Medical image classification method based on quantum convolutional neural network and transfer learning

Abstract

The invention provides a medical image classification method based on a quantum convolutional neural network and transfer learning, and relates to the technical field of medical image classification. According to the method, QCNN and transfer learning are fused, and a parameter-based quantum transfer learning algorithm QTL is provided. The method specifically comprises the steps that firstly, a quantum-classical parameter mapping scheme based on angle coding is designed, and classical parameters are mapped to a quantum space, so that a quantum convolutional neural network QCNN can fully utilize the advantages of quantum calculation, and the expression ability of a model is enhanced; secondly, a transfer learning algorithm is introduced, domain self-adaption is achieved through a maximum mean value difference method, the distribution difference between a medical image big data set and small sample data is recognized, the difference is reduced by adjusting model parameters, and therefore the generalization ability of the model in the small sample field is further improved. According to the method, the classification performance and generalization ability of the model in a small sample medical image classification task are remarkably improved.

Description

Technical Field

[0001] This application relates to the technical field of medical image classification, and particularly to a medical image classification method based on quantum convolutional neural network and transfer learning. Background Art

[0002] With the improvement and enhancement of medical devices, medical imaging technology has developed rapidly, and medical images also play an extremely important role in medical data. Different modality imaging, including X-ray imaging, magnetic resonance imaging (MRI), etc., has been widely used in the early detection, diagnosis, and treatment of diseases. Clinically, the interpretation of medical images is mainly completed by medical experts, such as radiologists and internists. However, with the continuous development of computer-aided means and quantum technology, using computer image processing technology to process medical images and assist doctors in qualitative and quantitative analysis of medical images, thereby promoting the accuracy and reliability of medical diagnosis, has gradually become the mainstream.

[0003] Among them, the intelligent classification of medical images based on computer vision is an important step for computer-aided doctors to understand medical images. However, due to the huge amount of medical image data, problems such as unbalanced image data, difficult-to-distinguish categories caused by complex pathological image features, and insufficient classification performance and generalization ability of existing medical image classification models exist. To solve the above problems, existing research mostly focuses on classical computers and classical machine learning algorithms, such as methods like convolutional neural network CNN, federated learning, etc. However, as Moore's Law gradually fails, the computing power of classical computers faces a bottleneck. Therefore, it has become particularly urgent to seek new and efficient computing methods to meet the growing demand for accurate classification tasks of medical images.

[0004] Quantum computing, as an emerging computing paradigm, its potential for massive information storage and efficient parallel computing capabilities has opened up new possibilities for the field of machine learning. Quantum machine learning, as an emerging research field at the intersection of quantum computing and machine learning, aims to utilize the advantages of quantum computing to improve the performance of machine learning algorithms.

[0005] The present invention proposes a medical image classification method based on a quantum convolutional neural network (QCNN) and transfer learning. On the one hand, the quantum convolutional neural network cleverly integrates the core concept of the convolutional neural network into quantum computing, and uses the superposition and entanglement characteristics of quantum states to achieve efficient extraction and representation of image features. Compared with the traditional convolutional neural network, the quantum convolutional neural network has advantages in reducing algorithm complexity, improving computing speed, and reducing energy consumption. On the other hand, considering that the classification performance of CNN and QCNN still has room for improvement when dealing with small-sample image classification tasks, a transfer learning algorithm is introduced. Transfer learning borrows the knowledge learned by the pre-trained model in the source domain, which can significantly improve the learning speed and performance of the target domain task. In addition, quantum transfer learning also has great advantages in dealing with challenging problems such as unbalanced data. Therefore, by integrating the quantum convolutional neural network and quantum transfer learning, making full use of the complementary advantages of the quantum convolutional neural network and quantum transfer learning, the processing speed, model complexity, and small-sample problems of the medical image classification model are improved, and further, a more fine-grained classification study of medical images is realized.

[0006] After retrieval, the application publication number is CN113592027B, a medical image classification method based on transfer learning, including step 1: obtaining medical images and preprocessing them; step 2: constructing different classification models based on the convolutional neural network and training them; step 3: based on the classification model in step 2, using the transfer learning algorithm to construct a transfer learning model; step 4: using the transfer learning model to classify the medical images in step 1. The present invention can be used in medical image classification and feature extraction. A trained high-quality classification model is selected as the source model to construct a transfer learning model, and then the transfer learning algorithm is used to fine-tune the parameters of the model to achieve image classification.

