Channel weighting and bayesian optimization for medical image processing method and system

Abstract

The application relates to the field of image processing and discloses a medical image processing method and system based on channel weighting and Bayesian optimization. An original high-resolution three-dimensional image is compressed into low-resolution compressed video information through a compression video information acquisition module, the calculation data amount of a first boundary extraction model is greatly reduced, and the extraction speed of a coarse boundary target region is significantly improved. A high-dimensional weighted feature matrix constructed by the application can retain the differentiated information value of each sequence in a clinical image, and the utilization rate of effective features is significantly improved. A Bayesian iteration framework is constructed based on a binary mask feature subset, a Gaussian process proxy model is used to fit the probability distribution between the feature subset and a target function, an expected gain acquisition function is combined to intelligently guide the search direction, a global optimal feature subset can be efficiently screened from thousands of features, the feature redundancy is low, the compactness is high, and the nonlinear synergistic effect between features can be captured.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a medical image processing method and system using channel weighting and Bayesian optimization. Background Technology

[0002] With the widespread application of multimodal medical imaging technology in the assessment of complex diseases, the amount of image data acquired by multiparameter magnetic resonance imaging (MRI) is growing exponentially. How to extract stable and interpretable key features from massive, high-dimensional raw images and construct accurate image processing and assessment models has become an important research direction in the field of medical image processing.

[0003] In existing technologies, common medical image processing methods typically involve preprocessing multiple MRI sequences, extracting radiomics features based on regions of interest (ROIs), and using machine learning models for feature selection and prediction. In the feature selection stage, the widely used method is LASSO regression, which treats all features extracted from the image sequences as equally important and performs dimensionality reduction based on the linear assumption.

[0004] The existing technology has the following problems:

[0005] 1. Relying entirely on the statistical distribution of data makes it impossible to distinguish the differentiated information value of different image sequences in clinical diagnosis, resulting in a large amount of computing power being consumed on redundant features, and it is impossible to incorporate prior weights at the sequence level in the early stages of screening.

[0006] 2. When faced with thousands of features, it tends to randomly retain one feature from highly correlated feature groups and discard the rest. This easily leads to the loss of feature groups that perform poorly individually but have strong representational capabilities under specific nonlinear combinations, thus reducing the accuracy of feature representation.

[0007] 3. Traditional methods such as grid search or random search have exponentially increasing computational costs in high-dimensional omics feature spaces, making it difficult to efficiently converge to the globally optimal feature combination. Summary of the Invention

[0008] The purpose of this invention is to solve the technical problems of existing MRI image processing technology, such as lack of prior knowledge guidance for feature selection, neglect of nonlinear synergistic effects between features in linear dimensionality reduction, and low efficiency of feature optimization in high-dimensional space. The invention provides a medical image processing method and system based on channel weighting and Bayesian optimization.

[0009] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0010] A channel-weighted and Bayesian optimization method for medical image processing includes the following sub-steps:

[0011] Bias field correction, anisotropic diffusion filtering, B-spline interpolation resampling, and normalization were performed on the original multi-sequence MRI images to obtain standardized multi-sequence image data.

[0012] Based on lesion masking and multi-sequence normalized image data, features are extracted to obtain an initial feature set. Unstable features in the initial feature set are then removed to obtain a stable feature set.

[0013] The GBDT model is used to process the stable feature set to obtain information gain. The information gain is aggregated to obtain the original representative weight of each sequence. The original representative weight is smoothed and normalized to obtain the prior weight. The prior weight and the stable feature set are used element by element to obtain the high-dimensional weighted feature matrix.

[0014] A feature subset is constructed based on the high-dimensional weighted feature matrix, and a Bayesian iteration is performed on the feature subset to obtain the globally optimal feature subset;

[0015] The candidate classifier set is processed through multi-fold cross-validation to process the globally optimal feature subset. The performance score of the candidate classifier set on the globally optimal feature subset is comprehensively evaluated, and the best classifier is selected to output the predicted probability of objective mitigation tendency.

[0016] The risk groups of case samples are divided by the predicted probability of objective mitigation tendency and the optimal cutoff threshold. The marginal contribution of each feature in the globally optimal feature subset is obtained by using the XGBoost classifier and then integrated to obtain the feature attribution map.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: By assigning prior weights to different image sequences and multiplying them element-wise with a stable feature set, the high-dimensional weighted feature matrix constructed by this invention can retain the differentiated information value of each sequence in clinical images, significantly improving the utilization rate of effective features; This invention constructs a Bayesian iterative framework based on binary mask feature subsets, uses a Gaussian process surrogate model to fit the probability distribution between the feature subset and the objective function, and combines the expected gain acquisition function to intelligently guide the search direction, which can efficiently screen out the globally optimal feature subset from thousands of features, with low feature redundancy, strong compactness, and the ability to capture the nonlinear synergistic effect between features; While outputting the objective mitigation tendency prediction probability, this invention automatically searches for the optimal truncation threshold to transform the probability value into a hierarchical risk group, and uses an XGBoost tree structure to generate an attribution map of the marginal contribution of each feature, showing the direction and strength of the effect of each weighted feature on the prediction result, providing a white-box interpretation basis, and significantly enhancing the transparency and credibility of image processing results.

[0018] Furthermore, a channel-weighted and Bayesian optimization method for medical image processing, wherein the process of performing bias field correction, anisotropic diffusion filtering, B-spline interpolation resampling, and normalization on multi-sequence raw MRI images to obtain multi-sequence standardized image data includes the following sub-steps:

[0019] The intensity of the original MRI image at the voxel location was corrected by bias field using the N4ITK method to obtain the intensity of the bias field corrected MRI image at the voxel location.

[0020] The intensity of the bias-field corrected MRI image at the voxel location is obtained by suppressing background noise through anisotropic diffusion filtering.

