A medical image classification method combining quantum integrated learning and MedMamba
By combining a hybrid quantum ensemble learning approach with MedMamba, and integrating multi-core quantum support vector machines and quantum convolutional neural networks, the problem of low accuracy and poor reliability of single models in medical image classification is solved, achieving efficient classification of complex medical images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING NORMAL UNIVERSITY
- Filing Date
- 2025-07-31
- Publication Date
- 2026-06-02
AI Technical Summary
In existing medical image classification technologies, single models have limited ability to capture complex and subtle lesion features, high computational costs, and insufficient sensitivity to noise, resulting in poor classification accuracy and reliability.
We employ a hybrid quantum ensemble learning approach combined with MedMamba. We extract interpretable features using the U-Net feature extractor, combine multi-core quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) as base learners, utilize logistic regression for meta-learning, and combine the patch embedding layer and SS-Conv-SSM block of the MedMamba model for feature processing. Finally, we calculate the final classification probability through probability fusion.
It improves the accuracy and reliability of medical image classification, especially its adaptability and generalization ability to complex medical images, and reduces the limitations and risk of misjudgment of a single model.
Smart Images
Figure CN120912982B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image classification technology, specifically to a medical image classification method that combines hybrid quantum ensemble learning with MedMamba collaboration. Background Technology
[0002] Medical image classification is a key technology in modern medical diagnosis and decision support. It uses computer algorithms to automatically analyze medical images (such as X-rays, CT scans, MRIs, and pathological slides) to achieve early disease screening, lesion localization, 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 Convolutional Neural Networks (CNNs) and Transformers are 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, existing medical image classification technologies still have the following shortcomings:
[0004] Traditional deep learning models (such as CNNs) rely on local feature extraction, which limits their ability to capture complex and subtle lesion features (such as early tumors and small lesions) in medical images. While pure Transformer-type models can model global dependencies, their computational cost is too high when processing high-resolution medical images, and they lack sensitivity to local details. In addition, medical images often have large modal differences and strong noise interference, making it difficult for a single model to stably extract discriminative and interpretable features, thus limiting classification accuracy.
[0005] Medical diagnosis requires extremely high reliability of classification results. The prediction results of a single model may be affected by noise and feature bias, resulting in misjudgment.
[0006] In summary, existing medical image classification methods suffer from low accuracy and poor reliability due to the reliance on single models. Summary of the Invention
[0007] One of the objectives of this invention is to provide a medical image classification method that combines quantum ensemble learning with MedMamba collaboration, thereby addressing the problems of low accuracy and poor reliability of single-model classification in existing technologies.
[0008] To achieve the above objectives, a medical image classification method combining hybrid quantum ensemble learning and MedMamba collaboration is provided, comprising the following steps:
[0009] S1. Input the original medical image into the preset U-Net feature extractor, extract interpretable features related to a specific disease from the original medical image, and output the corresponding feature vector;
[0010] S2. Input the output feature vector into the pre-trained hybrid quantum ensemble learning model and output the corresponding first predicted classification probability; the hybrid quantum ensemble learning model uses multi-core quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) as base learners and logistic regression as meta learner;
[0011] 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;
[0012] S4. Based on the output first and second predicted classification probabilities, and using the preset medical image classification probability fusion calculation formula, calculate the final classification probability corresponding to the original medical image.
[0013] The technical principle and effects of this scheme: In this scheme, U-Net feature extraction is used to extract interpretable features related to a specific disease from raw medical images. Through its symmetrical network structure, U-Net can effectively capture the contextual and detailed information of the image, and the output feature vector contains key disease-related features in the image, providing a foundation for subsequent classification.
[0014] This study utilizes the properties of quantum computing by employing Multi-Kernel Quantum Support Vector Machine (MKQSVM) and Quantum Convolutional Neural Network (QCNN) as base learners. MKQSVM combines quantum mechanics principles with support vector machines, potentially offering advantages in handling high-dimensional data and complex classification problems; QCNN introduces quantum convolution operations into traditional convolutional neural networks, promising more efficient feature extraction. Logistic regression serves as a meta-learner, integrating the outputs of the base learners. By learning the relationship between the predictions of the base learners and the true labels, it outputs the first predicted classification probability.
