Mixed quantum ensemble learning and MedMama collaborative medical image classification method

Abstract

The invention relates to the technical field of medical image classification, in particular to a mixed quantum ensemble learning and MedMama collaborative medical image classification method, which comprises the following steps: inputting an original medical image into a preset U-Net feature extractor, extracting interpretable features related to a specific disease from the original medical image, and outputting corresponding feature vectors; inputting the output feature vector into a pre-trained hybrid quantum ensemble learning model, and outputting a corresponding first prediction classification probability; inputting the medical original image into a preset MedMama model, and outputting a corresponding second prediction classification probability; and calculating a final classification probability corresponding to the medical original image based on a preset medical image classification probability fusion calculation formula according to the output first prediction classification probability and the second prediction classification probability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image classification, and in particular to a medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration. BACKGROUND

[0002] Medical image classification is one of the key technologies for modern medical diagnosis and decision support. It automatically analyzes medical images (such as X-rays, CT, MRI, and pathological sections) through computer algorithms to achieve early screening, lesion positioning, and disease assessment, which is of great significance for improving diagnostic efficiency and reducing human error. With the development of deep learning technology, models such as CNN, Transformer, and others have been widely used in medical image classification tasks. For example, models such as U-Net, ResNet, and MedMamba have achieved certain classification results in specific scenarios by extracting deep features from images.

[0003] However, the existing medical image classification technology still has the following shortcomings: Deep learning models (such as CNN) rely on local feature extraction and have limited ability to capture complex and subtle lesion features (such as early tumors and micro lesions) in medical images. While pure Transformer models can model global dependencies, they have high computational costs when processing high-resolution medical images and lack sensitivity to local details. In addition, medical images often have large modality differences and strong noise interference, making it difficult for a single model to consistently extract discriminative and interpretable features, which limits classification accuracy.

[0004] Medical diagnosis requires high reliability of classification results, and the prediction results of a single model may be affected by noise and feature bias, leading to misjudgment.

[0005] In summary, the existing medical image classification has the problems of low accuracy and poor reliability of single model classification. SUMMARY

[0006] One of the purposes of the present application is to provide a medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration, which solves the problem of low accuracy and poor reliability of single model classification in the prior art.

[0007] In order to achieve the above purpose, a medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration is provided, comprising the following steps: S1, inputting a medical original image into a preset U-Net feature extractor to extract interpretable features related to a specific disease from the medical original image, and outputting a corresponding feature vector; S2, input the output feature vector into a pre-trained hybrid quantum integrated learning model, output the corresponding first predicted classification probability; the hybrid quantum integrated learning model takes multi-kernel quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) as base learners, and takes logistic regression as meta-learner; S3, input the medical original image into a preset MedMamba model, output the corresponding second predicted classification probability; the MedMamba model includes a patch embedding layer, a stacked SS-Conv-SSM block, a patch merging layer for down-sampling, and a feature classifier; S4, according to the output first predicted classification probability and second predicted classification probability, based on the preset medical image classification probability fusion calculation formula, calculate the final classification probability corresponding to the medical original image.

[0008] The technical principle and effect of the scheme: in the scheme, U-Net feature extraction is used to extract interpretable features related to specific diseases from medical original images. Through its symmetrical network structure, U-Net can effectively capture the context information and detail information of the image, and the output feature vector contains the key features related to the disease in the image, providing a basis for subsequent classification.

[0009] Taking multi-kernel quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) as base learners, the characteristics of quantum computing are used. MKQSVM combines quantum mechanics principles and support vector machines, which may have advantages in handling high-dimensional data and complex classification problems; QCNN introduces quantum convolution operation into traditional convolutional neural networks, which is expected to more efficiently extract features. Logistic regression as a meta-learner is used to integrate the output of the base learner, and by learning the relationship between the prediction results of the base learner and the true labels, the first predicted classification probability is output.

[0010] The MedMamba model includes a patch embedding layer, which is used to divide the input image into small blocks and map them to a specific dimension, preserving the 2D structure of the image. The stacked SS-Conv-SSM block is the core component, which divides the feature map into two groups, extracts local and global information through Conv-Branch and SSM-Branch respectively, and then restores the channel dimension and avoids information loss through channel connection and channel shuffle. The patch merging layer is used for down-sampling to reduce the data dimension, and finally the feature classifier outputs the second predicted classification probability according to the extracted features.

[0011] According to the first prediction classification probability output by the hybrid quantum ensemble learning model and the second prediction classification probability output by the MedMamba model, a preset medical image classification probability fusion calculation formula is used for fusion. This formula may be based on weighted average or other more complex fusion strategies to combine the advantages of the two models to obtain the final classification probability to determine the category to which the medical image belongs.