[0007] The present invention introduces quantum image technology compared with this patent. This patent uses traditional image processing methods and uses the convolutional neural network to extract image features, while this patent uses the quantum image representation model INCQI, and INCQI is more helpful for image compression and processing. Combining with the quantum convolutional neural network QCNN to process images, the total complexity of QCNN is O(n), while the complexity of the corresponding classical algorithm is O(2 2n ). Therefore, the complexity of the QCNN used in the present invention is much lower than that of the classical convolutional neural network. The performance differences between QCNN and the classical neural network in image classification tasks are obvious: 1. The image accuracy is improved. The performance of the quantum representation model in image classification tasks is significantly higher than that of the classical image classification method. The improvement of the classification accuracy reflects the quantum representation ability; 2. The time required to train the model is shortened. The shorter model training time indicates that the quantum image representation model has high efficiency in storing image information. Summary of the Invention

[0008] The present invention aims to solve the above problems of the prior art. A medical image classification method based on a quantum convolutional neural network and transfer learning is proposed. The technical solution of the present invention is as follows:

[0009] A medical image classification method based on a quantum convolutional neural network and transfer learning, comprising the following steps:

[0010] Convert the weight and bias parameters of the pre-trained convolutional neural network CNN model into quantum parameters through an angle-encoding-based quantum-classical parameter mapping scheme, initialize the quantum convolutional neural network QCNN model with the converted parameters using the quantum transfer learning algorithm QTL, reduce the distribution difference between the training data set and the target data set through the maximum mean discrepancy MMD algorithm, and perform medical image classification and remote diagnosis assistance in the Web medical expert system.

[0011] Further, the conversion of the weight and bias parameters of the pre-trained convolutional neural network CNN model into quantum parameters through the angle-encoding-based quantum-classical parameter mapping scheme specifically includes:

[0012] S11. Load the pre-trained CNN model: Let the weight parameter matrix of the model be W CNN , and the bias parameter vector be B CNN , extract the corresponding weight parameter matrix W m×n and bias parameter matrix B m of each layer of the convolutional neural network; (m and n respectively represent the rows and columns of the matrix)

[0013] S12. Convert the parameter range: Normalize each element of W m×n to the interval [0, 1]. Since the rotation angle of the quantum bit is in the interval [0, 2π], convert the weight parameters of the pre-trained CNN model to an appropriate range through the formula ; Similarly, let the classical bias parameter vector be b m , normalize each element of b m to the interval [0, 1], and then map it to the interval [0, 2π], and convert the bias parameters of the pre-trained CNN model to an appropriate range through the formula ;

[0014] S13. Apply the parameter mapping scheme: Map the converted weight and bias parameters to the quantum bits; Based on the angle-encoding-based mapping scheme, apply an R y rotation gate to the weight parameters, whose rotation angle is A(i, j), and implement angle encoding in the quantum circuit, that is ( represents the quantum state of the image data) Similarly, apply an R yA revolving door with a revolving angle of B(i) realizes angle encoding in a quantum circuit, that is

[0015] Furthermore, initializing the quantum convolutional neural network QCNN model with the converted parameters by using the quantum transfer learning algorithm QTL includes the following steps:

[0016] S21. Construct a quantum convolutional neural network: Use the quantum convolutional neural network QCNN model as the medical image classification model in the QTL target domain;

[0017] S22. Initialize QCNN parameters: Use transfer learning to map the weight and bias parameters of the pre-trained CNN model onto qubits as the initial parameters of QCNN;

[0018] S23. Quantum circuit training and optimization: After initializing the QCNN parameters, use the quantum natural gradient method to optimize the rotation angle in the quantum circuit to adapt to the new medical image dataset or task; in each iteration, calculate the gradient of the loss function and continuously update the parameters in the quantum circuit, that is where L is the loss function, θ i is the rotation angle of the i-th revolving door, and μ is the learning rate;

[0019] S24. Verification and evaluation: During the process of training the QCNN model, a medical image test set is needed to verify the performance of the model, including the accuracy rate and the loss function value. When the performance of the model on the validation set reaches the expectation or the training process converges, stop training and evaluate the performance of the model on the test set.