[0015] The MedMamba model includes patch embedding layers to segment the input image into patches and map them to specific dimensions, preserving the image's 2D structure. Stacked SS-Conv-SSM blocks are the core component; they split the feature map into two groups, extracting local and global information through Conv-Branch and SSM-Branch respectively, and then restoring channel dimensions and avoiding information loss through channel concatenation and channel shuffling. Patch merging layers are used for downsampling to reduce data dimensionality, and finally, the feature classifier outputs a second predicted classification probability based on the extracted features.
[0016] The first predicted classification probability output by the hybrid quantum ensemble learning model and the second predicted classification probability output by the MedMamba model are fused using a pre-defined formula for calculating medical image classification probabilities. This formula may be based on weighted averaging or other more complex fusion strategies, combining the strengths of the two models to obtain the final classification probability, thereby determining the category to which the medical image belongs.
[0017] This approach extracts effective features using U-Net and combines a hybrid quantum ensemble learning model and a MedMamba model to process and classify these features from different perspectives. Finally, probabilities are fused, fully leveraging the advantages of each model and promising higher accuracy in medical image classification tasks, especially for complex medical images containing multiple lesions or diverse imaging modalities. By combining the hybrid quantum ensemble learning path based on U-Net interpretable features with the direct classification path based on the advanced state-space model (MedMamba), and achieving complementary advantages through weighted fusion, this approach overcomes the limitations of single methods. The hybrid quantum ensemble learning model and the MedMamba model each have different feature extraction and classification mechanisms; their fusion reduces the limitations of a single model, making the overall model more adaptable to different types of medical images and improving its generalization ability. It performs well in different datasets and real-world application scenarios, solving the problems of low accuracy and poor reliability of single-model classification in existing technologies.
[0018] This approach integrates clinically meaningful, interpretable feature extraction (reflecting domain knowledge) with the potential of QML and MedMamba's efficient global modeling capabilities into a unified framework, aiming to achieve a dual improvement in performance and interpretability.
[0019] Furthermore, S1 includes the following steps:
[0020] Acquire raw medical images, and identify the specific disease type corresponding to the acquired raw medical images;
[0021] Based on the identified specific disease type, the U-Net feature extractor corresponding to that specific disease type is retrieved. The training process of the U-Net feature extractor corresponding to different specific disease types is different.
[0022] Based on the retrieved U-Net feature extractor, the original medical image is segmented into regions of interest, and at least one interpretable feature among shape features, texture features, and frequency domain features is extracted from the regions of interest. The obtained interpretable features are normalized and finally combined into the corresponding feature vector.
[0023] Beneficial effects: By first identifying the specific disease type corresponding to the original medical image, and then retrieving the disease-specific U-Net feature extractor, the unique characteristics of different diseases in image representation can be fully utilized. Since the training process of the U-Net feature extractor differs for different diseases (optimized for the image characteristics of each disease), key image information related to the disease can be captured more accurately, avoiding omissions or misjudgments of specific disease features by general models, thus laying a more reliable foundation for subsequent analysis.
[0024] Normalizing interpretable features eliminates dimensional differences between features, preventing any single feature from dominating the model analysis due to its excessively large numerical range, and ensuring that all features have equal weight and influence in subsequent modeling. Combining the processed features into feature vectors facilitates their use as input data for subsequent tasks such as disease diagnosis, staging, or prognosis prediction, thereby improving the overall performance and stability of the model.
[0025] Furthermore, the preset formula for calculating the probability fusion of medical image classification is as follows:
[0026]
[0027] In the formula, The final classification probability of the original medical image. The first predicted classification probability, For the second predicted classification probability, This is the fusion weight value for the first predicted classification probability.
[0028] Beneficial effect: This formula integrates the first predicted classification probabilities. Second predicted classification probability This approach combines the strengths of two models. Different models may be based on different algorithmic principles, data focuses, or feature learning methods, and each has its own strengths in classifying medical images (for example, one model may be better at capturing detailed texture features, while another is better at global structure analysis). The final classification probability after fusion... It can compensate for the limitations of a single model, reduce misjudgments caused by inherent model biases, and thus improve the overall classification accuracy.