[0021] The intensity of the filtered MRI image at the voxel location is resampled to an isotropic voxel using B-spline interpolation to obtain the intensity of the spatially aligned MRI image at the target voxel location.

[0022] The intensity of the spatially aligned MRI image at the target voxel location was normalized using the Z-score method to obtain the intensity of the normalized MRI image at the target voxel location.

[0023] The intensity of each normalized MRI image sequence at the target voxel location is integrated to obtain multi-sequence normalized image data.

[0024] In the above scheme, by sequentially performing bias field correction, anisotropic diffusion filtering, B-spline spatial resampling, and Z-score gray-level normalization, the heterogeneity of equipment and scanning parameters among multi-sequence MRI images is effectively eliminated, and high-precision alignment and gray-level distribution standardization of each sequence are achieved in a unified physical space, providing a high-quality and consistent image data foundation for cross-sequence feature fusion.

[0025] Furthermore, a channel-weighted and Bayesian optimization method for medical image processing, wherein the extraction of features based on lesion masking and multi-sequence normalized image data to obtain an initial feature set, and the removal of unstable features from the initial feature set to obtain a stable feature set, includes the following sub-steps:

[0026] Through lesion masking First-order statistical features, shape features, texture features, and frequency domain features are extracted from the corresponding multi-sequence standardized image data and integrated to form an initial feature set;

[0027] The stability of the initial feature set is screened using the intragroup correlation coefficient, unstable features are removed, and a stable feature set is constructed.

[0028] By precisely defining the region of interest through lesion masking, multi-dimensional features such as first-order statistical features, shape features, texture features, and frequency domain features are extracted. Compared with single-type features, it can provide more comprehensive and complementary image information representation. Furthermore, unstable features are eliminated by using intragroup correlation coefficients, which significantly reduces the sensitivity of lesion segmentation boundaries and the interference of repeated measurement noise on feature quality, ensuring the high stability and high repeatability of the constructed initial feature set.

[0029] Furthermore, a medical image processing method based on channel weighting and Bayesian optimization, wherein the lesion masking... The process of extracting first-order statistical features, shape features, texture features, and frequency domain features from the corresponding multi-sequence standardized image data and integrating them to form an initial feature set includes the following sub-steps:

[0030] Through lesion masking For each sequence of standardized image data, the mean, variance, skewness, kurtosis, and entropy of each sequence of standardized image data are extracted and integrated to obtain first-order statistical features;

[0031] Through lesion masking The volume, surface area, sphericity, and compactness of the formed three-dimensional region are extracted and integrated to obtain shape features;

[0032] Based on lesion masking For each sequence of normalized image data, a gray-level co-occurrence matrix and a gray-level run-length matrix are constructed using the corresponding sequence of normalized image data. Contrast, energy, correlation, short run-length emphasis, and long run-length emphasis are obtained using the gray-level co-occurrence matrix and the gray-level run-length matrix, and texture features are obtained by integration.

[0033] Perform a 3D wavelet transform on each sequence of normalized image data to obtain multiple wavelet sub-band images. Then, perform a Laplace-Gaussian transform on each sequence of normalized image data to obtain LoG-filtered images at multiple scales. Finally, use lesion masking... The mean and variance of the wavelet subband image and the LoG filtered image are extracted and integrated to obtain the frequency domain features.

[0034] First-order statistical features, shape features, texture features, and frequency domain features are integrated to form an initial feature set.

[0035] In the above scheme, first-order statistical features quantify the overall attributes of grayscale distribution, shape features objectively describe three-dimensional geometric morphology, texture features capture pixel spatial dependencies and microscopic heterogeneity through co-occurrence matrices and run-length matrices, and frequency domain features extract multi-scale energy distribution using wavelet and LoG transforms. These four types of features complement each other from macroscopic to microscopic and from spatial to frequency domains. The integrated initial feature set can comprehensively characterize lesion heterogeneity, avoiding the limitations of single-feature descriptions and significantly improving the depth and expressive power of information mining for complex image patterns.

[0036] Furthermore, a channel-weighted and Bayesian optimization method for medical image processing, which involves processing a stable feature set using a GBDT model to obtain information gain, aggregating the information gain to obtain the original representative weights for each sequence, smoothing and normalizing the original representative weights to obtain prior weights, and using the prior weights and the stable feature set element-wise to obtain a high-dimensional weighted feature matrix, includes the following sub-steps:

[0037] The GBDT model processes the j-th feature of sequence i in the stable feature set and the clinical label corresponding to the feature to obtain the information gain of the j-th feature of sequence i in the stable feature set. The information gain is aggregated to obtain the original representative weight of sequence i.

[0038] The Softmax function with a temperature coefficient smoothly normalizes the original representative weights of sequence i to obtain the prior weights of sequence i.

[0039] The prior weights of each sequence and the stable feature set are multiplied element-wise to obtain a high-dimensional weighted feature matrix.

[0040] In the above scheme, the information gain of each feature is extracted by the GBDT model and aggregated into the original representative weight of each sequence, which can quantify the predictive contribution of different image sequences to clinical labels. The Softmax function with temperature coefficient is used for smooth normalization, which not only preserves the importance differences between sequences, but also prevents the absolute concentration of weight of a certain sequence from overwhelming the information of other sequences. The prior weights are multiplied element-wise with the stable feature set, so that each feature carries the differentiated weight of its sequence in numerical terms. This realizes the embedding of sequence-level prior knowledge for clinical data distribution and significantly improves the information fidelity of the high-dimensional weighted feature matrix.

[0041] Furthermore, a channel-weighted and Bayesian optimization method for medical image processing, wherein constructing a feature subset based on a high-dimensional weighted feature matrix and performing Bayesian iteration on the feature subset to obtain the globally optimal feature subset includes the following sub-steps:

[0042] The binary mask vector is defined by the dimension of the high-dimensional weighted feature matrix, and the feature subset is determined by the binary mask vector;

[0043] An XGBoost classifier is trained using feature subsets and their corresponding clinical labels. The trained XGBoost classifier is then subjected to repeated multi-fold cross-validation to obtain the objective function.