[0012] In this scheme, effective features are extracted by U-Net, and the hybrid quantum ensemble learning model and the MedMamba model are combined to process and classify features from different angles, and finally the probabilities are fused. It can fully utilize the advantages of various models and is expected to obtain higher accuracy in medical image classification tasks, especially for complex medical images such as images containing multiple lesions or multiple imaging modalities. The hybrid quantum ensemble learning path based on U-Net interpretable features is combined with the direct classification path based on advanced state space model (MedMamba), and the advantages are complementary through weighted fusion, breaking through the limitations of single method. The hybrid quantum ensemble learning model and the MedMamba model each have different feature extraction and classification mechanisms, and after fusion, the limitations of single model can be reduced, making the overall model have better adaptability to different types of medical images, improving the generalization ability of the model, and having good performance in different data sets and actual application scenarios, solving the problem of low accuracy and poor reliability of single model classification in the prior art.

[0013] The interpretable feature extraction with clinical significance (reflecting domain knowledge) and the potential of QML and the efficient global modeling ability of MedMamba are fused in a unified framework to achieve the dual improvement of performance and interpretability.

[0014] Further, the S1 comprises the following steps: Obtaining a medical original image, identifying a specific disease type corresponding to the medical original image according to the obtained medical original image; According to the identified specific disease type, the U-Net feature extractor corresponding to the specific disease type is retrieved, and the training process of the U-Net feature extractor corresponding to different specific disease types is different; According to the retrieved U-Net feature extractor, the medical original image is segmented into a region of interest, and at least one interpretable feature in the region of interest is extracted, including shape feature, texture feature, and frequency domain feature. The obtained interpretable features are normalized and finally combined into a corresponding feature vector.

[0015] Beneficial effects: By identifying the specific disease type corresponding to the medical original image first, and then calling the U-Net feature extractor dedicated to this disease, the specificity of different diseases in image performance can be fully utilized. Because the U-Net feature extractors of different diseases are trained differently (optimized for the image characteristics of each disease), they can more accurately capture key image information related to the disease, avoiding the omission or misjudgment of specific disease characteristics by general models, and laying a more reliable foundation for subsequent analysis.

[0016] Normalizing the interpretable features can eliminate the dimensional differences between different features, avoid a certain feature dominating the model analysis due to a too large value range, and ensure that each feature has equal weight and influence in subsequent modeling. Combining the processed features into a feature vector facilitates the use of the feature vector as input data for subsequent disease diagnosis, staging, or prognosis prediction tasks, improving the overall performance and stability of the model.

[0017] Further, the preset medical image classification probability fusion calculation formula is:

[0018] In the formula, is the final classification probability of the medical original image, is the first predicted classification probability, is the second predicted classification probability, is the fusion weight value of the first predicted classification probability.

[0019] Beneficial effects: This formula fuses the first predicted classification probability and the second predicted classification probability , which can combine the advantages of the two models. Different models may be based on different algorithm principles, data emphasis, or feature learning methods, and have their own strengths in medical image classification (for example, one model may be better at capturing detailed texture features, and the other may be better at global structure analysis). The final classification probability after fusion can compensate for the limitations of a single model, reduce misjudgments caused by inherent model bias, and thus improve the accuracy of overall classification.

[0020] Further, the training of the pre-trained hybrid quantum integrated learning model in S2 includes the following steps: S200, obtain different types of medical original historical image data sets, segment the region of interest through the U-Net feature extractor, and extract at least one interpretable feature in the shape feature, texture feature, and frequency domain feature in the region of interest. Normalize the obtained interpretable features to form a corresponding historical feature vector set, and divide the historical feature vector set into a training set and a validation set; S300, determining, according to the training set and the validation set, a quantum feature mapping function set corresponding to a multi-kernel quantum support vector machine (MKQSVM) and an optimal function combination weight corresponding to a quantum feature mapping function in the quantum feature mapping function set based on a preset multi-kernel learning optimization strategy and a plurality of to-be-selected quantum feature mapping functions, and completing training of the multi-kernel quantum support vector machine (MKQSVM); S400, training a quantum convolutional neural network (QCNN) according to the training set and the validation set, obtaining the trained multi-kernel quantum support vector machine (MKQSVM) and the quantum convolutional neural network (QCNN) after completing the training, combining prediction outputs of the multi-kernel quantum support vector machine (MKQSVM) and the quantum convolutional neural network (QCNN) on the validation set into a new feature set, learning the new feature set by taking a logistic regression as a meta-learner, and obtaining an HQEL model.