[0029] Furthermore, the training of the pre-trained hybrid quantum ensemble learning model in S2 includes the following steps:
[0030] S200. Obtain different types of original historical medical image datasets, segment the region of interest using the U-Net feature extractor, and extract at least one interpretable feature from the shape feature, texture feature, and frequency domain feature within the region of interest. Normalize the obtained interpretable features and combine them to form the corresponding historical feature vector set. Divide the historical feature vector set into a training set and a validation set.
[0031] S300. Based on the training set and validation set, and using a preset multi-core learning optimization strategy and a variety of quantum feature mapping functions to be selected, determine the set of quantum feature mapping functions corresponding to the multi-core quantum support vector machine (MKQSVM) and the optimal function combination weights corresponding to the quantum feature mapping functions in the set of quantum feature mapping functions, and complete the training of the multi-core quantum support vector machine (MKQSVM).
[0032] S400. Based on the training set and validation set, train the quantum convolutional neural network (QCNN). After training, obtain the predicted outputs of the trained multi-kernel quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) on the validation set and combine them into a new feature set. Use logistic regression as a meta-learner to learn the new feature set to obtain the HQEL model.
[0033] Beneficial effects: By employing a pre-defined multi-core learning optimization strategy, the optimal combination is selected from a variety of candidate quantum feature mapping functions. This allows for more efficient nonlinear transformations of features using quantum mechanical properties (such as superposition and entanglement), uncovering complex patterns hidden in the data and improving the accuracy of classification boundaries. The weighted combination of different quantum feature mapping functions can take into account the advantages of various feature transformation methods, adapting to the diversity of medical image data.
[0034] The predicted outputs of the trained MKQSVM and QCNN on the validation set are combined to form a new feature set, which is then further learned through a logistic regression meta-learner to form the HQEL model. This ensemble learning approach combines the advantages of two different models: MKQSVM excels at handling small sample sizes and high-dimensional data, while QCNN performs exceptionally well in extracting deep image features. By fusing the prediction results of both models, the risk of overfitting from a single model can be reduced, the model's adaptability to different medical image scenarios can be enhanced, and its generalization ability and prediction stability can be improved.
[0035] Furthermore, the preset multi-core learning optimization strategy includes the following steps:
[0036] S301. Employ multiple quantum feature mapping functions to form a corresponding set of quantum feature mapping functions;
[0037] S302, Randomly generate multiple sets of function weight combinations , where n is the initial search point and n is the total number of quantum feature mapping functions in the set of quantum feature mapping functions;
[0038] S303. Based on the function weight combination of each group, map the historical feature vectors in the training set to the quantum state space. Different quantum feature mapping functions capture the different quantum properties of the historical feature vectors. For each quantum feature mapping, calculate the quantum kernel matrix corresponding to each function weight combination. and quantum combination nuclear matrix ;
[0039] S304. Train the quantum support vector separator using the corresponding quantum combination kernel matrix, and perform k-fold cross-validation on the validation set to calculate the cross-validation accuracy under each set of function weight combinations.
[0040] S305. Based on the evaluated function weight combination and the corresponding cross-validation accuracy, construct a Gaussian process surrogate model. The Gaussian surrogate model is used to predict the objective function value and its uncertainty at each point in the entire weight search space, and output the corresponding prediction result data.
[0041] S306. Based on the prediction results data and the preset data of the collected functions, select the next set of function weight combinations to be evaluated, repeat S303 in a loop until the preset stopping condition is met, and select the function weight combination with the highest cross-validation accuracy from all the function weight combinations that have been evaluated as the optimal function combination weight.
[0042] Beneficial effects: By employing multiple quantum feature mapping functions to form a function set, it is possible to capture the quantum properties of historical feature vectors from different perspectives and uncover richer intrinsic structures and patterns in the data.
[0043] Multiple initial weight combinations are randomly generated as the starting point for the search, and iterative optimization is performed using a Gaussian process surrogate model, 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 information from already evaluated points, guiding the intelligent selection of the next set of weight combinations to be evaluated. This significantly reduces computational overhead and improves optimization efficiency while ensuring search quality.
