Brain image analysis method and system based on multi-modal fusion

Abstract

The invention discloses a brain image analysis method and system based on multi-modal fusion, and relates to the technical field of brain image processing. A brain image analysis system based on multi-modal fusion comprises a data acquisition module, a data preprocessing module, a feature extraction and fusion module, a brain network analysis module, a comprehensive analysis module and a focus detection module. The comprehensive analysis model adopts a double-branch structure, deep processing is performed on multi-modal features and brain network features, and interactive fusion of the two types of features is realized through a cross-modal attention mechanism; the model is further combined with a classification branch and a regression branch to cooperatively complete brain disease classification and focus quantitative analysis, and the adaptive capacity and diagnosis performance of an existing model in a complex task scene are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of brain image analysis, and particularly to a brain image analysis method and system based on multimodal fusion. Background Art

[0002] Brain image analysis technology plays an important role in the diagnosis, treatment evaluation and scientific research of nervous system diseases. With the rapid development of medical imaging equipment, multimodal imaging technology has gradually become an important tool in clinical and scientific research. Through different modalities of image data, multi-dimensional information such as brain anatomy, function and metabolism can be provided, thus providing support for the accurate diagnosis and personalized treatment of brain diseases.

[0003] Currently, the analysis methods of multimodal images mainly focus on single-modal processing and simple modal feature superposition. Some studies evaluate brain structure changes through texture analysis of structural images, or use metabolic features of metabolic images to judge the metabolically active regions of diseases. In addition, functional images have been widely used to evaluate the functional connectivity between brain regions. However, these methods are usually based on single-modal information and lack in-depth mining of the potential associations between multimodal data, which may lead to one-sided interpretation of brain lesion information.

[0004] On the other hand, multimodal image analysis technology faces many challenges in practical applications: it is difficult to align the space of images between modalities; multimodal image data has different resolutions, signal-to-noise ratios and intensity distributions, and it is difficult to directly fuse different modal features, and the existing simple superposition methods are prone to losing key information; the analysis of functional images and structural images is usually independent, but brain diseases involve the combined effects of functional disorders and structural abnormalities, and it is difficult for existing technologies to simultaneously characterize the association between the two; the current lesion detection methods mainly rely on the global features of images or simple region segmentation algorithms, and fail to combine multimodal data for refined analysis.

[0005] In the above background, multimodal image analysis technology urgently needs to be further improved to solve problems such as difficult data alignment between modalities, insufficient feature fusion, and isolated analysis of functional and structural information. The existence of these problems limits the application effect of existing technologies in the diagnosis of complex brain diseases, and also makes it a challenge to accurately locate and quantitatively analyze specific lesions. Summary of the Invention

[0006] A brain image analysis method based on multimodal fusion includes the following steps:

[0007] Step S1, data acquisition: acquiring image data containing multimodal brain images through a medical imaging device or an image database, where the image data includes structural images, functional images and metabolic images;

[0008] Step S2, Data Preprocessing: Denoise, intensity normalize, and perform inter-modal registration on the brain image data. The inter-modal registration is implemented using an elastic registration algorithm, and a brain region is extracted based on a template-based segmentation method;

[0009] Step S3, Feature Extraction and Fusion: Extract multi-modal features from the preprocessed image data, including texture features, morphological features, and metabolic features. The texture features are extracted by the gray-level co-occurrence matrix method, the morphological features are obtained by extracting the volume information of key brain structures from the structural image through a segmentation algorithm, and the metabolic features are obtained by statistically analyzing the intensity index of the metabolic hot spot regions in the metabolic image. Standardize the extracted multi-modal features, assign modal feature weights based on the entropy weight method, and fuse the multi-modal features through the principal component analysis method;

[0010] Step S4, Brain Network Analysis: Based on the functional image and the structural image, construct a brain functional connection network and a brain structural connection network. Among them, the brain functional connection network is obtained by calculating the correlation of time series of different brain regions; the brain structural connection network is obtained by generating a white matter connection matrix through fiber tracking technology. Extract the graph features of the brain functional connection network and the brain structural connection network, including node degree, clustering coefficient, and characteristic path length;