[0020] Furthermore, reducing the distribution difference between the training dataset and the target dataset by using the maximum mean discrepancy MMD algorithm. MMD is a method for measuring the distance between two probability distributions. Given the source dataset S and the target dataset T, MMD aims to find a mapping function F such that the distance between the mapped source data distribution F(S) and the target data distribution F(T) is minimized, including the following steps:

[0021] S31 Add the loss function L new to MMD, that is L new = L + γ * MMD(F(S), F(T)), where E represents the expectation, represents the mapping function, x s and x t represent the samples of the training dataset and the target dataset respectively; L is the original loss function, and γ is a hyperparameter used to balance the weights between the original loss term L and the MMD loss term L new ;

[0022] S32 Update and optimization method; using the quantum natural gradient method to optimize the rotation angle θ in the quantum circuit i for optimization. In each iteration process, the angular variable of the loss function is updated accordingly, that is simultaneously optimize the original loss term L and the MMD loss term L new ;

[0023] S33 When the number of iterations is completed or convergence occurs, obtain the angular variable of the optimal loss function, and realize the self - adaptation of the model and the accurate classification of small - sample medical images.

[0024] Furthermore, perform medical image classification and remote diagnosis assistance in the Web medical expert system

[0025] including the following steps:

[0026] S41. Medical users use intelligent devices pre - installed with application programs to capture medical images of the lesion area, including skin lesion tissues;

[0027] S42. Directly send the medical images to the expert system based on the above algorithm architecture through the application program;

[0028] S43. Perform medical image classification according to the QCNN model in the expert system, and generate feedback to the user to obtain preliminary diagnostic opinions including the type of skin disease and whether urgent medical treatment is needed, so as to realize intelligent diagnosis of assisted remote medical images;

[0029] S44. At the same time, this system is also used for auxiliary diagnosis within the hospital after doctors and patients seek medical treatment; that is, for the imaging diagnosis made by patients in the hospital, the medical images can be automatically transmitted to this expert system for medical image classification by the model. At the same time, as the number of medical images increases, the model performs continuous learning and update, combined with the adaptive algorithm, to improve the classification accuracy of medical images and assist doctors in intelligent diagnosis of medical images.

[0030] The advantages and beneficial effects of the present invention are as follows:

[0031] Innovation points:

[0032] 1. Quantum convolutional neural network and classical parameter fusion: The present invention first converts the parameters of the classical convolutional neural network (CNN) into quantum parameters through an angle - encoding - based quantum - classical parameter mapping scheme for the construction of the quantum convolutional neural network (QCNN). This innovation point cleverly utilizes the advantages of quantum computing to transfer the knowledge of the classical model to the quantum model, enhancing the expressive ability and computational efficiency of the QCNN.

[0033] 2. Quantum Transfer Learning Algorithm QTL Based on Parameters: Through parameter transfer, the present invention proposes a quantum transfer learning algorithm based on the weights and bias parameters of a pre-trained model, which accelerates the training process of QCNN and improves the performance of the model in small-sample medical image classification tasks. This is the first attempt of quantum transfer learning in the field of medical image classification, filling the gap in the combination of quantum computing and transfer learning.

[0034] 3. Maximum Mean Discrepancy Algorithm MMD in Quantum Transfer Learning: Incorporating MMD into quantum transfer learning is used to reduce the distribution difference between the training data set and the target data set, thereby achieving regional adaptability and enhancing the generalization ability of the QCNN model in the target domain. This is the first combination of quantum transfer learning and domain adaptation algorithms, providing a new perspective for solving the small-sample problem in the field of medical images.

[0035] Beneficial Effects:

[0036] 1. Improved Classification Performance: The solution proposed by the present invention can significantly improve the model performance and generalization ability in small-sample medical image classification tasks. Experimental results show that the average classification accuracy has increased by 4.24%. For a field like medical images that is sensitive to data, this improvement is particularly crucial.