[0044] Performing k-fold cross-validation on the validation set allows for a more comprehensive and reliable evaluation of the performance of each weight combination. By repeatedly splitting the validation set and calculating the average accuracy, evaluation bias caused by the randomness of data partitioning is reduced, ensuring that the selected optimal weight combination not only performs well on the training data but also has strong generalization ability, enabling it to adapt to unseen medical image data and reducing the risk of overfitting.
[0045] Gaussian process surrogate models not only predict the objective function value (cross-validation accuracy) but also provide information on the uncertainty of the prediction. This uncertainty quantification helps to balance "exploration" (selecting weight combinations in known high-performance regions) and "exploration" (trying unknown regions to discover potentially better solutions) during the weight search process, making the optimization process more robust, avoiding getting trapped in local optima, and thus increasing the probability of finding the globally optimal weight combination.
[0046] The entire optimization process is based on feedback from real data (cross-validation accuracy), without relying on human experience or preset fixed parameters. By continuously iterating and updating the Gaussian process surrogate model and selecting evaluation points, the optimization strategy can adaptively adjust the search direction, automatically focusing on the most promising region in the weight space. This makes the final optimal function combination weights more closely match the characteristics of the specific medical image dataset, improving the model's specificity and accuracy.
[0047] Furthermore, the specific prediction logic of the MedMamba model is as follows:
[0048] 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;
[0049] 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.
[0050] 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.
[0051] 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;
[0052] 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.
[0053] Beneficial effects: The patch embedding layer achieves efficient image transformation from spatial to feature dimensions by dividing the original medical image (H×W×3) into sub-blocks and mapping them to the channel dimension through linear transformation, outputting a feature map of size H / 4×W / 4×C. This processing preserves the key spatial information of the original image, reduces the complexity of subsequent calculations through downsampling (reducing the size to 1 / 4), and unifies the channel dimensions of the features, providing more adaptable input data for subsequent parallel processing modules.
[0054] The SS-Conv-SSM block splits the input features into and The model processes features in parallel using SSM-Branch (a branch based on a state-space model) and Conv-Branch (a convolutional branch). This global + local parallel design allows the model to simultaneously take into account feature information at different scales in medical images, avoiding the omission of key features by a single branch and improving the comprehensiveness of feature learning.
[0055] The outputs of the parallel branches are fused through concatenation, shuffle operations (disrupting channel order to facilitate cross-branch information interaction), element-wise multiplication, and element-wise addition: concatenation preserves the complete features of both branches; shuffle operations break the inherent correlation between channels, promoting the mixing and complementarity of features from different branches; element-wise multiplication and addition further enhance effective features and suppress redundant information through non-linear interactions. Features are path-merged and downsampled through a patch merging layer, and then multiple SS-Conv-SSM blocks are executed sequentially, forming a progressive processing flow of "feature extraction-fusion-downsampling-re-extraction". Finally, the features are mapped to a class probability distribution by a classifier, directly outputting the predicted probability of each class, meeting the need for "probabilistic results" in medical diagnosis. Attached Figure Description
[0056] Figure 1 This is a flowchart of the medical image classification method combining hybrid quantum integrated learning and MedMamba in Embodiment 1 of the present invention;
[0057] Figure 2 This is a diagram showing the overall architecture of hybrid quantum integrated learning and MedMamba collaboration in Embodiment 1 of the present invention.
[0058] Figure 3 This is a flowchart illustrating the implementation of the multi-core quantum support vector machine in Embodiment 1 of the present invention.
[0059] Figure 4 This is the complete quantum circuit diagram of QCNN in Embodiment 1 of the present invention;
[0060] Figure 5This is an overall architecture diagram of the MedMamba model in Embodiment 1 of the present invention;
[0061] Figure 6 This is a comparison diagram between the hybrid quantum integrated learning model in Embodiment 1 of the present invention and other models. Detailed Implementation
[0062] The following detailed description illustrates the specific implementation method:
[0063] Example 1
[0064] A medical image classification method combining hybrid quantum ensemble learning and MedMamba collaboration, basically as follows: Figure 1 and Figure 2 As shown, it includes the following steps:
[0065] S1. Input the original medical image into the preset U-Net feature extractor, extract interpretable features related to a specific disease from the original medical image, and output the corresponding feature vector;
[0066] S1 includes the following steps:
[0067] Acquire raw medical images, and identify the specific disease type corresponding to the acquired raw medical images;
[0068] Based on the identified specific disease type, the U-Net feature extractor corresponding to that specific disease type is retrieved. The training process of the U-Net feature extractor corresponding to different specific disease types is different.