[0011] Step S5, Comprehensive Analysis: Input the fused multi-modal features and graph features into the comprehensive analysis model to output the classification of brain diseases and related evaluation results. The comprehensive analysis model adopts a dual-branch structure and realizes the deep fusion of the two types of features through a cross-modal attention mechanism;

[0012] Step S6, Lesion Detection: Based on the comprehensive analysis results, use the region growing algorithm to detect and segment the lesion region, quantify the volume, morphology, and metabolic features of the lesion, and generate an auxiliary diagnosis report.

[0013] As a preferred technical solution of the present invention, the denoising, intensity normalization, and inter-modal registration of the brain image data include:

[0014] Denoising Processing: Adopt the non-local means denoising algorithm. For each pixel point of the image, calculate the gray similarity between it and the surrounding pixel blocks, and perform weighted averaging on the pixel values based on the similarity weights to remove high-frequency noise while retaining texture features and edge information;

[0015] Intensity normalization: Linear normalization. For structural image data, calculate the maximum and minimum values, and perform linear mapping on all pixel values to normalize the corresponding pixel intensities to the range of [0, 1] to eliminate the influence of equipment or scanning parameters during the data acquisition process. For functional images, based on the z-score normalization method, calculate the mean and standard deviation of the whole-brain functional signals, and standardize the signal values corresponding to each pixel to a distribution with a mean of 0 and a standard deviation of 1. Piecewise normalization. For the metabolic hot spot regions of metabolic images, adopt the piecewise linear normalization method to map the high-dynamic-range hot spot intensity values to the interval of 0.8 - 1 respectively, and map the low-intensity background values to the interval of 0 - 0.2 to retain the metabolic characteristics;

[0016] Inter-modal registration: Initial registration. Taking the structural image as the reference modality, use the rigid registration algorithm based on mutual information to preliminarily align the functional image and the metabolic image by optimizing the rotation and translation parameters. Nonlinear registration. On the basis of rigid registration, adopt the elastic registration algorithm based on the multi-resolution strategy to optimize the detail differences layer by layer. Resampling and standardization. Use the cubic B-spline interpolation method to resample the registered images, interpolate the registered image data to a unified resolution, and align all modal images to the standard anatomical template to ensure that all modal images have consistent spatial coordinates and resolutions;

[0017] The elastic configuration algorithm includes: In the low-resolution stage, downsample the image data, calculate the coarse-grained global deformation field, and adjust the overall structure alignment between modalities. In the high-resolution stage, at the fine-grained resolution, locally optimize the deformation field to correct the minor structural differences between modalities.

[0018] As a preferred technical solution of the present invention, the brain region extraction by the template-based segmentation method includes:

[0019] Select a standard anatomical template containing brain anatomical structure partition information, and the template is the MNI152 template or the ICBM152 template; perform spatial standardization processing on the brain image, align the brain image with the template through the elastic registration algorithm so that the two are in the same anatomical space; based on the probability atlas in the template, perform pixel-level probability calculation on the aligned image data to obtain the probability values of each pixel point belonging to different brain regions; apply the segmentation method based on the probability atlas, combine the preset probability threshold, segment the brain region in the image, and identify the gray matter, white matter and cerebrospinal fluid regions; use the region growing algorithm to optimize the boundary of the segmentation result, remove the noise regions and refine the structure boundary; output the segmentation mask file, and the mask file marks different brain regions and their corresponding anatomical structures.

