear interpolation upsampling calculation on the calculated high-resolution spectral image features of the context patient, and integrate the results with the high-resolution spectral image features of the patient of the first dilated convolutional layer and the high-resolution spectral image features of the patient of the second dilated convolutional layer in a concatenated manner to obtain the integrated high-resolution spectral image features of the patient; input the obtained integrated high-resolution spectral image features of the patient into a convolutional layer, use the exponential normalization function to obtain the probability map of the high-resolution spectral image of the patient predicted by the model, and calculate the cross-entropy classification cost function between the predicted probability map of the high-resolution spectral image of the patient and the true label; use the gradient descent optimization algorithm to iterate the cross-entropy classification cost function; repeat learning the overall model until the overall model is robust to obtain a well-learned overall model; input the target high-resolution spectral image of the patient to be recognized into the well-learned overall model to complete the classification of the high-resolution spectral image of the high-resolution spectral patient.;

[0041] By creating an overall model, which includes a backbone model, a self-attention mechanism unit, and a context encoder unit, it can effectively improve the model's resistance to adversarial samples. Compared with the existing method of directly applying the original hyperspectral remote sensing data for classification, the present invention constructs the spatial contrast relationship between pixels in the patient's high-resolution spectral image through the self-attention mechanism unit and the context encoder unit, obtains the global context features of the patient's high-resolution spectral image, and can still maintain excellent recognition accuracy on the high-resolution spectral data contaminated by adversarial attacks.

[0042] When storing the patient's registration information, the analyzed medical information of the patient, and the patient's exclusive number on the blockchain, a timestamp and a hash value corresponding to this storage are also generated.

[0043] A medical management method based on blockchain includes the following steps:

[0044] Receive the registration information sent by the patient through the patient registration module, create a login account for the patient and a login password corresponding to the login account according to the registration information, and send them to the patient; collect the medical information of the patient and the patient's exclusive number through the patient information acquisition module, analyze the medical information of the patient, and upload the analyzed medical information and the corresponding patient's exclusive number to the patient information storage module; store the patient's registration information, the analyzed medical information of the patient, and the corresponding patient's exclusive number on the blockchain through the patient information storage module; retrieve the medical information of the patient corresponding to the patient's exclusive number in the blockchain according to the patient's exclusive number input by the patient received through the patient information query module.