[0021] Beneficial effects: By the preset multi-kernel learning optimization strategy, the optimal combination is selected from a plurality of to-be-selected quantum feature mapping functions, quantum mechanics characteristics (such as superposition and entanglement) can be used to perform more efficient nonlinear transformation on features, hidden complex patterns in data can be mined, and the accuracy of a classification boundary can be improved. The weighted combination of different quantum feature mapping functions can take into account the advantages of a plurality of feature transformation modes and adapt to the diversity of medical image data.

[0022] The prediction outputs of the trained MKQSVM and QCNN on the validation set are combined into a new feature set, and then a logistic regression meta-learner is used for secondary learning to form an HQEL model. This integrated learning method combines the advantages of two different models: MKSVM is good at processing small samples and high-dimensional data, and QCNN performs outstandingly in extracting deep features of images. By fusing the prediction results of the two, the overfitting risk of a single model can be reduced, the adaptability of the model to different medical image scenarios can be enhanced, and the generalization ability and prediction stability can be improved.

[0023] Further, the preset multi-kernel learning optimization strategy includes the following steps: S301, a plurality of quantum feature mapping functions are used to form a corresponding quantum feature mapping function set; S302, a plurality of groups of function weight combinations are randomly generated , as initial search points, and n is the total number of quantum feature mapping functions in the quantum feature mapping function set; S303, according to the function weight combination of each group, the historical feature vectors in the training set are mapped into a quantum state space, different quantum feature mapping functions capture different quantum characteristics of the historical feature vectors; for each quantum feature mapping, a quantum kernel matrix and a quantum combination kernel matrix corresponding to each function weight combination are calculated; S304, train the quantum support vector separator using the corresponding quantum combined kernel matrix, and perform k-fold cross-validation on the validation set to calculate the cross-validation accuracy of each set of function weight combinations; S305, based on the evaluated function weight combinations and the corresponding cross-validation accuracy, construct a Gaussian process surrogate model, which is used to predict the target function value and its uncertainty at each point in the entire weight search space, and output the corresponding prediction result data; S306, according to the prediction result data, select the next set of function weight combinations to be evaluated based on the preset acquisition function data, and repeat S303 until the preset stopping condition is met, and select the function weight combination with the highest cross-validation accuracy from all the evaluated function weight combinations as the optimal function combination weight.

[0024] Beneficial effects: By adopting multiple quantum feature mapping functions to form a function set, the quantum characteristics of historical feature vectors can be captured from different angles, and more rich internal structures and patterns can be mined from data.

[0025] Randomly generate multiple sets of initial weight combinations as search starting points, and combine the Gaussian process surrogate model for iterative optimization, avoiding the blindness of traditional grid search or random search. The Gaussian process surrogate model can predict the performance and uncertainty of each point in the entire weight search space based on the information of the evaluated points, guide the intelligent selection of the next set of weight combinations to be evaluated, significantly reduce the computational overhead while ensuring the search quality, and improve the optimization efficiency.

[0026] Perform k-fold cross-validation on the validation set to more comprehensively and reliably evaluate the performance of each set of weight combinations. By dividing the validation set multiple times and calculating the average accuracy, the evaluation bias caused by data division randomness is reduced, ensuring that the selected optimal weight combination not only performs well on training data, but also has strong generalization ability, can adapt to unseen medical image data, and reduces the risk of overfitting.

[0027] The Gaussian process surrogate model not only predicts the target function value (cross-validation accuracy), but also provides the uncertainty information of the prediction. This uncertainty quantification helps to balance "exploitation" (selecting weight combinations in known high-performance areas) and "exploration" (trying unknown areas to find potentially better solutions) during the weight search process, making the optimization process more robust and avoiding getting stuck in local optimal solutions, thereby improving the probability of finding the global optimal weight combination.

[0028] The entire optimization process is based on the feedback of actual data (cross-validation accuracy) without relying on artificial experience or preset fixed parameters. By continuously updating the Gaussian process proxy model and selecting the points to be evaluated, the optimization strategy can adaptively adjust the search direction, automatically focus on the most potential area in the weight space, and make the final optimal function combination weight more suitable for the characteristics of the specific medical image data set, thereby improving the relevance and accuracy of the model.

[0029] Further, the specific prediction logic of the MedMamba model is as follows: The patch embedding layer receives a medical original image with a size of The medical original image is divided into a feature map with a size of by linear transformation to map the channel dimension, and the output dimension is The SS-Conv-SSM block splits the input feature X into and , and and are processed in parallel by SSM-Branch and Conv-Branch respectively, and the output results corresponding to each are output. The output results of SSM-Branch and Conv-Branch are spliced, the channel order is shuffled, and further fusion is performed through element multiplication (Element-wise Product) and element addition (Element-wise Addition), and the feature after this stage processing is output. The patch merging layer performs path merging and down-sampling on the feature after this stage processing, and sequentially completes the execution of the next SS-Conv-SSM block. After completing the execution of all SS-Conv-SSM blocks, the feature output by the last SS-Conv-SSM block is input into the feature classifier, the feature is mapped to the class probability distribution, and the prediction probability of each class is output.