[0020] As a preferred technical solution of the present invention, the feature extraction includes:

[0021] Texture feature extraction: For the preprocessed structural images, calculate the contrast, entropy, and homogeneity features using the gray-level co-occurrence matrix method; adopt the sliding window technique to extract local texture features on the two-dimensional slices of the structural images, and statistically summarize the feature values of all slices to generate a three-dimensional texture feature matrix;

[0022] Morphological feature extraction: Based on the segmentation results of brain regions, calculate the volume features of the gray matter, white matter, and cerebrospinal fluid regions of the structural images; extract the shape features of the segmented regions, including surface area, compactness, and aspect ratio, to generate the morphological feature vectors corresponding to the regions;

[0023] Metabolic feature extraction: For the metabolic hot spots in the metabolic images, calculate the SUVmax, SUVmean, and standard deviation features; generate the metabolic intensity matrix of specific brain regions based on the distribution of metabolic hot spots in different intervals, and normalize it into a standardized metabolic feature vector;

[0024] Group all the extracted features according to the modality and store them in matrix form, where each row represents the feature vector of a segmented region or brain region, and each column represents different feature types for subsequent feature fusion.

[0025] As a preferred technical solution of the present invention, the fusion of multi-modal features includes:

[0026] Feature standardization: Perform standardization processing on the feature vectors extracted from different modality images. Adopt the z-score normalization method to convert each feature into a distribution with a mean of 0 and a standard deviation of 1;

[0027] Feature weight assignment: Calculate the importance weights of different modality features based on the entropy weight method, including constructing an information entropy model for each modality feature to evaluate the difference of feature data; assign corresponding weights to each modality according to the information entropy value, and the weight values are used to adjust the contribution ratio of modality features in the fusion process;

[0028] Feature splicing: Splice the texture features, morphological features, and metabolic features in the column direction according to the feature type and modality to generate a unified high-dimensional feature matrix;

[0029] Dimensionality reduction processing: Use the principal component analysis method to perform dimensionality reduction processing on the high-dimensional feature matrix, extract the main components with a cumulative contribution rate exceeding 95%, and generate low-dimensional fusion feature vectors;

[0030] Store the dimensionality-reduced fusion feature vectors as a two-dimensional matrix according to the patient individuals, where each row represents the features of a patient, and each column represents different fusion feature types for subsequent comprehensive analysis.

[0031] As a preferred technical solution of the present invention, the brain network analysis includes:

[0032] Construction of brain functional connectivity network: Based on functional imaging, time series are extracted for whole-brain partitioning; the time series correlation between every two brain regions is calculated to construct a functional connectivity matrix, where the elements of the matrix represent the correlation coefficients between brain regions;

[0033] Construction of brain structural connectivity network: Based on structural imaging, a white matter fiber connection map is generated using fiber tracking technology; a structural connectivity matrix is constructed, where the elements of the matrix represent the number of fibers or fiber density between brain regions;

[0034] Extraction of brain network features: Graph features are extracted from the brain functional connectivity network and the brain structural connectivity network, including: Node degree: representing the number of connections of each brain region with other brain regions; Clustering coefficient: representing the tightness of the local network of each brain region; Characteristic path length: representing the average length of the shortest path between any two nodes in the network;

[0035] The graph features of the brain functional connectivity network and the brain structural connectivity network are stored in vector format, with each row representing the features of a brain region and each column representing different types of graph features for subsequent comprehensive analysis.

[0036] As a preferred technical solution of the present invention, the structure of the comprehensive analysis model includes:

[0037] Input layer: Used to receive multi-modal fusion features and graph features. Among them, the multi-modal fusion features are a two-dimensional matrix, with rows representing each sample and columns representing multi-modal features; the graph features are graph structure data, including a functional connectivity matrix and a structural connectivity matrix;

[0038] Modal feature processing branch:

[0039] For the multi-modal fusion features, MLP is used to process different modal features respectively. The MLP of each modality consists of two fully connected layers. The first layer uses the ReLU activation function, and the second layer outputs the modal feature embeddings after Batch-ormalization normalization; the modal embedding features are concatenated and input into the feature interaction module;

[0040] For the brain network graph features, the functional connectivity matrix and the structural connectivity matrix are processed through a graph neural network to extract the local and global features of the graph structure; a graph convolutional layer is adopted in the graph neural network to aggregate node information layer by layer and calculate the high-dimensional feature embeddings of nodes; the embedding features of the functional graph and the structural graph are fused into a unified graph feature vector;