[0044] Assuming the objective function follows a Gaussian distribution, construct a Gaussian process surrogate model;

[0045] The expected gain is obtained based on the predicted mean and predicted standard deviation of the Gaussian process.

[0046] Repeatedly perform expected gain acquisition and objective function evaluation, update the Gaussian process surrogate model as a Bayesian iteration, and stop the iteration when the termination condition is met to obtain the globally optimal feature subset.

[0047] In the above scheme, the high-dimensional weighted feature matrix is ​​transformed into a searchable feature subset space through a binary mask vector. The cross-validation AUC of the XGBoost classifier is used as the objective function. A Gaussian process surrogate model is used to fit the probability distribution between the feature subset and the objective function. The search direction is intelligently guided based on the expected gain acquisition function, thus achieving global and efficient optimization in the high-dimensional space. It can quickly converge to the optimal feature combination with fewer iterations. At the same time, it avoids invalid computation by setting a preset termination condition. The final output global optimal feature subset has significant low redundancy and high representation ability.

[0048] Furthermore, a channel-weighted and Bayesian optimization method for medical image processing, wherein the termination condition is:

[0049] If the maximum number of Bayes iterations If the iteration stops, then stop.

[0050] If the improvement of the objective function within 100 consecutive Bayesian iterations If the iteration stops, then stop.

[0051] If the maximum expected gain is within 50 consecutive Bayesian iterations If the iteration stops, then stop.

[0052] If the Jaccard similarity coefficient between the optimal feature subsets during continuous Bayes iteration is... If the iteration stops, then stop.

[0053] in, The optimal feature subset for the current features. It is the optimal feature subset of historical features.

[0054] In the above scheme, by setting an upper limit on the maximum number of iterations, the waste of computational resources caused by infinite loops is avoided; by monitoring the improvement of the objective function, the system automatically terminates when the performance approaches saturation to prevent invalid iterations; by limiting the maximum expected gain, the system exits early when the acquisition function cannot find a more promising subset of candidate features; by calculating the Jaccard similarity coefficient between the optimal subsets of features, the system determines that the search has reached stable convergence; the synergistic effect of multiple termination strategies ensures that the global optimization process achieves the best balance between efficiency and convergence.

[0055] Furthermore, a channel-weighted and Bayesian optimization method for medical image processing is provided, wherein the candidate classifier set includes logistic regression, support vector machine, random forest, and XGBoost classifiers.

[0056] The performance score includes accuracy, sensitivity, specificity, precision, and objective function.

[0057] In the above scheme, by integrating multiple candidate classifiers such as logistic regression, support vector machine, random forest and XGBoost, and comprehensively evaluating them based on five indicators: accuracy, sensitivity, specificity, precision and AUC, the generalization ability of each classifier can be comprehensively measured from the perspective of different decision boundaries and model complexity. This allows for the automatic selection of the optimal prediction engine that best suits the current data features, effectively avoiding the risks of overfitting or underfitting that may be caused by a single model.

[0058] Furthermore, a channel-weighted and Bayesian optimization method for medical image processing, which divides case samples into risk groups based on the predicted probability of objective mitigation tendency and the optimal cutoff threshold, uses an XGBoost classifier to obtain the marginal contribution of each feature in the globally optimal feature subset, and integrates them to obtain a feature attribution map, includes the following sub-steps:

[0059] By using feature subsets and clinical labels, the optimal cutoff threshold is obtained. The predicted probability of objective remission tendency is compared with the optimal cutoff threshold to classify the case samples into risk groups.

[0060] The XGBoost classifier processes each feature in the globally optimal feature subset using a tree structure to obtain the marginal contribution of each feature, and integrates the marginal contributions of each feature to generate a feature attribution map.

[0061] In the above scheme, the optimal truncation threshold is automatically determined using the training set, and the predicted probability is transformed into high and low risk groups, giving the single probability value a clear hierarchical interpretation capability. At the same time, the marginal contribution of each feature is calculated based on the tree structure of the XGBoost classifier and integrated to generate a feature attribution map, which intuitively shows the direction and strength of the effect of each feature on the prediction result, providing a quantifiable white-box interpretation basis for the image processing results, and significantly improving the transparency of the processing process and the richness of the output information.

[0062] A medical image processing system with channel weighting and Bayesian optimization includes an image preprocessing module, a feature extraction module, a feature weighting module, a Bayesian iterative optimization module, a probability acquisition module, a risk classification module, and a map generation module.

[0063] The image preprocessing module performs bias field correction, anisotropic diffusion filtering, B-spline interpolation resampling, and normalization on the multi-sequence raw MRI images to obtain multi-sequence standardized image data.

[0064] The feature extraction module extracts features based on lesion masks and multi-sequence normalized image data to obtain an initial feature set, and removes unstable features from the initial feature set to obtain a stable feature set;

[0065] The feature weighting module processes the stable feature set through the GBDT model to obtain information gain, aggregates the information gain to obtain the original representative weight of each sequence, performs smooth normalization on the original representative weight to obtain the prior weight, and uses the prior weight and the stable feature set element by element to obtain the high-dimensional weighted feature matrix.

[0066] The Bayesian iterative optimization module constructs a feature subset based on the high-dimensional weighted feature matrix, and performs Bayesian iteration on the feature subset to obtain the globally optimal feature subset;

[0067] The probability acquisition module uses the candidate classifier set to process the globally optimal feature subset through multi-fold cross-validation, comprehensively evaluates the performance score of the candidate classifier set on the globally optimal feature subset, selects the best classifier, and outputs the predicted probability of objective mitigation tendency.

[0068] The risk classification and attribution map generation module divides case samples into risk groups based on the predicted probability of objective mitigation tendency and the optimal cutoff threshold. It then uses the XGBoost classifier to obtain the marginal contribution of each feature in the globally optimal feature subset and integrates them to obtain a feature attribution map.