[0045] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A blockchain-based medical management system, characterized in that: It includes a patient registration module, a patient information acquisition module, a patient information storage module, and a patient information query module; The patient registration module is used to receive the registration information sent by the patient, create a login account for the patient and a corresponding login password according to the registration information, and send them to the patient; The patient information acquisition module is used to collect the medical information of the patient and the exclusive number of the patient, analyze the medical information of the patient, and upload the analyzed medical information and the corresponding exclusive number of the patient to the patient information storage module; The patient information storage module is used to store the registration information of the patient, the analyzed medical information of the patient, and the corresponding exclusive number of the patient on the blockchain; The patient information query module is used to retrieve the medical information of the patient corresponding to the exclusive number of the patient in the blockchain according to the exclusive number of the patient input by the received patient; The registration information includes the patient's ID number, the patient's facial image, and the patient's name to ensure that the patient's medical information corresponds to the patient himself; Perform high-resolution processing on the facial image of the patient, specifically as follows: Extract the low-level and low-resolution facial image features of the patient using the facial image sharing source network of the patient. The input of the facial image sharing source network of the patient is the low-resolution facial image of the patient. Use the convolutional layer and the feature extraction function of the low-resolution facial image of the patient to extract the low-resolution facial image feature map from the input low-resolution facial image of the patient; the facial image sharing source network of the patient includes a deep facial image feature extraction module connected in sequence; the connected-in-sequence block includes at least one mobile adaptive ratio residual small module of the facial image of the patient, and the mobile adaptive ratio residual small module of the facial image of the patient is composed of pointwise convolution of the facial image of the patient, the parameter normalization layer of the facial image of the patient, the maximum suppression activation function of the facial image of the patient, and the depth convolution of the facial image of the patient; first, use the pointwise convolution of the facial image of the patient to perform a convolution operation on the upper-layer output facial image feature map of the previous layer to obtain the first pointwise convolution facial image feature map of the patient, then perform parameter normalization on the first pointwise convolution facial image feature map of the patient, decompose the obtained ratio vector into two parts: the direction vector and the vector norm. This way of ratio decomposition fixes the distance norm of the ratio vector and achieves the effect of regularization, obtaining the regularized facial image feature map of the patient. Then, use the maximum suppression activation function of the facial image of the patient to process the obtained regularized facial image feature map of the patient to obtain the first activated facial image feature map of the patient; use the depth convolution of the facial image of the patient to perform a convolution operation on the first activated facial image feature map of the patient to obtain the depth facial image feature map of the patient; after obtaining the depth facial image feature map of the patient, perform parameter normalization and the maximum suppression activation function processing of the facial image of the patient again to obtain the second activated facial image feature map of the patient; use the pointwise convolution to perform a convolution operation on the second activated facial image feature map of the patient to obtain the second pointwise convolution facial image feature map of the patient; after obtaining the second pointwise convolution facial image feature map of the patient, perform parameter normalization processing; multiply the obtained facial image feature map of the patient by a corresponding scale function coefficient of the adaptive ratio and add it to the upper-layer output facial image feature map of the patient multiplied by a corresponding scale function coefficient of the adaptive ratio to obtain the output facial image feature map of the patient; The scale function coefficient parameters will be updated automatically during learning; the patient's facial images are moved to adaptively scale the residual small modules and connected in sequence to obtain a sequentially connected block; the patient's facial image sharing source network includes a patient's facial image adaptive scale sharing source module, which is used for non-linear transformation processing. Among them, the patient's facial image adaptive scale sharing source module is created by cascading and sequentially connecting the obtained sequentially connected blocks. There is an adaptive scale function coefficient between the low-resolution patient's facial image feature map and the output patient's facial image feature map of each sequentially connected block; the patient's facial image feature map obtained by multiplying the low-resolution patient's facial image feature map by the corresponding adaptive scale function coefficient and adding it to the patient's facial image feature map obtained by multiplying the output patient's facial image feature map of the sequentially connected block by the corresponding adaptive scale function coefficient is the input patient's facial image feature map of the next sequentially connected block; the patient's facial image sharing source network is recreated; the low-resolution patient's facial image input to the model is processed using a convolutional layer and then upsampled using a sub-pixel convolutional layer; the extracted low-resolution patient's facial image feature map is input to the patient's facial image adaptive scale sharing source module, and the output of the patient's facial image adaptive scale sharing source module is operated using a convolutional layer and then upsampled using a sub-pixel convolutional layer; the results of the two upsamplings are superimposed to obtain the recreated super-resolution patient's facial image.

2. The medical management system based on blockchain according to claim 1, wherein The cost function of the sparse regular operator is calculated between the recreated super-resolution patient's facial image and the standard high-resolution patient's facial image, and the parameters of the model are iterated using a supervised learning algorithm; then, the model parameters are iterated using different pairs of high- and low-resolution patient's facial images to obtain a well-trained model. Finally, the patient's facial image that needs to be recreated with super-resolution is input to the well-trained model for operation, and the enlarged patient's facial image after recreation can be obtained.

3. The blockchain-based medical management system according to claim 1, characterized in that, The patient registration module is specifically used to receive the registration information sent by the patient, create a login account for the patient and a login password corresponding to the login account, store the patient's login account and login password in the form of a mapping table, and send them to the patient.