[0069] Based on the retrieved U-Net feature extractor, the original medical image is segmented into regions of interest, and at least one interpretable feature among shape features, texture features, and frequency domain features is extracted from the regions of interest. The obtained interpretable features are normalized and finally combined into the corresponding feature vector.
[0070] S2. Input the output feature vector into the pre-trained hybrid quantum ensemble learning model and output the corresponding first predicted classification probability; the hybrid quantum ensemble learning model uses multi-core quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) as base learners and logistic regression as meta learner;
[0071] The training of the pre-trained hybrid quantum ensemble learning model in S2 includes the following steps:
[0072] S200. Obtain different types of original historical medical image datasets. Segment the region of interest using the U-Net feature extractor, and extract at least one interpretable feature from the shape, texture, and frequency domain features within the region of interest. Normalize the obtained interpretable features and combine them to form a corresponding historical feature vector set. Divide the historical feature vector set into a training set and a validation set. In this embodiment, the required interpretable features are mainly derived using the encoder part of the U-Net feature extractor or by combining feature maps from specific layers of the decoder. The extracted features are dataset- and task-specific. For example:
[0073] For the dermoscopy image dataset HAM10000, we extracted lesion-related features based on the clinically commonly used ABCDPT rules (Asymmetry, Borderirregularity, Color variegation, Diameter, Patterns, Texture). First, we segmented the lesion region using U-Net, measured lesion asymmetry using the principal axis symmetry index, and assessed borderirregularity using fractal dimension and the ratio of perimeter squared to area. Color variegation was represented by color histograms and color standard deviations in RGB / HSV space, and lesion diameter was estimated using the minimum circumcircle. Image patterns were extracted from U-Net feature maps and analyzed in conjunction with edge and structure distributions; texture features were described by combining the gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), and wavelet energy distribution to depict the internal structure of the lesion. This feature system achieves a good balance between medical interpretability and image classification effectiveness.
[0074] For the Brain Tumor Classification (MRI) image dataset, we employed the U-Net model to extract multi-level image features. Texture features are crucial for differentiating different tumor types. We constructed various texture representations, including Gray-Level Co-occurrence Matrix (GLCM) features (such as contrast, homogeneity, energy, and entropy), Local Binary Pattern (LBP) histograms, Haralick texture features (13-dimensional descriptors), and Tamura texture indices (roughness, contrast, and directionality), to comprehensively depict the structural details and directional characteristics of tumor tissue. For intensity distribution features, we statistically analyzed the mean gray level, maximum, minimum, median, variance, skewness, and kurtosis of the lesion region, and used gray-level histograms to analyze the brightness structure distribution of the tumor region, providing the model with richer gray-level information. Meanwhile, we extract frequency domain features, evaluate the proportion of high-frequency components in the image through Fourier transform, and further enhance the model's ability to perceive structural hierarchical changes by combining the multi-scale energy distribution after wavelet decomposition. We also calculate the coefficient statistics after discrete cosine transform (DCT), such as the mean and standard deviation of the first few principal coefficients, as a compressed expression of the local and global spectral patterns.
[0075] For the ECG Images Dataset of Cardiac Patients, we extracted key waveform structural features (such as P wave width, QRS width, T wave amplitude, and waveform spacing) to characterize ECG cycle changes. Morphological features, including asymmetry, sharpness, and edge density, were used to identify abnormal waveform shapes. For texture features, we constructed local gray-level consistency and LBP histograms to reflect the microstructure of image texture. Frequency domain features included wavelet transform analysis of multi-scale energy distribution and low-frequency energy density changes, and the statistical values of the DCT principal coefficients were extracted to improve the model's ability to perceive rhythm and spectral anomalies.
[0076] The interpretable features obtained from the quantization are normalized as necessary and finally combined into a d-dimensional feature vector X as the input to the subsequent HQEL model.