[0041] Feature interaction module: For the modal features and the graph features, a cross-modal attention mechanism is adopted to construct query, key, and value matrices for the modal features and the graph features respectively; calculate the attention weight matrix between the modal features and the graph features, and generate an interaction-enhanced feature representation through a weighting mechanism;

[0042] Feature Fusion Module: Input the interaction-enhanced feature representation into a multi-layer stacked fully-connected network to extract comprehensive disease-related features layer by layer; Each layer of the fully-connected network uses the ReLU activation function and adds a Dropout mechanism to avoid overfitting;

[0043] Task Branch Module: Classification Branch, classify brain diseases for the comprehensive features through the Softmax classification layer; Regression Branch, calculate the lesion size, disease severity or progression prediction score through the fully-connected layer;

[0044] Data Processing Flow: After the multi-modal fusion features and brain network graph features are respectively processed by branches, embedded feature vectors are generated; In the feature interaction module, the two types of features are fused to capture the interaction relationship between the modality and the graph features; The comprehensive features are further dimension-reduced and extracted by the fusion module, and finally the classification results and relevant evaluation indicators are output.

[0045] As a preferred technical solution of the present invention, the training of the comprehensive analysis model includes:

[0046] Collect multi-modal imaging data and brain network data, and extract multi-modal fusion features and brain network graph features as inputs; Divide the data set into a training set, a validation set and a test set, with a ratio of 8:1:1;

[0047] Define the objective function of the classification task as the cross-entropy loss function, and the calculation formula is:

[0048] where N is the number of samples, K is the number of categories, y ik is the true label of sample i, is the probability value predicted by the model;

[0049] Define the objective function of the regression task as the mean square error loss function, and the calculation formula is: where z i and are the true value and the value predicted by the model respectively;

[0050] The comprehensive loss function is the multi-task weighted loss: L = αL cls + βL reg , where α and β are the loss weight coefficients of the classification and regression tasks;

[0051] Perform model training, initialize the model parameters, and use the stochastic gradient descent method for optimization; Set the learning rate and update it using the cosine annealing learning rate adjustment strategy, and the formula is: where η t represents the learning rate of the t-th iteration, T is the total number of training rounds, η min and η maxThey are the minimum and maximum values of the learning rate during the training process, respectively;

[0052] In each training iteration, the training set data is input into the model to calculate the prediction results of the classification and regression tasks; calculate the loss function; calculate the gradients through the backpropagation algorithm; and finally update the model parameters: where θ represents the parameters of the model;

[0053] After each training round, the validation set is used to evaluate the model performance, and the classification accuracy, recall rate, and regression mean squared error are recorded; the hyperparameters are adjusted according to the validation results;

[0054] Set an early stopping mechanism to terminate the training when the validation loss does not decrease for several consecutive rounds; save the model parameters with the best performance on the validation set as the final trained model;

[0055] Use the test set to finally evaluate the trained model, calculate the accuracy, AUC value of the classification task, and the mean squared error of the regression task, and output the evaluation results to verify the performance of the model in practical applications.

[0056] As a preferred technical solution of the present invention, the lesion detection includes:

[0057] Receive the segmentation mask data of the gray matter, white matter, and cerebrospinal fluid regions extracted by the template segmentation method; combine the classification results of the comprehensive analysis model to determine the target regions with high disease risks and generate the initial mask of the target regions;

[0058] In the classification results of the comprehensive analysis model, identify the initial seed points within the target regions by locating the high-risk regions that contribute the most to the classification results;

[0059] The selection of the seed points is based on the voxels with significantly abnormal multimodal features, specifically including the statistical threshold conditions of texture, morphology, or metabolic features;