4. The medical management system based on blockchain according to claim 1, wherein: The patient information acquisition module analyzes the patient's medical information, including classifying the patient's high-resolution spectral image in the medical information, specifically as follows: First, create an overall model. The overall model includes a backbone model, a self-attention mechanism unit, and a context encoder unit. The backbone model obtains hierarchical high-resolution spectral image features of patients through three dilated convolutional layers and an average pooling layer. Subsequently, the hierarchical high-resolution spectral image features of patients obtained by the backbone model are used as the input of the self-attention mechanism unit for self-attention iteration to create a spatial contrast relationship between pixels, resulting in self-attention high-resolution spectral image features of patients. The self-attention high-resolution spectral image features of patients are then used as the input of the context encoder unit to iterate the global context high-resolution spectral image features of patients. Specifically: Initialize and assign the parameters in the overall model to fit the Gaussian distribution. Perform a normalization operation on the original high-resolution spectral image of the patient to obtain a standardized high-resolution spectral image of the patient. Input the standardized high-resolution spectral image of the patient into the first dilated convolutional layer of the backbone model for operation to obtain the high-resolution spectral image features of the patient in the first dilated convolutional layer. Then, input the obtained high-resolution spectral image features of the patient in the first dilated convolutional layer into the second dilated convolutional layer for operation to obtain the high-resolution spectral image features of the patient in the second dilated convolutional layer. Next, input the obtained high-resolution spectral image features of the patient in the second dilated convolutional layer into the first pooling layer for average pooling operation to obtain the high-resolution spectral image features of the patient in the first pooling layer. Finally, input the obtained high-resolution spectral image features of the patient in the first pooling layer into the third dilated convolutional layer for operation to obtain the high-resolution spectral image features of the patient in the third dilated convolutional layer. Input the obtained high-resolution spectral image features of the patient in the third dilated convolutional layer into the self-attention mechanism unit to iterate the self-attention high-resolution spectral image features of patients. Specifically: To reduce the computational burden during the iteration of the self-attention high-resolution spectral image features of patients, use an average pooling layer to perform an average pooling operation on the input high-resolution spectral image features of the patient in the third dilated convolutional layer to halve its spatial size and obtain the high-resolution spectral image features of the patient in the second pooling layer. Input the obtained high-resolution spectral image features of the patient in the second pooling layer into three convolutional layers respectively to obtain the corresponding first convolutional high-resolution spectral feature map, second convolutional high-resolution spectral feature map, and third convolutional high-resolution spectral feature map. Adjust the first convolutional high-resolution spectral feature map, second convolutional high-resolution spectral feature map, and third convolutional high-resolution spectral feature map to the preset size, and perform a product calculation on the first convolutional high-resolution spectral feature map and the second convolutional high-resolution spectral feature map to obtain a spatial attention high-resolution spectral image. Perform a product calculation on the obtained spatial attention high-resolution spectral image and the third convolutional high-resolution spectral feature map to obtain a new first high-resolution spectral feature map of the patient. Subsequently, adjust the new first high-resolution spectral feature map of the patient to the preset size;Perform double bilinear interpolation acquisition calculation on the newly adjusted first patient's high-resolution spectral feature map to obtain a new second patient's high-resolution spectral feature map, and then use a convolutional layer to perform non-linear mapping calculation on the new second patient's high-resolution spectral feature map to obtain the self-attention patient's high-resolution spectral image features; Input the self-attention patient high-resolution spectral image features iterated by the self-attention mechanism unit into the context encoder unit to iterate the context patient high-resolution spectral image features. Specifically: Use a convolutional layer to perform dimensionality reduction calculation on the input self-attention patient high-resolution spectral image features to obtain a dimensionality-reduced patient high-resolution spectral feature map, and adjust the obtained dimensionality-reduced patient high-resolution spectral feature map to a preset size; Use the global statistical information data in the dimensionality-reduced patient high-resolution spectral feature map with adjusted size to iterate the coding table of the visual center, and calculate the normalized residual between the dimensionality-reduced patient high-resolution spectral feature map and the coding table; Perform batch normalization calculation on the calculated normalized residual to obtain a global context vector; Use a fully connected layer to increase the dimension of the obtained global context vector to a preset dimension; Calculate the global context vector after dimension increase through dot multiplication on the channel dimension to obtain context patient high-resolution spectral image features; Perform double bilinear interpolation upsampling calculation on the calculated context patient high-resolution spectral image features, and integrate the results with the patient high-resolution spectral image features of the first dilated convolutional layer and the patient high-resolution spectral image features of the second dilated convolutional layer in a concatenated manner to obtain the integrated patient high-resolution spectral image features; Input the obtained integrated patient high-resolution spectral image features into a convolutional layer, and use the exponential normalization function to obtain the patient high-resolution spectral image probability map predicted by the model, and calculate the cross-entropy classification cost function between the predicted patient high-resolution spectral image probability map and the true label; Use the gradient descent optimization algorithm to iterate the cross-entropy classification cost function; Repeat learning the overall model until the overall model is robust to obtain a well-learned overall model; Input the target patient high-resolution spectral image to be recognized into the well-learned overall model to complete the classification of the high-resolution spectral patient high-resolution spectral image.

5. The medical management system based on blockchain according to claim 1, wherein When storing the patient's registration information, the analyzed patient's medical information, and the corresponding patient's exclusive number on the blockchain, a timestamp and a hash value corresponding to this storage are also generated.

Images