[0069] In the above scheme, the system, through the collaborative work of image preprocessing, feature extraction, feature weighting, Bayesian iterative optimization, probability acquisition, risk classification, and attribution map generation modules, forms a complete automated processing chain from raw multi-sequence MRI image input to predicted probability output and interpretable attribution map generation. The feature weighting module introduces sequence-level prior weights, enabling the system to differentiate the information value of different image sequences; the Bayesian iterative optimization module achieves global and efficient search in the high-dimensional feature space based on a Gaussian process surrogate model, ensuring low redundancy and strong representational ability of the optimal feature subset; the risk classification and attribution map generation module simultaneously outputs hierarchical risk groups and feature marginal contribution attribution maps, significantly improving the information dimension and interpretability of the system processing results; the data transfer between the various modules of the system is smooth and the functional decoupling is clear, demonstrating a comprehensive advantage of high accuracy, high robustness, and high transparency when processing high-dimensional multimodal medical image data. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a structural diagram of a magnetic resonance-guided target region boundary extraction system.

[0072] Figure 2 A radar chart showing the performance of the candidate classifier set.

[0073] Figure 3 This is a feature attribution map. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0075] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements or an indirect connection via other elements.

[0076] like Figure 1 As shown, a channel-weighted and Bayesian optimization method for medical image processing includes the following sub-steps:

[0077] S1: Perform bias field correction, anisotropic diffusion filtering, B-spline interpolation resampling, and normalization on the original multi-sequence MRI images to obtain multi-sequence standardized image data.

[0078] S11: The intensity of the original MRI image at the voxel location is corrected using the N4ITK method to obtain the intensity of the MRI image at the voxel location after bias field correction. The formula is:

[0079] ;

[0080] in, Original MRI image at voxel location The strength, Bias field estimated by the N4ITK method at voxel location The value, For bias field corrected MRI images at voxel positions The intensity.

[0081] S12: The intensity of the bias-field corrected MRI image at the voxel location is obtained by suppressing background noise through anisotropic diffusion filtering, using the following formula:

[0082] ;

[0083] in, For diffusion time 0, the MRI image at the voxel location The strength, For diffusion, the MRI image at time t is located at the voxel position. The intensity, where t is the diffusion time index, t=0,1,2...T, and T is the total diffusion time. For divergence operators, The diffusion coefficient function, For diffusion, the MRI image at time t is located at the voxel position. Spatial gradient, For edge-sensitive thresholds, For diffusion T-time MRI images at voxel locations The strength, For the filtered MRI image at voxel locations The intensity.

[0084] S13: Resample the intensity of the filtered MRI image at the voxel location to isotropic voxels using B-spline interpolation. The intensity of the spatially aligned MRI image at the target voxel location is obtained using the following formula:

[0085] ;

[0086] in, voxel position Index in the x-axis direction, voxel position Index in the y-axis direction, voxel position Index in the z-axis direction For spatially aligned MRI images at the target voxel location The intensity, where Z is the set of integers. For filtered MRI images in voxel index (voxel position) The corresponding voxel index is The strength of ) For cubic B-spline basis functions, Location of the target voxel The floating-point coordinates along the x-axis after mapping to the original space. Location of the target voxel The floating-point coordinates along the y-axis after mapping to the original space. Location of the target voxel The floating-point coordinates along the z-axis after mapping to the original space. To control the grid spacing using B-splines, Location of the target voxel Floating-point coordinates mapped to the original space. Let be the coordinate mapping function from the target space to the original space. Location of the target voxel Index in the x-axis direction, Location of the target voxel Index in the y-axis direction, Location of the target voxel Index in the z-axis direction.

[0087] It should be noted that the summation range only needs to satisfy the following conditions: , , .

[0088] S14: The intensity of the spatially aligned MRI image at the target voxel location is normalized using the Z-score method to obtain the normalized intensity of the MRI image at the target voxel location, as shown in the formula:

[0089] ;

[0090] in, To determine the location of the target voxel in the standardized MRI image The strength, This represents the mean intensity of the MRI image at all target voxel locations after spatial alignment. This represents the standard deviation of intensity at all target voxel locations in the spatially aligned MRI image.

[0091] S15: Integrate the intensity of the MRI images from each sequence at the target voxel location after standardization to obtain multi-sequence standardized image data, using the following formula:

[0092] ;

[0093] in, For multi-sequence standardized image data, Location of the target voxel in the MRI image after T1-weighted sequence normalization The strength, Location of the target voxel in the MRI image after T2-weighted sequence normalization The strength, To determine the location of the target voxel in the standardized MRI images of the dynamically enhanced arterial phase sequence. The strength, Location of target voxel in MRI images after portal venous phase sequence standardization The strength, , Weighting for T1, Weighting for T2, To dynamically enhance the arterial phase, This is the portal venous phase.

[0094] S2: Based on lesion masking and multi-sequence normalized image data, extract features to obtain an initial feature set, remove unstable features from the initial feature set, and obtain a stable feature set.

[0095] S21: Through lesion masking The first-order statistical features, shape features, texture features, and frequency domain features are extracted from the corresponding multi-sequence standardized image data and integrated to form an initial feature set.

[0096] S211: Through lesion masking For each sequence of standardized image data, the mean, variance, skewness, kurtosis, and entropy of each sequence of standardized image data are extracted and integrated to obtain the first-order statistical features, as shown in the formula:

[0097] ;

[0098] in, Here, we define the first-order statistical characteristics of the standardized image data for sequence i, where i is the MRI image sequence index. , The mean of the image data for sequence i is standardized. Standardize the variance of the image data for sequence i. To standardize the skewness of the image data for sequence i, Normalize the kurtosis of the image data for sequence i. The entropy of the image data is normalized for sequence i.