[0037] 2. Optimized Computational Efficiency: By constructing a quantum convolutional neural network, the present invention utilizes the superposition and entanglement characteristics of quantum states to achieve efficient extraction and representation of image features, reduces the algorithm complexity, improves the computational speed, and reduces energy consumption. This is particularly important for processing large data sets and high-dimensional images.

[0038] 3. Enhanced Small-Sample Learning Ability: In the case of data imbalance or limited sample quantity, the solution of the present invention can better identify and adjust the distribution differences between different data sets, improving the learning and diagnostic capabilities of the model in the small-sample field.

[0039] 4. Achieved Remote Intelligent Diagnosis Assistance: The Web-based medical expert system can conveniently obtain images from medical users for real-time classification and preliminary diagnosis, providing effective technical support for telemedicine. Especially in some areas with scarce medical resources, it can improve the diagnosis efficiency and quality.

[0040] Not Easily Thought of:

[0041] 1. Integration of Quantum Computing and Classical Deep Learning: Quantum computing and classical deep learning are two relatively independent fields. Combining the two requires a profound understanding of the principles of quantum computing and the architecture of deep learning models, which is not common in the current interdisciplinary field.

[0042] 2. Implementation of Quantum Transfer Learning: The mechanism of quantum transfer learning is not yet fully mature. Especially in the specific application of medical image classification, challenges such as quantum state encoding, quantum circuit design, and optimization need to be addressed.

[0043] 3. Design of Quantum Algorithms for Domain Adaptation: Combining the Maximum Mean Discrepancy (MMD) algorithm with the quantum transfer learning algorithm to achieve domain adaptation requires a deep understanding of MMD and, at the same time, innovative design of quantum circuits to adapt to the computational requirements of MMD.

[0044] 4. Integration of Cross - domain Knowledge and Technologies: This invention involves multiple fields such as medical image analysis, quantum computing, deep learning, and transfer learning, and requires cross - domain knowledge integration and technological innovation, which is not easily thought of by researchers in traditional single fields.

[0045] 5. Combination of Theory and Practice: Transforming these theoretical innovations into a practical Web - based medical expert system requires overcoming numerous technical problems from theoretical design to system implementation, including data pre - processing, model deployment, user interface design, etc., which is also not easily thought of and implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flowchart of the medical image classification method based on quantum convolutional neural network and transfer learning proposed in the preferred embodiment provided by the present invention;

[0047] Figure 2 is a flowchart of the medical expert system based on quantum medical images and the intelligent diagnosis of assisted remote medical images proposed by the present invention;

[0048] Figure 3 is an example diagram of three categories of positive images, normal lung images, and other pneumonia images in the image data set in the example of the present invention;

[0049] Figure 4 is a comparison table of the accuracy of the model in the example of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0051] The technical solution for the present invention to solve the above - mentioned technical problems is:

[0052] This embodiment provides a medical image classification method based on quantum convolutional neural network and transfer learning. The specific process is as Figure 1 shown, including the following steps:

[0053] S1 proposes a quantum-classical parameter mapping scheme based on angle encoding. Using quantum angle encoding, the pre-trained model weights and bias parameters are converted into quantum states, that is, the medical image bias parameters are converted into quantum parameters, and the medical image classification model weight parameters are converted into quantum parameters;

[0054] S2 proposes a quantum transfer learning algorithm QTL based on parameters. Using the transfer learning method, pre-trained weights are obtained from a large medical image dataset and transferred to the QCNN model in the construction stage to achieve parameter transfer from CNN to QCNN;

[0055] S3 incorporates the maximum mean discrepancy algorithm MMD into QTL to achieve regional adaptability, thereby improving the generalization ability of QCNN on the target tasks of small sample datasets;

[0056] S4 builds a Web medical expert system based on the above algorithm architecture to realize the classification of medical images and assist in remote medical image diagnosis.