[0077] S300. Based on the training and validation sets, and using a pre-defined multi-kernel learning optimization strategy and various quantum feature mapping functions to be selected, determine the set of quantum feature mapping functions corresponding to the multi-kernel quantum support vector machine (MKQSVM) and the optimal function combination weights corresponding to the quantum feature mapping functions in the set, thus completing the training of the multi-kernel quantum support vector machine (MKQSVM). In this embodiment, to enhance the model's expressive power and adaptability, we adopt a multi-kernel method. This may involve combining multiple different quantum kernels. By trying different quantum feature mappings and corresponding quantum kernel functions, we can find the combination most suitable 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 need to capture more complex interactions. Therefore, we use Bayesian optimization and apply multiple mapping methods corresponding to different quantum kernel functions to find the optimal weights, thereby realizing a multi-kernel quantum support vector machine under multiple mappings. In this embodiment, as shown... Figure 3 The diagram shows the implementation flowchart of the multi-core quantum support vector machine in this embodiment, which combines four quantum kernel functions (ZFeatureMap, ZZFeatureMap, BSPFeatureMap and PauliFeatureMap) and obtains the optimal combination through a Bayesian optimization algorithm.
[0078] The preset multi-core learning optimization strategy includes the following steps:
[0079] S301. Employ multiple quantum feature mapping functions to form a corresponding set of quantum feature mapping functions;
[0080] S302, Randomly generate multiple sets of function weight combinations , where n is the initial search point and n is the total number of quantum feature mapping functions in the set of quantum feature mapping functions;
[0081] S303. Based on the function weight combination of each group, map the historical feature vectors in the training set to the quantum state space. Different quantum feature mapping functions capture the different quantum properties of the historical feature vectors. For each quantum feature mapping, calculate the quantum kernel matrix corresponding to each function weight combination. and quantum combination nuclear matrix ;
[0082] In this embodiment, the quantum core matrix and quantum combination nuclear matrix The calculation formula is:
[0083]
[0084]
[0085] In the formula, To incorporate the features of the j-th classic sample Through the mapping of the i-th quantum feature The quantum state obtained after the transformation The weight coefficients of the i-th quantum kernel matrix are... The i-th quantum core matrix is calculated from the i-th quantum feature mapping.
[0086] S304. Train the quantum support vector separator using the corresponding quantum combination kernel matrix, and perform k-fold cross-validation on the validation set to calculate the cross-validation accuracy under each set of function weight combinations.
[0087] S305. Based on the evaluated function weight combination and the corresponding cross-validation accuracy, construct a Gaussian process surrogate model. The Gaussian surrogate model is used to predict the objective function value and its uncertainty at each point in the entire weight search space, and output the corresponding prediction result data.
[0088] S306. Based on the prediction results data and the preset data of the collected functions, select the next set of function weight combinations to be evaluated, repeat S303 in a loop until the preset stopping condition is met, and select the function weight combination with the highest cross-validation accuracy from all the function weight combinations that have been evaluated as the optimal function combination weight.
[0089] S400. Based on the training and validation sets, the quantum convolutional neural network (QCNN) is trained. After training, the predicted outputs of the trained multi-kernel quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) on the validation set are obtained and combined into a new feature set. Logistic regression is used as a meta-learner to learn the new feature set, resulting in the HQEL model. In this embodiment, the cuTensorNet framework is used to accelerate high-dimensional tensor operations and optimize the computational efficiency of quantum kernels. In this embodiment, the quantum convolutional neural network (QCNN) combines the local feature extraction and dimensionality reduction operations of classical convolutional neural networks with the entanglement and superposition characteristics of quantum computing. Through quantum convolutional layers and quantum pooling layers, QCNN gradually extracts local features of quantum states and reduces the number of qubits. Trainable quantum gate parameters help 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 and pooling layers. The complete quantum circuit diagram of the QCNN we use is as follows: Figure 4 As shown
[0090] In this embodiment, to evaluate the performance of the proposed hybrid quantum ensemble learning model, we designed a series of experiments to test the model's performance on multiple public datasets. For example... Figure 6 As shown, we selected three datasets for comparison. Experimental results demonstrate that the hybrid quantum ensemble learning model significantly improves accuracy, robustness, and generalization ability compared to a single quantum model. Furthermore, we compared the proposed hybrid quantum ensemble learning model with traditional machine learning models (such as classical SVM and Random Forest). Classical ensemble learning uses classical SVM, Random Forest, and Gradient Boosting Decision Tree as base learners, and logistic regression as the meta-learner. Tables 1, 2, and 3 show the accuracy, recall, F1-score, and other performance metrics of different models on each dataset.