[0030] Beneficial effects: The patch embedding layer divides the medical original image (HxWx3) into sub-blocks and maps them to the channel dimension through linear transformation, outputting a feature map with a size of H / 4xW / 4xC, which realizes efficient conversion of the image from spatial dimension to feature dimension. This processing not only preserves the key spatial information of the original image, but also reduces the complexity of subsequent calculations through down-sampling (size reduced by 1 / 4), while unifying the channel dimension of the feature, providing more adaptive input data for the subsequent parallel processing module.

[0031] The SS-Conv-SSM block splits the input feature into and The global and local parallel design through the SSM-Branch (state space model-based branch) and the Conv-Branch (convolution branch) in parallel enables the model to simultaneously consider the feature information of different scales in the medical image, avoids the omission of key features by a single branch, and improves the comprehensiveness of feature learning.

[0032] The output results of the parallel branches are fused through splicing, shuffle operation (disordering the channel order to promote cross-branch information interaction), element multiplication and element addition: the splicing operation retains the complete features of the two branches; the shuffle operation breaks the fixed association between channels, promotes the mixing and complementation of different branch features; the element multiplication and element addition further strengthen effective features and suppress redundant information through nonlinear interaction. The features are path-merged and down-sampled through the patch merging layer, and then a plurality of SS-Conv-SSM blocks are executed in sequence, forming a gradual processing procedure of "feature extraction-fusion-down-sampling-re-extraction". Finally, the features are mapped to a category probability distribution by the classifier, and the prediction probability of each category is directly output, which meets the demand for "probabilistic results" in medical diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 FIG. 1 is a flowchart of a medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration in an embodiment of the present application; Figure 2 FIG. 2 is a whole architecture diagram corresponding to the hybrid quantum ensemble learning and MedMamba collaboration in the embodiment of the present application; Figure 3 FIG. 3 is an implementation flowchart of a multi-core quantum support vector machine in the embodiment of the present application; Figure 4 FIG. 4 is a complete quantum circuit diagram of a QCNN in the embodiment of the present application; Figure 5 FIG. 5 is a whole architecture diagram of a MedMamba model in the embodiment of the present application; Figure 6 FIG. 6 is a comparison diagram between the hybrid quantum ensemble learning model and other models in the embodiment of the present application. DETAILED DESCRIPTION

[0034] The following will be further described in detail through specific embodiments: Embodiment One A medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration, substantially as shown in Figure 1 and Figure 2 , comprises the following steps: S1, input the medical original image into a preset U-Net feature extractor, extract the explainable features related to the specific disease from the medical original image, and output the corresponding feature vector; The S1 includes the following steps: Obtain a medical original image, and identify a specific disease type corresponding to the medical original image according to the obtained medical original image; According to the identified specific disease type, the U-Net feature extractor corresponding to the specific disease type is called, and the training process of the U-Net feature extractor corresponding to different specific disease types is different; According to the U-Net feature extractor called, the medical original image is segmented into a region of interest, and at least one of shape features, texture features, and frequency domain features is extracted in the region of interest. The obtained explainable features are normalized and finally combined into a corresponding feature vector.

[0035] S2, input the output feature vector into a pre-trained hybrid quantum integrated learning model, and output a corresponding first prediction classification probability; the hybrid quantum integrated learning model takes a multi-kernel quantum support vector machine (MKQSVM) and a quantum convolutional neural network (QCNN) as a base learner, and takes a logistic regression as a meta-learner; The training of the pre-trained hybrid quantum integrated learning model in S2 includes the following steps: S200, obtain different types of medical original historical image data sets, segment the region of interest through the U-Net feature extractor, and extract at least one of shape features, texture features, and frequency domain features in the region of interest. The obtained explainable features are normalized and combined to form a corresponding historical feature vector set. The historical feature vector set is divided into a training set and a validation set. In this embodiment, the encoder part of the U-Net feature extractor or the feature map of the decoder specific layer is mainly used to derive the required explainable features. The extracted features have data set and task specificity. For example: For the dermoscopy image dataset HAM10000, we extract lesion-related features according to the ABCDPT rule (Asymmetry, Border irregularity, Color variegation, Diameter, Patterns, Texture) commonly used in clinical practice. First, we use the U-Net to segment the lesion area, and measure the lesion asymmetry (Asymmetry) by the principal axis symmetry index, and evaluate the border irregularity (Border irregularity) by the fractal dimension and the ratio of the square of the perimeter to the area. Color variegation (Color variegation) is represented by the color histogram and color standard deviation in the RGB / HSV space, and the lesion diameter (Diameter) is estimated by the minimum circumscribed circle. Image patterns (Patterns) are extracted from the U-Net feature maps, combined with edge and structure distribution for analysis; texture features (Texture) are described by the gray level co-occurrence matrix (GLCM), local binary pattern (LBP), and wavelet energy distribution of the lesion internal structure. This feature system achieves a good balance between medical interpretability and image classification effectiveness.