[0057] In this embodiment, in step S1, the medical image classification method based on quantum convolutional neural network and transfer learning is characterized in that the quantum-classical parameter mapping scheme based on angle encoding uses the angle encoding method to map the classical weight parameters and bias parameters of medical images to the rotation angles of quantum bits. The steps are as follows:

[0058] (1) Load the pre-trained CNN model: Let the weight parameter matrix of the model be W CNN , and the bias parameter vector be B CNN . It is necessary to extract the corresponding weights W m×n and bias parameters B m of each layer of the convolutional neural network;

[0059] (2) Convert the parameter range: Normalize each element of W m×n to the interval [0,1]. Since the rotation angle of the quantum bit is in the interval [0,2π], through the formula the weight parameters of the pre-trained CNN model are converted to an appropriate range. In the same way, the bias vector B m is normalized and mapped, and the bias parameters are converted to an appropriate range through the formula ;

[0060] (3) Apply the parameter mapping scheme: Map the converted weights and bias parameters to the quantum bits. Based on the angle encoding mapping scheme, apply an R y rotation gate to the weight parameters, and its rotation angle is A(i,j), which realizes angle encoding in the quantum circuit, that is . In the same way, apply an R yA revolving door with a revolving angle of B(i) realizes angle encoding in a quantum circuit.

[0061] In this embodiment, in step S2, the medical image classification method based on a quantum convolutional neural network and transfer learning is characterized in that the parameter-based quantum transfer learning algorithm QTL initializes the QCNN by parameter transfer using the weights and bias parameters of a pre-trained model, accelerating the model convergence process and improving the model performance. It includes the following steps:

[0062] (1) Construct a quantum convolutional neural network: Use the quantum convolutional neural network QCNN model as the medical image classification model in the target domain of QTL.

[0063] (2) Initialize the QCNN parameters: Use transfer learning to map the weights and bias parameters of the pre-trained CNN model onto qubits as the initial parameters of the QCNN.

[0064] (3) Quantum circuit training and optimization: After initializing the QCNN parameters, use the quantum natural gradient method to optimize the rotation angles in the quantum circuit to adapt to a new medical image dataset or task. In each iteration, the gradient of the loss function needs to be calculated, and the parameters in the quantum circuit are continuously updated. That is where L is the loss function, θ i is the rotation angle of the i-th revolving door, and μ is the learning rate;

[0065] (4) Verification and evaluation: During the process of training the QCNN model, a medical image test set is needed to verify the performance of the model, such as accuracy, loss function value, etc. When the performance of the model on the validation set reaches the expectation or the training process converges, stop training and evaluate the performance of the model on the test set.

[0066] In this embodiment, in step S3, the medical image classification method based on a quantum convolutional neural network and transfer learning is characterized in that the maximum mean discrepancy algorithm MMD is incorporated into QTL, aiming to reduce the distribution difference between the training dataset and the target dataset, achieve regional adaptability, improve the performance of the QCNN model on the target task, and achieve accurate classification of medical images. It includes the following steps:

[0067] (1) Add the loss function L new to MMD, that is, L new = L + γ * MMD(F(S), F(T)), where E represents expectation, represents the mapping function, x s and x trepresent samples of the training dataset and the target dataset respectively; L is the original loss function, and γ is a hyperparameter used to balance the weights between the original loss term L and the MMD loss term L new between;

[0068] (2) Update and optimization method. Use the quantum natural gradient method to optimize the rotation angle θ in the quantum circuit i During each iteration, the angle variable of the loss function is updated accordingly, that is Optimize both the original loss term L and the MMD loss term L new ;

[0069] (3) When the number of iterations is completed or convergence occurs, the angle variable of the optimal loss function is obtained, realizing the self - adaptation of the model and the accurate classification of small - sample medical images.

[0070] In this embodiment, in step S4, the medical image classification method based on the quantum convolutional neural network and transfer learning is characterized in that the quantum medical image medical expert system based on the above algorithm architecture aims to realize the classification of medical images and assist in the intelligent diagnosis of remote medical images. As Figure 2 shown, it includes the following steps:

[0071] (1) Medical users use intelligent devices pre - installed with application programs to capture medical images of the lesion area, such as skin lesion tissues, etc.;

[0072] (2) Directly send the medical image to the expert system based on the above algorithm architecture through the application program;