[0060] Starting from the seed points, use the region growing algorithm to expand the lesion regions, specifically including checking whether the eigenvalue of the neighboring voxels around the seed points meets the expansion conditions, including intensity value proximity and spatial connectivity; the voxels that meet the conditions are included in the lesion regions, and the newly added voxels are marked as new seed points and continue to expand; repeat the above steps until the lesion regions stop expanding;

[0061] For the finally expanded lesion regions, calculate their morphological features, including volume, surface area, aspect ratio, and compactness; for metabolic images, calculate the metabolic feature values of the lesion regions, including SUVmax, SUVmean, and the standard deviation of metabolic intensity; superimpose the lesion regions on the original images to generate three-dimensional visualization lesion images; output an auxiliary diagnosis report containing the lesion location, volume, metabolic features, and morphological features for further clinical diagnosis and evaluation.

[0062] A brain image analysis system based on multimodal fusion, comprising the following modules:

[0063] Data acquisition module: used to acquire image data containing multimodal brain images through medical imaging devices or image databases;

[0064] Data preprocessing module: used to denoise, intensity normalize and register between modalities for the brain image data, and extract brain regions based on a template-based segmentation method;

[0065] Feature extraction and fusion module: used to extract multimodal features from the preprocessed image data, perform standardization processing on the extracted multimodal features, and fuse the multimodal features by principal component analysis method;

[0066] Brain network analysis module: used to construct a brain functional connection network and a brain structural connection network based on functional images and structural images, and extract graph features of the brain functional connection network and the brain structural connection network;

[0067] Comprehensive analysis module: used to input the fused multimodal features and graph features into a comprehensive analysis model, and output brain disease classification and related evaluation results;

[0068] Lesion detection module: used to detect and segment the lesion area based on the comprehensive analysis results by using the region growing algorithm, quantify the volume, morphology and metabolic characteristics of the lesion, and generate an auxiliary diagnosis report.

[0069] The present invention has the following advantages:

[0070] The present invention performs standardization processing and fusion analysis on multimodal image data, dynamically assigns importance weights to different modal features by using the entropy weight method, and combines principal component analysis for dimensionality reduction, which improves the efficiency and accuracy of feature fusion while reducing redundant information, and overcomes the problem of insufficient fusion of multimodal features in the prior art.

[0071] The present invention constructs a brain functional connection network and a brain structural connection network at the same time, uses time series correlation and fiber tracking technology to extract functional connection and structural connection matrices, and combines graph features such as node degree, clustering coefficient and characteristic path length, which makes up for the information loss problem caused by independent analysis of functional images and structural images in the prior art, and significantly improves the dimension and accuracy of brain disease diagnosis.

[0072] The present invention combines the classification results of the comprehensive analysis model, intelligently locates high-risk areas as the initial seed points for lesion detection, and uses the region growing algorithm to expand the lesion area, solves the problems of single-modal dependence and inaccurate positioning in lesion detection in the prior art, and further refines the diagnosis results of the lesion through quantitative analysis of morphological features and metabolic features.

[0073] The comprehensive analysis model of the present invention adopts a dual-branch structure, which deeply processes multi-modal features and brain network features respectively, and realizes the interactive fusion of the two types of features through a cross-modal attention mechanism; the model further combines a classification branch and a regression branch to jointly complete brain disease classification and lesion quantification analysis, improving the adaptability and diagnostic performance of the existing model in complex task scenarios.

[0074] The present invention constructs a complete analysis process from data preprocessing to lesion detection, solves the problem of data alignment between modalities through rigid registration based on mutual information and elastic registration based on thin plate spline interpolation, and realizes the optimization of the whole process by feature extraction, fusion and comprehensive analysis model, which has strong robustness and practicability.