[0099] It is important to note the lesion mask. At that time, determine the location of the corresponding target voxel. To find the location of each target voxel in the normalized MRI image of each sequence. The intensity is calculated, including mean, variance, skewness, kurtosis, and entropy.

[0100] S212: Via lesion masking The resulting three-dimensional region is analyzed by extracting its volume, surface area, sphericity, and compactness. These parameters are then integrated to obtain shape features, as shown in the formula:

[0101] ;

[0102] in, For shape features, For the volume of the three-dimensional region, The surface area of ​​the three-dimensional region. For the sphericity of the three-dimensional region, The compactness of the three-dimensional region.

[0103] It should be noted that the lesion mask is shared by multiple sequences, so it does not distinguish between sequences.

[0104] S213: Based on lesion masking For each sequence of normalized image data, a gray-level co-occurrence matrix and a gray-level run-length matrix are constructed using the corresponding normalized image data. Contrast, energy, correlation, short run-length emphasis, and long run-length emphasis are obtained using the gray-level co-occurrence matrix and the gray-level run-length matrix, and integrated to obtain texture features. The formula is as follows:

[0105] ;

[0106] in, To normalize the texture features of the image data for sequence i, To normalize the gray-level co-occurrence matrix contrast of the image data for sequence i, The gray-level co-occurrence matrix energy of the i-sequence image data is normalized. To normalize the gray-level co-occurrence matrix correlation of the image data for sequence i, For the gray-level run-length matrix of the i-sequence normalized image data, short run-length emphasis is required. Long run emphasis is applied to the grayscale run matrix of the i-sequence normalized image data.

[0107] S214: Perform a 3D wavelet transform on each sequence of normalized image data to obtain multiple wavelet subband images. Perform a Gaussian-Laplace transform on each sequence of normalized image data to obtain LoG filtered images at multiple scales. Then, use a lesion mask. For the corresponding wavelet subband image and LoG filtered image, extract the mean and variance of the wavelet subband image and LoG filtered image, and integrate them to obtain the frequency domain features. The formula is as follows:

[0108] ;

[0109] in, To normalize the frequency domain features of the image data for sequence i, Let be the mean intensity of the k-th wavelet sub-band image after the normalized image data transformation of sequence i, where k is the sequential index of the wavelet sub-band image. The total number of wavelet subband images. Let i be the variance of the image intensity of the k-th wavelet subband of the normalized image data transformation. Let m be the mean LoG filtered image intensity at the m-th scale of the normalized image data transformation of sequence i, where m is the LoG filtered image scale index. The total number of scales in the LoG-filtered image. Let be the variance of the LoG-filtered image intensity at the m-th scale of the image data transformation for the i-th sequence normalized image data.

[0110] In this example, k=8.

[0111] S215: Integrate first-order statistical features, shape features, texture features, and frequency domain features to form an initial feature set, using the following formula:

[0112] ;

[0113] in, This is the initial feature set.

[0114] It is important to note the shape features Shared by each sequence.

[0115] S22: Use the within-group correlation coefficient (ICC) to screen the stability of the initial feature set, remove unstable features, and construct a stable feature set. The formula is as follows:

[0116] ;

[0117] in, For a stable feature set, Let i be the j-th feature of the i-th sequence in the initial feature set. For the j-th feature of sequence i The intragroup correlation coefficient, If j is the feature index within the sequence i.

[0118] S3: Process the stable feature set through the GBDT model to obtain information gain, aggregate the information gain to obtain the original representative weight of each sequence, smooth and normalize the original representative weight to obtain the prior weight, and use the prior weight and the stable feature set element by element to obtain the high-dimensional weighted feature matrix.

[0119] S31: The GBDT model processes the j-th feature of sequence i in the stable feature set and the clinical label corresponding to the feature, obtains the information gain of the j-th feature of sequence i in the stable feature set, and aggregates the information gain to obtain the original representative weight of sequence i, as shown in the formula:

[0120] ;

[0121] in, Let i be the original representative weight of sequence i. Let i be the total number of features of sequence i. The information gain of the j-th feature of sequence i in the stable feature set.

[0122] S32: The Softmax function with a temperature coefficient smoothly normalizes the original representative weights of sequence i to obtain the prior weights of sequence i, as shown in the formula:

[0123] ;

[0124] in, Let i be the prior weights of sequence i. For the summation variable, It is a natural exponential function. Let be the original representative weight of the z-th sequence. This is the temperature coefficient.

[0125] In the embodiments, The results showed that the portal venous phase sequence had the highest prior weight, approximately 0.35-0.45. This is because during the portal venous phase, the tumor presents a low signal due to rapid inflow and outflow, which contrasts sharply with the surrounding significantly enhanced normal liver parenchyma. This makes the tumor boundary, the integrity of the pseudocapsule, and the abnormal perfusion zone around the tumor most clearly visible during the portal venous phase. The next highest prior weight was for the dynamic enhancement arterial phase sequence, approximately 0.30-0.40, followed by the T2-weighted sequence, approximately 0.15-0.25, and the T1-weighted sequence had the lowest prior weight, approximately 0.10-0.20.

[0126] It is important to note that smoothing normalization is performed to prevent the i-sequence (e.g., The original representative weights of the sequence are too concentrated, causing other sequences (e.g.) to be affected. Valuable supplementary information was obscured when As the weight increases, the weight distribution of each sequence becomes smoother.

[0127] S33: Multiply the prior weights of each sequence element-wise with the stable feature set to obtain a high-dimensional weighted feature matrix, as shown in the formula:

[0128] ;

[0129] in, It is a high-dimensional weighted feature matrix. For element-wise multiplication, It can be any.