[0073] (3) Classify the medical image according to the QCNN model in the expert system and generate feedback to the user to obtain preliminary diagnostic opinions such as the type of skin disease and whether urgent medical treatment is needed, realizing the assistance in the intelligent diagnosis of remote medical images;

[0074] (4) At the same time, this system is also used for auxiliary diagnosis within the hospital after doctors and patients seek medical treatment. That is, for the imaging diagnosis made by patients in the hospital, the medical image can be automatically transmitted to this expert system for medical image classification by the model. At the same time, as the number of medical images increases, the model conducts continuous learning and updating, combined with the adaptive algorithm, to improve the classification accuracy of medical images and assist doctors in the intelligent diagnosis of medical images.

[0075] Furthermore, to verify the superiority of the algorithm, we selected 600 pneumonia positive images, 1300 normal lung images, and 1300 other pneumonia images from the image library. Among them, the training set I consists of 500 positive, 1200 normal, and 1200 other pneumonia images, and the test set C consists of 100 positive, 100 normal, and 100 other pneumonia images. Figure 3An example of an image dataset. The experimental environment is set with an AMD Ryzen7 PRO 6850HS processor and an Ubuntu 18.04 operating system. An experimental simulation test is conducted on Python: taking the training set I as the source domain and the test set C as the target domain, a transfer learning task is constructed, denoted as I→C

[0076] Furthermore, VGG16, Res Net18, Xception, and Efficient Net are used as pre-trained models. By comparing the performance with and without QTL processing, the experimental results show that for the four models without QTL processing, the average accuracy of the pre-trained models for the I→C task is 88.30%, and the accuracy of the model with relatively poor performance is only 86.17%;

[0077] The above models are processed using the QTL algorithm. First, the VGG16, ResNet18, Xception, and Efficient Net models pre-trained on the dataset I are loaded. Subsequently, the parameters of the pre-trained models are frozen and used as the initial parameters of the QCNN network in the target domain, and parameter fine-tuning is performed. Then, the performance of the QCNN is verified using the validation set C, and the iteration stops after the model converges. The experimental results are as follows, where V_QTL, R_QTL, X_QTL, and E_QTL respectively represent the models processed by the QTL algorithm. The experimental results are as Figure 4 shown:

[0078] For the task I→C, the accuracy of V_QTL is 92.12%, the accuracy of R_QTL is 90.82%, the accuracy of X_QTL is 92.90%, and the accuracy of E_QTL is 94.32%. Generally speaking, in the task I→C, the average accuracy of the four models is 92.54%, and the accuracy of the model with relatively strong performance reaches 94.32%. Compared with the models without QTL, the average classification accuracy is improved by 4.24%. The experimental results show that the quantum transfer learning algorithm based on parameters can significantly improve the classification performance and generalization ability of the model in the small-sample medical image classification task.

[0079] The systems, devices, modules, or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions.

[0080] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

[0081] The above embodiments should be understood as being only for the purpose of illustrating the present invention and not for limiting the scope of protection of the present invention. After reading the content described in the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A medical image classification method based on quantum convolutional neural network and transfer learning, characterized in that: The following steps are involved: The weights and bias parameters of the pre-trained convolutional neural network (CNN) model are converted into quantum parameters through a quantum-classical parameter mapping scheme based on angle encoding. The quantum convolutional neural network (QCNN) model is initialized with the converted parameters using the quantum transfer learning algorithm QTL. The distribution difference between the training data set and the target data set is reduced through the maximum mean difference (MMD) algorithm. Medical image classification and remote diagnosis assistance are also performed in the Web medical expert system.