[0091]
[0092] Table 1. Accuracy of different models on various datasets
[0093]
[0094] Table 2. Recall rates of different models on various datasets
[0095]
[0096] Table 3. F1 scores of different models on various datasets
[0097] 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.
[0098] 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;
[0099] like Figure 5 As shown, the specific prediction logic of the MedMamba model is as follows:
[0100] 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;
[0101] 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.
[0102] 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.
[0103] 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;
[0104] 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.
[0105] S4. Based on the output first and second predicted classification probabilities, and using the preset medical image classification probability fusion calculation formula, calculate the final classification probability corresponding to the original medical image.
[0106] The preset formula for calculating the probability fusion of medical image classification is as follows:
[0107]
[0108] In the formula, The final classification probability of the original medical image. The first predicted classification probability, For the second predicted classification probability, This is the fusion weight value for the first predicted classification probability.
[0109] In this embodiment, the proposed hybrid quantum ensemble learning-MedMamba collaborative framework is compared with traditional deep learning models (such as Xception, ResNet, etc.). Tables 4, 5, and 6 show the performance metrics of different models on various datasets, including accuracy, recall, and F1-score.
[0110] In this implementation, the datasets are HAM10000 (skin images), Chest X-Ray, FETAL_PLANES_DB (fetal images), ECG Images, and DatasetofCardiac Patients. HAM10000: This dataset collects dermoscopic images from different populations, acquired and stored using various methods. The final dataset consists of 10,015 dermoscopic images. The cases include a representative set of all important diagnostic categories in the field of pigmented lesions: actinic keratosis and intraepithelial carcinoma / Bowen's disease (akiec), basal cell carcinoma (bcc), benign keratodermoid lesions (solar freckles / seborrheic keratosis and lichen planus-like keratosis, BKL), dermatofibroma (DF), melanoma (MEL), melanocytic nevus (NV), and vascular lesions (hemangioma, angiokeratoma, pyogenic granuloma, and hemorrhage, vasc).
[0111] Chest X-Ray Images: This dataset contains 5863 X-ray images and two categories (pneumonia / normal). Chest X-ray images were selected from a retrospective cohort of pediatric patients aged 1 to 5 years at Guangzhou Women and Children's Medical Center. All chest X-ray imaging was performed as part of the patients' routine clinical care.
[0112] FETAL_PLANES_DB
[50] : This dataset is a large dataset of routinely acquired maternal-fetal screening ultrasound images collected from two different hospitals by several staff members and ultrasound machines. All images were manually labeled by professional maternal-fetal clinicians. The images are divided into 6 categories: the four most widely used fetal anatomical planes (abdomen, brain, femur, and chest), the maternal cervix (widely used for preterm infant screening), and a general category that includes any other less common image planes.
[0113] ECG Images Dataset of Cardiac Patients: This dataset is a comprehensive collection of electrocardiogram (ECG) images designed to aid research and advancements in the field of cardiovascular medicine. It provides rich data for various analyses, including the development of diagnostic tools and research into different heart diseases. The dataset is divided into four main categories, each representing a different heart disease. Brain Tumor Classification (MRI): This dataset consists of images, as magnetic resonance imaging (MRI) is the best technique for detecting brain tumors. A large amount of image data is generated through scans. Radiologists examine these images. Brain tumors are categorized as benign, malignant, pituitary adenomas, etc. This dataset classifies images into four categories: benign, malignant, no tumor, and pituitary adenoma.
[0114]
[0115] Table 4. Accuracy of different models on various datasets
[0116]
[0117] Table 5. Recall rates of different models on various datasets
[0118]
[0119] Table 6. F1 scores of different models on various datasets.
[0120] The experimental results show that the hybrid quantum ensemble learning-MedMamba collaborative framework outperforms traditional deep learning models on almost all test datasets. This demonstrates the significant advantages of the proposed model in handling different types of datasets.