[0130] In existing image group techniques, LASSO regression is commonly used for feature selection and dimensionality reduction. The optimal regression coefficients are obtained by adding an L1 regularization penalty term to the squared error loss function, as shown in the formula:

[0131] ;

[0132] in, The optimal regression coefficients are... To return To reach the minimum , For regression coefficients, The total number of case samples, For case sample index, For the first Clinical labels of individual case samples, This is a radiomics feature index (or, if implemented in this invention, a feature index for a stable feature set). This represents the total number of radiomics features (or, in the case of this invention, the total number of features in a stable feature set). For the first The first case sample One radiomics feature For the first Regression coefficients of radiomics features This is the regularization hyperparameter.

[0133] However, LASSO regression, which typically treats all features equally, fails to incorporate prior clinical knowledge. For example, in liver cancer diagnosis, blood supply information during the dynamically enhanced arterial phase is far more important than T1-weighted anatomical information, but LASSO cannot achieve differentiated priors. Furthermore, LASSO is essentially a linear model, tending to randomly retain one feature from highly correlated feature groups while discarding the rest, easily losing feature groups with non-linear predictive power. Additionally, when the total number of radiomics features reaches thousands, LASSO's regularization hyperparameters become problematic. Grid search is inefficient and cannot actively explore higher-order interactions between image omics features. Therefore, this invention uses Bayesian global optimization to solve the technical problem of difficulty in meeting the requirements of high-dimensional, nonlinear, and multimodal data.

[0134] S4: Construct a feature subset based on the high-dimensional weighted feature matrix, and perform Bayesian iteration on the feature subset to obtain the globally optimal feature subset.

[0135] S41: Define a binary mask vector using the dimension of the high-dimensional weighted feature matrix, and determine the feature subset using the binary mask vector. The formula is as follows:

[0136] ;

[0137] in, For the d-th weighted feature of the high-dimensional weighted feature matrix, The dimension of the high-dimensional weighted features. ,when When selecting the d-th weighted feature of the high-dimensional weighted feature matrix, For feature subset, Let d be the d-th mask component of the binary mask vector. It is a binary mask vector.

[0138] S42: Train an XGBoost classifier using feature subsets and their corresponding clinical labels. Perform repeated 5-fold cross-validation on the trained XGBoost classifier to obtain the objective function, as shown in the formula:

[0139] ;

[0140] in, Binary mask vector The objective function, To calculate the average AUC operation, To train the XGBoost classifier, For clinical labeling.

[0141] S43: Assuming the objective function follows a Gaussian distribution, construct a Gaussian process surrogate model, the formula of which is:

[0142] ;

[0143] in, To obey, For Gaussian processes, Binary mask vector The predicted mean, For the covariance function, the Matern5 / 2 kernel function is selected. For another binary mask vector, For signal variance, For two binary mask vectors and Manhattan distance between them For length dimensions, This is a Gaussian process proxy model.

[0144] It is important to note that real-world data (such as loss surfaces optimized by hyperparameters, blood perfusion dynamics inside tumors, or spatial distribution of pathological features) are not smooth. When the smoothing parameter is 5 / 2, second-order differentiability means that the function not only has a continuous slope (velocity) but also a continuous curvature (acceleration). This allows for relatively abrupt changes in the function curve locally while ensuring overall smoothness and coherence.

[0145] S44: Obtain the expected gain based on the predicted mean and predicted standard deviation of the Gaussian process, using the following formula:

[0146] ;

[0147] in, Binary mask vector Expected gain, For the optimal mask vector The objective function, For the optimal mask vector, To explore parameters (which can be 0, 0.005, or 0.01), The cumulative distribution function of the standard normal distribution. Let be the probability density function of the standard normal distribution. Binary mask vector The standard deviation of the prediction This represents the improvement amount after standardization.

[0148] S45: Repeat the expected gain acquisition and objective function evaluation, update the Gaussian process surrogate model as a Bayesian iteration, and stop the iteration when the termination condition is met to obtain the globally optimal feature subset. The termination condition is:

[0149] If the maximum number of Bayes iterations If the iteration stops, then stop.

[0150] If the improvement of the objective function within 100 consecutive Bayesian iterations If the iteration stops, then stop.

[0151] If the maximum expected gain is within 50 consecutive Bayesian iterations If the iteration stops, then stop.

[0152] If the Jaccard similarity coefficient between the optimal feature subsets during continuous Bayes iteration is... If the iteration stops, then stop.

[0153] in, The optimal feature subset for the current features. It is the optimal feature subset of historical features.

[0154] In this embodiment, the initial number of random sampling points is set to 50, and the maximum number of Bayesian iterations is set to 1500. To avoid increasing model complexity and overfitting due to selecting too many features, each... The number of features can be set to upper and lower limits, with a lower limit of 3 and an upper limit of 20.

[0155] S5: The candidate classifier set is processed through 5-fold cross-validation on the globally optimal feature subset. The performance scores of the candidate classifier set on the globally optimal feature subset are comprehensively evaluated, and the best classifier is selected to output the predicted probability of the objective tendency to mitigate (ORR). The formula is as follows:

[0156] ;

[0157] in, To objectively predict the probability of mitigation tendency, For the best classifier, .

[0158] The candidate classifier set includes Logistic Regression, Support Vector Machine (SVM), Random Forest, and XGBoost classifiers;

[0159] The performance rating includes accuracy, sensitivity, specificity, precision, and objective function (AUC).

[0160] In this embodiment, the performance radar chart of the candidate classifier set selecting the best classifier is as follows: Figure 2 As shown, the XGBoost classifier was ultimately selected as the best classifier, with an objective function of 0.932.

[0161] S6: Divide the case samples into risk groups by the predicted probability of objective mitigation tendency and the optimal cutoff threshold, use the XGBoost classifier to obtain the marginal contribution of each feature in the globally optimal feature subset, and integrate them to obtain the feature attribution map.

[0162] S61: Obtain the optimal cutoff threshold using feature subsets and clinical labels. The predicted probability of objective remission tendency is compared with the optimal cutoff threshold to classify the case samples into risk groups, using the following formula:

[0163] ;

[0164] It is important to note that the risk group can be mapped to the difference between progression-free survival (PFS) and overall survival (OS) in long-term statistics.