2. The medical image classification method based on quantum convolutional neural network and transfer learning according to claim 1 is characterized in that: The weights and bias parameters of the pre-trained convolutional neural network (CNN) model are converted into quantum parameters through a quantum-classical parameter mapping scheme based on angle encoding, specifically including: S11. Load the pre-trained CNN model: Let the weight parameter matrix of the model be W CNN , the bias parameter vector is B CNN , extract the corresponding weight W of each layer of convolutional neural network m×n and the bias parameter B m ; m and n represent the rows and columns of the matrix respectively; S12. Conversion parameter range: W m×n Each element of is normalized to the interval [0,1]. Since the rotation angle of the quantum bit is in the interval [0,2π], the formula Convert the weight parameters of the pre-trained CNN model to an appropriate range; in the same way, let the classic bias parameter vector be b m , b m Each element of is normalized to the interval [0,1] and then mapped to the interval [0,2π] by the formula Convert the bias parameters of the pre-trained CNN model to the appropriate range; S13. Apply parameter mapping scheme: map the converted weight and bias parameters to the qubits; apply an R to the weight parameters based on the angle-encoded mapping scheme. y The revolving door, whose rotation angle is A(i,j), implements angle encoding in the quantum circuit, that is, Represents the quantum state of the image data. In the same way, an R is applied to the bias parameter y The revolving door, whose rotation angle is B(i), implements angle encoding in the quantum circuit, that is, 3. The medical image classification method based on quantum convolutional neural network and transfer learning according to claim 1 is characterized in that: The method of using the quantum transfer learning algorithm QTL to initialize the quantum convolutional neural network QCNN model with the converted parameters includes the following steps: S21. Constructing quantum convolutional neural network: Using quantum convolutional neural network QCNN model as the medical image classification model in the QTL target field; S22. Initialize QCNN parameters: Use transfer learning to map the pre-trained CNN model to the weights and bias parameters on the quantum bits as the initial parameters of QCNN; S23. Quantum circuit training and optimization: After initializing the QCNN parameters, the quantum natural gradient method is used to optimize the rotation angle in the quantum circuit to adapt to the new medical image dataset or task; in each iteration, the gradient of the loss function is calculated and the parameters in the quantum circuit are continuously updated, that is, Where L is the loss function, θ i is the rotation angle of the i-th revolving door, μ is the learning rate; S24. Verification and evaluation: During the training of the QCNN model, a medical image test set is needed to verify the performance of the model, including accuracy and loss function value. When the performance of the model on the verification set reaches expectations or the training process converges, stop training and evaluate the performance of the model on the test set.

4. The medical image classification method based on quantum convolutional neural network and transfer learning according to claim 1 is characterized in that: The maximum mean difference (MMD) algorithm is used to reduce the distribution difference between the training data set and the target data set. MMD is a method for measuring the distance between two probability distributions. Given a source data set S and a target data set T, MMD aims to find a mapping function F so that the distance between the source data distribution F(S) and the target data distribution F(T) after mapping is minimized, including the following steps: S31. Add loss function L to MMD new , that is, L new =L+γ*MMD(F(S),F(T)), where E stands for expectation. represents the mapping function, x s and x t Represent the samples of the training dataset and the target dataset respectively; L is the original loss function, and γ is a hyperparameter used to balance the original loss term L and the MMD loss term L new The weight between S32, update optimization method; use quantum natural gradient method to calculate the rotation angle θ in quantum circuit i Optimize, and in each iteration, the angle variable of the loss function is updated accordingly, that is, Simultaneously optimize the original loss term L and the MMD loss term L new ; S33. When the number of iterations is completed or converged, the angle variable of the optimal loss function is obtained to achieve model adaptability and accurate classification of small sample medical images.

5. The medical image classification method based on quantum convolutional neural network and transfer learning according to claim 1 is characterized in that: The medical image classification and remote diagnosis assistance in the Web medical expert system includes the following steps: S41. A medical user uses a smart device with a pre-installed application to capture medical images of a lesion area, including skin lesion tissue; S42, directly sending the medical image to the expert system based on the above algorithm architecture through the application; S43. Classify medical images according to the QCNN model in the expert system and generate feedback to the user to obtain preliminary diagnosis opinions including the type of skin disease and whether medical treatment is urgently needed, so as to realize intelligent diagnosis of auxiliary remote medical images; S44. At the same time, the system is also used within the hospital to assist in the diagnosis of doctors and patients after they receive medical treatment. That is, when patients undergo imaging diagnosis in the hospital, the medical images can be automatically transmitted to the expert system for model-based medical image classification. At the same time, as the number of medical images increases, the model continues to learn and update, and combined with adaptive algorithms, the classification accuracy of medical images is improved to assist doctors in intelligent diagnosis of medical images.

Images