[0121] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, based on the guidance provided in this application, improve and implement this solution in combination with their own capabilities. Some typical well-known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A medical image classification method of hybrid quantum integrated learning and MedMamba collaboration, characterized in that: Includes the following steps: S1. Input the original medical image into the preset U-Net feature extractor, extract interpretable features related to a specific disease from the original medical image, and output the corresponding feature vector; S2. Input the output feature vector into the pre-trained hybrid quantum ensemble learning model and output the corresponding first predicted classification probability; the hybrid quantum ensemble learning model uses multi-core quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) as base learners and logistic regression as meta learner; 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; 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 of each class. S4. Based on the output first and second predicted classification probabilities, and using the preset medical image classification probability fusion calculation formula, calculate the final classification probability corresponding to the original medical image.
2. The medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration according to claim 1, characterized in that: S1 includes the following steps: Acquire raw medical images, and identify the specific disease type corresponding to the acquired raw medical images; Based on the identified specific disease type, the U-Net feature extractor corresponding to that specific disease type is retrieved. The training process of the U-Net feature extractor corresponding to different specific disease types is different. Based on the retrieved U-Net feature extractor, the original medical image is segmented into regions of interest, and at least one interpretable feature among shape features, texture features, and frequency domain features is extracted from the regions of interest. The obtained interpretable features are normalized and finally combined into the corresponding feature vector.
3. The medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration according to claim 2, characterized in that: The preset formula for calculating the probability fusion of medical image classification is as follows: In the formula, The final classification probability of the original medical image. The first predicted classification probability, For the second predicted classification probability, This is the fusion weight value for the first predicted classification probability.
4. The medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration according to claim 3, characterized in that: The training of the pre-trained hybrid quantum ensemble learning model in S2 includes the following steps: S200. Obtain different types of original historical medical image datasets, segment the region of interest using the U-Net feature extractor, and extract at least one interpretable feature from the shape feature, texture feature, and frequency domain feature within the region of interest. Normalize the obtained interpretable features and combine them to form the corresponding historical feature vector set. Divide the historical feature vector set into a training set and a validation set. S300. Based on the training set and validation set, and using a preset multi-core learning optimization strategy and a variety of quantum feature mapping functions to be selected, determine the set of quantum feature mapping functions corresponding to the multi-core quantum support vector machine (MKQSVM) and the optimal function combination weights corresponding to the quantum feature mapping functions in the set of quantum feature mapping functions, and complete the training of the multi-core quantum support vector machine (MKQSVM). S400. Based on the training set and validation set, train the quantum convolutional neural network (QCNN). After training, obtain the predicted outputs of the trained multi-kernel quantum support vector machine (MKQSVM) and quantum convolutional neural network (QCNN) on the validation set and combine them into a new feature set. Use logistic regression as a meta-learner to learn the new feature set to obtain the HQEL model.
5. A medical image classification method based on hybrid quantum ensemble learning and MedMamba collaboration according to claim 4, characterized in that: The preset multi-core learning optimization strategy includes the following steps: S301. Employ multiple quantum feature mapping functions to form a corresponding set of quantum feature mapping functions; S302, Randomly generate multiple sets of function weight combinations , where n is the initial search point and n is the total number of quantum feature mapping functions in the set of quantum feature mapping functions; S303. Based on the function weight combination of each group, map the historical feature vectors in the training set to the quantum state space. Different quantum feature mapping functions capture the different quantum properties of the historical feature vectors. For each quantum feature mapping, calculate the quantum kernel matrix corresponding to each function weight combination. and quantum combination nuclear matrix ; S304. Train the quantum support vector separator using the corresponding quantum combination kernel matrix, and perform k-fold cross-validation on the validation set to calculate the cross-validation accuracy under each set of function weight combinations. S305. Based on the evaluated function weight combination and the corresponding cross-validation accuracy, construct a Gaussian process surrogate model. The Gaussian surrogate model is used to predict the objective function value and its uncertainty at each point in the entire weight search space, and output the corresponding prediction result data. S306. Based on the prediction results data and the preset data of the collected functions, select the next set of function weight combinations to be evaluated, repeat S303 in a loop until the preset stopping condition is met, and select the function weight combination with the highest cross-validation accuracy from all the function weight combinations that have been evaluated as the optimal function combination weight.