[0165] S62: Process each feature in the globally optimal feature subset through the tree structure (node ​​splitting path and leaf node weight) of the XGBoost classifier to obtain the marginal contribution of each feature, and integrate the marginal contributions of each feature to generate a feature attribution map.

[0166] It is important to note the optimal cutoff threshold. Determined by the ROC curves and Youden index of the training set, the feature attribution map can explain, for example, texture features representing the homogeneity of water signal distribution within the tumor.

[0167] In the above scheme, a 6-dimensional optimal feature subset was selected from a 4250-dimensional feature subset through 1500 maximum Bayes iterations, achieving a feature compression rate of 99.86%. Compared to the 11-dimensional feature set obtained by traditional LASSO, the number of features was further reduced by 45.45%, and the selected globally optimal feature subset had lower internal correlation and less redundancy. The XGBoost prediction model built based on the globally optimal feature subset achieved an objective function of 0.932 on the internal validation set, an improvement of 0.033 compared to the existing technology's 0.899. The performance difference between the training set and the validation set was controlled within a low range, effectively reducing the risk of overfitting and improving generalization ability. After risk stratification based on the predicted probability values, the median progression-free survival for the high-risk group and the low-risk group were 19.5 months and 7.2 months, respectively, with a statistically significant difference (p<0.001).

[0168] In the embodiments, such as Figure 3 As shown, each point represents a single feature of the case sample. Finally, from the feature attribution map, six omics features associated with the response to triple therapy for liver cancer were selected, namely:

[0169] 1. Log_sigma_3_0_mm_3D_firstorder_Entropy (from dynamically enhanced arterial phase sequences)

[0170] 2. Log_sigma_3_0_mm_3D_ngtdm_Busyness (from portal venous phase sequence)

[0171] 3. Log_sigma_5_0_mm_3D_firstorder_Kurtosis (from T1-weighted sequences)

[0172] 4. Log_sigma_5_0_mm_3D_gldm_LowGrayLevelEmphasis (from portal venous phase sequence)

[0173] 5. Original_firstorder_Kurtosis (from dynamic enhanced arterial phase sequence)

[0174] 6. Log_sigma_3_0_mm_3D_glcm_InverseVariance (from T2 weighted sequence).

[0175] It should be noted that the specific ways in which each module performs its operations in the above embodiments have been described in detail in the embodiments of the system, and will not be elaborated here.

[0176] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A medical image processing method using channel weighting and Bayesian optimization, characterized in that, Includes the following sub-steps: Bias field correction, anisotropic diffusion filtering, B-spline interpolation resampling, and normalization were performed on the original multi-sequence MRI images to obtain standardized multi-sequence image data. Based on lesion masking and multi-sequence normalized image data, features are extracted to obtain an initial feature set. Unstable features in the initial feature set are then removed to obtain a stable feature set. The GBDT model is used to process the stable feature set to obtain information gain. The information gain is aggregated to obtain the original representative weight of each sequence. The original representative weight is smoothed and normalized to obtain the prior weight. The prior weight and the stable feature set are used element by element to obtain the high-dimensional weighted feature matrix. A feature subset is constructed based on the high-dimensional weighted feature matrix, and a Bayesian iteration is performed on the feature subset to obtain the globally optimal feature subset; The candidate classifier set is processed through multi-fold cross-validation to process the globally optimal feature subset. The performance score of the candidate classifier set on the globally optimal feature subset is comprehensively evaluated, and the best classifier is selected to output the predicted probability of objective mitigation tendency. The risk groups of case samples are divided by the predicted probability of objective mitigation tendency and the optimal cutoff threshold. The marginal contribution of each feature in the globally optimal feature subset is obtained by using the XGBoost classifier and then integrated to obtain the feature attribution map.

2. The medical image processing method based on channel weighting and Bayesian optimization according to claim 1, characterized in that, The process of performing bias field correction, anisotropic diffusion filtering, B-spline interpolation resampling, and normalization on the multi-sequence raw MRI images to obtain multi-sequence standardized image data includes the following sub-steps: The intensity of the original MRI image at the voxel location was corrected by bias field using the N4ITK method to obtain the intensity of the bias field corrected MRI image at the voxel location. The intensity of the bias-field corrected MRI image at the voxel location is obtained by suppressing background noise through anisotropic diffusion filtering. The intensity of the filtered MRI image at the voxel location is resampled to an isotropic voxel using B-spline interpolation to obtain the intensity of the spatially aligned MRI image at the target voxel location. The intensity of the spatially aligned MRI image at the target voxel location was normalized using the Z-score method to obtain the intensity of the normalized MRI image at the target voxel location. The intensity of each normalized MRI image sequence at the target voxel location is integrated to obtain multi-sequence normalized image data.

3. The medical image processing method based on channel weighting and Bayesian optimization according to claim 1, characterized in that, The process of extracting features from lesion masks and multi-sequence normalized image data to obtain an initial feature set, and then removing unstable features from the initial feature set to obtain a stable feature set includes the following sub-steps: Through lesion masking First-order statistical features, shape features, texture features, and frequency domain features are extracted from the corresponding multi-sequence standardized image data and integrated to form an initial feature set; The stability of the initial feature set is screened using the intragroup correlation coefficient, unstable features are removed, and a stable feature set is constructed.