[0036] For the brain tumor MRI image dataset (Brain Tumor Classification (MRI)), we use the U-Net model to extract multi-level image features. Texture features are important for distinguishing different types of tumors, and we construct a variety of texture representations, including gray level co-occurrence matrix (GLCM) features (such as contrast, homogeneity, energy, and entropy), local binary pattern (LBP) histogram, Haralick texture features (13-dimensional descriptors), and Tamura texture indicators (roughness, contrast, and directionality), to fully characterize the structural details and directional characteristics of tumor tissue. For intensity distribution features, we calculate the average gray level, maximum value, minimum value, median, variance, skewness, and kurtosis of the lesion area, and analyze the brightness structure distribution of the tumor area through the gray level histogram, providing more rich gray level information for the model. At the same time, we extract frequency domain features, evaluate the proportion of high-frequency components in the image through Fourier transform, and further improve the model's perception of structural level changes by combining the multi-scale energy distribution after wavelet decomposition, and calculate the statistical values of the coefficients after discrete cosine transform (DCT), such as the mean and standard deviation of the first few main coefficients, as a compressed representation of local and global frequency patterns.

[0037] For the ECG Images Dataset of Cardiac Patients, we extract key waveform structural features such as P-wave width, QRS-wave width, T-wave amplitude, and waveform spacing to characterize the changes in the cardiac cycle. Morphological features include asymmetry, sharpness, and edge density to identify abnormal waveform shapes. In terms of texture features, we construct local gray-level consistency and LBP histograms to reflect the image texture microstructure. Frequency domain features are analyzed through wavelet transform to analyze multi-scale energy distribution and low-frequency energy density changes, and statistical values of DCT main coefficients are extracted to improve the model's perception of rhythm and spectral abnormalities.

[0038] The quantified interpretable features are normalized as necessary and finally combined into a d-dimensional feature vector X as the input of the subsequent HQEL model.

[0039] S300、According to the training set and the validation set, based on the preset multi-kernel learning optimization strategy and the plurality of selected quantum feature mapping functions, the quantum feature mapping function set corresponding to the multi-kernel quantum support vector machine (MKQSVM) and the optimal function combination weight corresponding to the quantum feature mapping function in the quantum feature mapping function set are determined, and the training of the multi-kernel quantum support vector machine (MKQSVM) is completed; in this embodiment, in order to enhance the expression ability and adaptability of the model, a multi-kernel method is adopted. This may involve combining multiple different quantum kernels. By trying different quantum feature mappings and corresponding quantum kernel functions, we can find the best combination for a specific dataset and task. For example, ZZFeatureMap may perform better for data with strong nonlinear features, while PauliFeatureMap may be more suitable for tasks that require capturing more complex interactions. Therefore, we use Bayesian optimization and apply multiple mapping methods corresponding to different quantum kernel functions to find the best weights, thereby realizing the multi-kernel quantum support vector machine under multiple mappings. In this embodiment, as shown in Figure 3 the implementation flowchart of the multi-kernel quantum support vector machine used in this embodiment is shown, which combines four quantum kernel functions (ZFeatureMap, ZZFeatureMap, BSPFeatureMap, and PauliFeatureMap) to obtain the optimal combination through the Bayesian optimization algorithm.

[0040] The preset multi-kernel learning optimization strategy includes the following steps: S301, a plurality of quantum feature mapping functions are used to form a corresponding quantum feature mapping function set; S302, a plurality of function weight combinations are randomly generated as initial search points, and n is the total number of quantum feature mapping functions in the quantum feature mapping function set. S303, mapping the historical feature vectors in the training set into the quantum state space according to the function weight combination of each group, different quantum feature mapping functions capturing different quantum characteristics of the historical feature vectors; for each quantum feature mapping, calculating the quantum kernel matrix corresponding to each function weight combination and the quantum combination kernel matrix ; In this embodiment, the calculation formula of the quantum kernel matrix and the quantum combination kernel matrix is:

[0041]

[0042] In the formula, is the quantum state obtained by transforming the jth classical sample feature through the ith quantum feature mapping , is the weight coefficient of the ith quantum kernel matrix, is the ith quantum kernel matrix calculated by the ith quantum feature mapping. S304, training the quantum support vector separator using the corresponding quantum combination kernel matrix, and performing k-fold cross-validation on the validation set to calculate the cross-validation accuracy under each group of function weight combinations;

[0043] S305, based on the evaluated function weight combinations and the corresponding cross-validation accuracy, constructing a Gaussian process surrogate model, the Gaussian process surrogate model being used to predict the objective function value and its uncertainty of each point in the entire weight search space, and outputting the corresponding prediction result data; S306, according to the prediction result data, selecting the next group of function weight combinations to be evaluated based on the preset acquisition function data, and repeatedly executing S303 until the preset stopping condition is met, and selecting the function weight combination with the highest cross-validation accuracy from all the evaluated function weight combinations as the optimal function combination weight.

[0044] ​S400, according to the training set and the validation set, the quantum convolutional neural network (QCNN) is trained and after the training is completed, the trained multi-core quantum support vector machine (MKQSVM) and the quantum convolutional neural network (QCNN) are obtained, the prediction output on the validation set is combined into a new feature set, and the new feature set is learned by using a logistic regression as a meta-learner to obtain an HQEL model. In this embodiment, the cuTensorNet framework is used to accelerate high-dimensional tensor operations and optimize the efficiency of quantum kernel calculations. In this embodiment, the quantum convolutional neural network (QCNN) combines the local feature extraction and dimension reduction operation of the classical convolutional neural network with the entanglement and superposition characteristics of quantum computing. Through the quantum convolutional layer and the quantum pooling layer, the QCNN gradually extracts the local features of the quantum state and reduces the number of quantum bits, and the trainable quantum gate parameters help to optimize the model. The behavior of the quantum convolutional neural network (QCNN) is similar to that of the classical convolutional neural network (CCNN). We encode the given dataset into a quantum circuit, and after encoding the dataset, we apply alternating convolutional layers and pooling layers. The complete quantum circuit diagram of the QCNN we use is as shown in Figure 4 In this embodiment, in order to evaluate the performance of the proposed hybrid quantum ensemble learning model, a series of experiments are designed to test the model's performance on multiple public datasets. As shown in Figure 6 The experimental results show that compared with a single quantum model, the hybrid quantum ensemble learning model has significantly improved accuracy, robustness and generalization ability. In addition, we compare the proposed hybrid quantum ensemble learning model with traditional machine learning models (such as classical SVM, Random Forest, etc.). The classical ensemble learning takes classical SVM, Random Forest and Gradient Boosting Decision Tree as base learners, and logistic regression as a meta-learner. Tables 1, 2 and 3 show the accuracy, recall rate, F1-score and other performance indicators of different models on each dataset.

[0045]

[0046] Table 1 Accuracy of different models on each dataset

[0047] Table 2 Recall rate of different models on each dataset

[0048] Table 3 F1-score of different models on each dataset ​The main advantages of the proposed model include: 1. By integrating multiple quantum base learners, the model's adaptability to different datasets is effectively improved. 2. The introduction of meta-learners further optimizes the model's ability to handle noise and outliers. 3. Although quantum computing itself has certain computational complexity, through reasonable kernel function combinations and model optimization, the proposed hybrid model still maintains a high level of classification performance. Experimental results show that the hybrid quantum ensemble learning model based on MKQSVM and QCNN generally outperforms traditional machine learning models on all test datasets. This demonstrates the significant advantages of the proposed model in handling different types of datasets.

[0049] S3. Input the original medical image into the preset MedMamba model and output the corresponding second predicted classification probability; the MedMamba model includes a patch embedding layer, stacked SS-Conv-SSM blocks, a patch merging layer for downsampling, and a feature classifier; like Figure 5 As shown, the specific prediction logic of the MedMamba model is as follows: The patch embedding layer receives size The original medical image is divided into channels by a linear transformation, and the output dimension is... Feature mapping; The SS-Conv-SSM block will input feature X Split into and ,Will and Perform SSM-Branch and Conv-Branch parallel processing respectively, and output the corresponding output results for each. The outputs of SSM-Branch and Conv-Branch are concatenated, the channel order is shuffled by a shuffle operation, and then further fused by element-wise product and element-wise addition to output the features processed in this stage. The patch merging layer performs path merging and downsampling on the features after stage processing, and sequentially completes the execution of the next SS-Conv-SSM block; After all SS-Conv-SSM blocks have been executed, the features output by the last SS-Conv-SSM block are input into the feature classifier, which maps the features to the class probability distribution and outputs the predicted probability for each class.