4. The medical image processing method based on channel weighting and Bayesian optimization according to claim 3, characterized in that, The lesion mask The process of extracting first-order statistical features, shape features, texture features, and frequency domain features from the corresponding multi-sequence standardized image data and integrating them to form an initial feature set includes the following sub-steps: Through lesion masking For each sequence of standardized image data, the mean, variance, skewness, kurtosis, and entropy of each sequence of standardized image data are extracted and integrated to obtain first-order statistical features; Through lesion masking The volume, surface area, sphericity, and compactness of the formed three-dimensional region are extracted and integrated to obtain shape features; Based on lesion masking For each sequence of normalized image data, a gray-level co-occurrence matrix and a gray-level run-length matrix are constructed using the corresponding sequence of normalized image data. Contrast, energy, correlation, short run-length emphasis, and long run-length emphasis are obtained using the gray-level co-occurrence matrix and the gray-level run-length matrix, and texture features are obtained by integration. Perform a 3D wavelet transform on each sequence of normalized image data to obtain multiple wavelet sub-band images. Then, perform a Laplace-Gaussian transform on each sequence of normalized image data to obtain LoG-filtered images at multiple scales. Finally, use lesion masking... The mean and variance of the wavelet subband image and the LoG filtered image are extracted and integrated to obtain the frequency domain features. First-order statistical features, shape features, texture features, and frequency domain features are integrated to form an initial feature set.

5. The medical image processing method based on channel weighting and Bayesian optimization according to claim 1, characterized in that, The process of processing the stable feature set using the GBDT model to obtain information gain, aggregating the information gain to obtain the original representative weights of each sequence, smoothing and normalizing the original representative weights to obtain prior weights, and using the prior weights and the stable feature set element-wise to obtain the high-dimensional weighted feature matrix includes the following sub-steps: The GBDT model processes the j-th feature of sequence i in the stable feature set and the clinical label corresponding to the feature to obtain the information gain of the j-th feature of sequence i in the stable feature set. The information gain is aggregated to obtain the original representative weight of sequence i. The Softmax function with a temperature coefficient smoothly normalizes the original representative weights of sequence i to obtain the prior weights of sequence i. The prior weights of each sequence and the stable feature set are multiplied element-wise to obtain a high-dimensional weighted feature matrix.

6. The medical image processing method based on channel weighting and Bayesian optimization according to claim 1, characterized in that, The step of constructing a feature subset based on a high-dimensional weighted feature matrix and then performing Bayesian iteration on the feature subset to obtain the globally optimal feature subset includes the following sub-steps: The binary mask vector is defined by the dimension of the high-dimensional weighted feature matrix, and the feature subset is determined by the binary mask vector; An XGBoost classifier is trained using feature subsets and their corresponding clinical labels. The trained XGBoost classifier is then subjected to repeated multi-fold cross-validation to obtain the objective function. Assuming the objective function follows a Gaussian distribution, construct a Gaussian process surrogate model; The expected gain is obtained based on the predicted mean and predicted standard deviation of the Gaussian process. Repeatedly perform expected gain acquisition and objective function evaluation, update the Gaussian process surrogate model as a Bayesian iteration, and stop the iteration when the termination condition is met to obtain the globally optimal feature subset.

7. The medical image processing method based on channel weighting and Bayesian optimization according to claim 6, characterized in that, The termination condition is: If the maximum number of Bayes iterations If the iteration stops, then stop. If the improvement of the objective function within 100 consecutive Bayesian iterations If the iteration stops, then stop. If the maximum expected gain is within 50 consecutive Bayesian iterations If the iteration stops, then stop. If the Jaccard similarity coefficient between the optimal feature subsets during continuous Bayes iteration is... If the iteration stops, then stop. in, The optimal feature subset for the current features. It is the optimal feature subset of historical features.

8. The medical image processing method based on channel weighting and Bayesian optimization according to claim 1, characterized in that, The candidate classifier set includes logistic regression, support vector machine, random forest, and XGBoost classifiers; The performance score includes accuracy, sensitivity, specificity, precision, and objective function.

9. The medical image processing method based on channel weighting and Bayesian optimization according to claim 1, characterized in that, The process of dividing case samples into risk groups based on the predicted probability of objective remission tendency and the optimal cutoff threshold, obtaining the marginal contribution of each feature in the globally optimal feature subset using the XGBoost classifier, and integrating them to obtain the feature attribution map includes the following sub-steps: By using feature subsets and clinical labels, the optimal cutoff threshold is obtained. The predicted probability of objective remission tendency is compared with the optimal cutoff threshold to classify the case samples into risk groups. The XGBoost classifier processes each feature in the globally optimal feature subset using a tree structure to obtain the marginal contribution of each feature, and integrates the marginal contributions of each feature to generate a feature attribution map.

10. A medical image processing system using channel weighting and Bayesian optimization, for implementing the medical image processing method using channel weighting and Bayesian optimization as described in any one of claims 1-9, characterized in that, It includes an image preprocessing module, a feature extraction module, a feature weighting module, a Bayesian iterative optimization module, a probability acquisition module, a risk classification module, and a map generation module; The image preprocessing module performs bias field correction, anisotropic diffusion filtering, B-spline interpolation resampling, and normalization on the multi-sequence raw MRI images to obtain multi-sequence standardized image data. The feature extraction module extracts features based on lesion masks and multi-sequence normalized image data to obtain an initial feature set, and removes unstable features from the initial feature set to obtain a stable feature set; The feature weighting module processes the stable feature set through the GBDT model to obtain information gain, aggregates the information gain to obtain the original representative weight of each sequence, performs smooth normalization on the original representative weight to obtain the prior weight, and uses the prior weight and the stable feature set element by element to obtain the high-dimensional weighted feature matrix. The Bayesian iterative optimization module constructs a feature subset based on the high-dimensional weighted feature matrix, and performs Bayesian iteration on the feature subset to obtain the globally optimal feature subset; The probability acquisition module uses the candidate classifier set to process the globally optimal feature subset through multi-fold cross-validation, comprehensively evaluates the performance score of the candidate classifier set on the globally optimal feature subset, selects the best classifier, and outputs the predicted probability of objective mitigation tendency. The risk classification and attribution map generation module divides case samples into risk groups based on the predicted probability of objective mitigation tendency and the optimal cutoff threshold. It then uses the XGBoost classifier to obtain the marginal contribution of each feature in the globally optimal feature subset and integrates them to obtain a feature attribution map.

Images