[0050] S4, according to the output of the first predicted classification probability and the second predicted classification probability, based on the preset medical image classification probability fusion calculation formula, the final classification probability corresponding to the medical original image is calculated.

[0051] The preset medical image classification probability fusion calculation formula is:

[0052] In the formula, is the final classification probability of the medical original image, is the first predicted classification probability, is the second predicted classification probability, is the fusion weight value of the first predicted classification probability.

[0053] In this embodiment, the proposed hybrid quantum integrated learning-MedMamba collaborative framework is compared with traditional deep learning models (such as Xception, ResNet, etc.). Tables 4, 5, and 6 show the accuracy, recall rate, F1-score, and other performance indicators of different models on each dataset.

[0054] In this embodiment, each dataset is HAM10000 (skin image), Chest X-Ray (chest X-ray), FETAL_PLANES_DB (fetal image), ECG Images (electrocardiogram), and Dataset of Cardiac Patients (cardiac patient dataset). HAM10000: This dataset collects skin mirror images from different populations, which are obtained and stored in different ways. The final dataset consists of 10,015 skin mirror images. The cases include a representative collection of all important diagnostic categories in the field of pigmented lesions: actinic keratosis and intraepithelial carcinoma / Bowen disease (akiec), basal cell carcinoma (bcc), benign keratotic lesions (solar lentigines / Seborrheic keratosis and lichen planus-like keratoses, BKL), dermal fibroma (DF), melanoma (MEL), melanocytic nevi (NV), and vascular lesions (angiomas, angiokeratomas, pyogenic granulomas, and hemorrhages, vasc).

[0055] Chest X-Ray Images: This dataset has 5863 X-ray images and 2 categories (pneumonia / normal). Chest X-ray images are selected from a retrospective cohort of pediatric patients aged 1 to 5 years at the Guangzhou Women and Children's Medical Center. All chest X-ray imaging was performed as part of routine clinical care for patients.

[0056] FETAL_PLANES_DB

[50] : This dataset is a large collection of fetal screening ultrasound images routinely acquired by sonographers and ultrasound machines from two different hospitals. All images are manually labeled by expert maternal-fetal clinicians. Images are classified into 6 categories: four most widely used fetal anatomical planes (abdomen, brain, femur, and chest), the mother's cervix (widely used for preterm birth screening), and a general category that includes any other less common image planes.

[0057] ECG Images Dataset of Cardiac Patients: This dataset is a wide collection of electrocardiogram (ECG) images aimed at aiding research and progress in the field of cardiovascular medicine. The dataset provides rich data that can be used for various analyses, including developing diagnostic tools and studying different heart conditions. The dataset is divided into four main categories, each representing a different heart condition.

[0058]

[0059] Table 4 Accuracy of different models on each dataset

[0060] Table 5 Recall of different models on each dataset

[0061] Table 6 F1-score of different models on each dataset From the experimental results, it can be seen that the hybrid quantum ensemble learning-MedMamba collaborative framework generally outperforms traditional deep learning models on all test datasets. This demonstrates the proposed model's clear advantage in handling different types of datasets.

[0062] The above-mentioned are only embodiments of the present application, and the common knowledge of the specific structure and characteristics in the scheme is described too much, the ordinary skilled in the art knows all the ordinary technical knowledge in the field of the present application before the application date or the priority date, can know all the prior art in the field, and has the ability to apply the conventional experimental means before the date, the ordinary skilled in the art can perfect and implement the scheme under the enlightenment given by the present application combined with their own ability, some typical known structure or known method should not become the obstacle for the ordinary skilled in the art to implement the present application. It should be pointed out that for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can also be made, which should also be considered as the protection scope of the present application, which will not affect the effect and practicality of the patent. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A medical image classification method of hybrid quantum integrated learning and MedMamba collaboration, characterized in that: The method comprises the following steps: S1, inputting a medical original image into a preset U-Net feature extractor to extract an interpretable feature related to a specific disease from the medical original image, and outputting a corresponding feature vector; S2, inputting the output feature vector into a pre-trained hybrid quantum integrated learning model to output a corresponding first predicted classification probability; the hybrid quantum integrated learning model takes a multi-kernel quantum support vector machine (MKQSVM) and a quantum convolutional neural network (QCNN) as base learners, and takes a logistic regression as a meta-learner; S3, inputting the medical original image into a preset MedMamba model to output a corresponding second predicted classification probability; the MedMamba model comprises a patch embedding layer, a stacked SS-Conv-SSM block, a patch merging layer for down-sampling, and a feature classifier; S4, calculating a final classification probability corresponding to the medical original image based on a preset medical image classification probability fusion calculation formula according to the output first predicted classification probability and the second predicted classification probability.

2. The method of claim 1, wherein the method is characterized in that: The S1 comprises the following steps: Obtaining a medical original image, and identifying a specific disease type corresponding to the medical original image according to the obtained medical original image; According to the identified specific disease type, the U-Net feature extractor corresponding to the specific disease type is called, and the training processes of the U-Net feature extractors corresponding to different specific disease types are different; According to the called U-Net feature extractor, the medical original image is segmented into a region of interest, and at least one interpretable feature of shape feature, texture feature and frequency domain feature is extracted in the region of interest, the obtained interpretable feature is normalized, and finally combined into a corresponding feature vector.

3. The method of claim 2, wherein the method is characterized in that: The preset medical image classification probability fusion calculation formula is: In the formula, is the final classification probability of the medical original image, is the first predicted classification probability, is the second predicted classification probability, is the fusion weight value of the first predicted classification probability.

4. The method of claim 3, wherein the method is characterized in that: The training of the pre-trained hybrid quantum integrated learning model in the S2 comprises the following steps: S200, obtaining different types of medical original historical image data sets, segmenting a region of interest through a U-Net feature extractor, and extracting at least one interpretable feature of shape feature, texture feature and frequency domain feature in the region of interest, normalizing the obtained interpretable feature, and combining to form a corresponding historical feature vector set, dividing the historical feature vector set into a training set and a validation set; S300, according to the training set and the validation set, based on a preset multi-kernel learning optimization strategy and a plurality of selected quantum feature mapping functions, determining a quantum feature mapping function set corresponding to a multi-kernel quantum support vector machine (MKQSVM) and an optimal function combination weight corresponding to a quantum feature mapping function in the quantum feature mapping function set, and completing the training of the multi-kernel quantum support vector machine (MKQSVM); S400, training a quantum convolutional neural network (QCNN) according to the training set and the validation set, and after the training is completed, obtaining a multi-kernel quantum support vector machine (MKQSVM) and a quantum convolutional neural network (QCNN) trained, a prediction output on the validation set, and combining into a new feature set, learning the new feature set by taking a logistic regression as a meta-learner, and obtaining an HQEL model.

5. The method of claim 4, wherein the method is characterized in that: The preset multi-kernel learning optimization strategy comprises the following steps: S301, a plurality of quantum feature mapping functions are used to form a corresponding quantum feature mapping function set; S302、Randomly generate multiple sets of function weight combinations , as the initial search point, n is the total number of quantum feature mapping functions in the set of quantum feature mapping functions; S303、According to the function weight combination of each group, map the historical feature vector in the training set to the quantum state space, and different quantum feature mapping functions capture different quantum characteristics of the historical feature vector; for each quantum feature mapping, calculate the quantum kernel matrix corresponding to each function weight combination and the quantum combination kernel matrix ; S304, a quantum support vector separator is trained using the corresponding quantum combination kernel matrix, and k-fold cross-validation is performed on the validation set to calculate the cross-validation accuracy under each group of function weight combinations; S305, based on the evaluated function weight combinations and the corresponding cross-validation accuracy, a Gaussian process proxy model is constructed, the Gaussian proxy model is used to predict the target function value and its uncertainty of each point in the entire weight search space, and a corresponding prediction result data is output; S306, according to the prediction result data, based on the preset acquisition function data, the next group of function weight combinations to be evaluated is selected, and S303 is repeatedly executed until the preset stopping condition is met, and the function weight combination with the maximum cross-validation accuracy is selected from all the evaluated function weight combinations as the optimal function combination weight.

6. The method of claim 5, wherein the method is characterized in that: The specific prediction logic of the MedMamba model is: The patch embedding layer receives a size is a medical raw image, and the medical raw image is divided into a channel dimension mapped by a linear transformation, and the output dimension is characteristic mapping The SS-Conv-SSM block processes the input feature X is split into and , respectively, and and are processed in parallel by the SSM-Branch and the Conv-Branch, respectively, and output the respective corresponding output results. The output results of the SSM-Branch and the Conv-Branch are spliced, the channel order is shuffled, and further fusion is performed through element multiplication (Element-wise Product) and element addition (Element-wise Addition), and the features processed in this stage are output; The patch merging layer performs path merging and down-sampling on the features processed in the stage, and sequentially completes the execution of the next SS-Conv-SSM block; After the execution of all SS-Conv-SSM blocks is completed, the features output by the last SS-Conv-SSM block are input into the feature classifier, the features are mapped to the category probability distribution, and the prediction probability of each category is output.

Images