[0075] During the model training process, the present invention adopts cosine annealing learning rate adjustment and multi-task weighted loss function, optimizes the training speed and performance stability of the model, overcomes the shortcomings of slow model training and insufficient optimization in the prior art, and ensures the comprehensive performance of the model in classification and regression tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings;

[0077] Figure 1 It is a schematic structural diagram of a brain image analysis system based on multi-modal fusion adopted in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0078] In order to make the objectives, technical solutions and advantages of the present invention clearer, the following will further describe the present invention in detail with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0079] Embodiment 1, a brain image analysis method based on multi-modal fusion, includes the following steps:

[0080] Step S1, data acquisition: Obtain image data containing multi-modal brain images through medical imaging equipment or image databases, and the image data includes structural images (MRI), functional images (fMRI) and metabolic images (PET);

[0081] The data is sourced from a standardized medical imaging database (ADNI database) or a brain imaging dataset provided by a medical institution.

[0082] The data types include structural images (MRI, T1-weighted and T2-weighted images), metabolic images (PET), and functional images (fMRI).

[0083] Data acquisition process:

[0084] MRI data acquisition: A 3T magnetic resonance imaging scanner is used to acquire T1-weighted images for describing brain anatomy. The acquisition parameters include: resolution: 1mm 3 Voxel size; scan time: 5 minutes; scan sequence: fast gradient echo (MPRAGE).

[0085] PET data acquisition: A positron emission tomography scanner (such as GE-Discovery-710) is used to measure brain metabolic activity by injecting a radioactive tracer (FDG). The acquisition parameters include: resolution: 3mm 3 Voxel size; data format: SUV (standard uptake value).

[0086] fMRI data acquisition: Functional magnetic resonance imaging technology is adopted to reflect the dynamic functional activities of brain regions through blood oxygenation level-dependent (BOLD) signals. The acquisition parameters include: temporal resolution: 2 seconds / frame; total scan time: 6 minutes; spatial resolution: 3mm 3 Voxel size.

[0087] The collected data is uniformly stored in DICOM format, containing complete scan parameters and patient information; before data processing, the DICOM data is desensitized to ensure patient privacy and security; it is converted to an analysis-friendly NIfTI format for subsequent preprocessing and feature extraction.

[0088] Check whether there are artifacts, motion artifacts, or incomplete acquisitions in the collected image data; re-acquire or exclude the data that does not meet the quality standards.

[0089] Step S2, data preprocessing: Denoise, intensity normalize, and perform inter-modal registration on the brain imaging data. The inter-modal registration is implemented using an elastic registration algorithm, and brain regions are extracted based on a template-based segmentation method;

[0090] The denoising, intensity normalization, and inter-modal registration of the brain imaging data include:

[0091] Denoising processing: The non-local means denoising algorithm is adopted. For each pixel point in the image, the gray-scale similarity between it and the surrounding pixel blocks is calculated, and the pixel values are weighted and averaged based on the similarity weights to remove high-frequency noise while retaining texture features and edge information;

[0092] Intensity normalization: Linear normalization. For structural image data, the maximum and minimum values are calculated, and all pixel values are linearly mapped to normalize the corresponding pixel intensities to the range of [0,1] to eliminate the influence of equipment or scanning parameters during the data acquisition process; for functional images, based on the z-score normalization method, the mean and standard deviation of the whole-brain functional signals are calculated, and the signal values corresponding to each pixel are standardized to a distribution with a mean of 0 and a standard deviation of 1; Piecewise normalization. For the metabolic hot-spot regions of metabolic images, the piecewise linear normalization method is adopted to map the high-dynamic-range hot-spot intensity values to the interval of 0.8-1 respectively, and the low-intensity background values are mapped to the interval of 0-0.2 to retain the metabolic characteristics;

[0093] Inter-modal registration: Preliminary registration. Taking the structural image as the reference modality, a rigid registration algorithm based on mutual information is used to preliminarily align the functional image and the metabolic image by optimizing the rotation and translation parameters; Non-linear registration. On the basis of rigid registration, an elastic registration algorithm based on a multi-resolution strategy is adopted to optimize the detailed differences layer by layer; Resampling and standardization. The registered images are resampled using the cubic B-spline interpolation method, and the registered image data is interpolated to a unified resolution, and all modal images are aligned to the standard anatomical template to ensure that all modal images have consistent spatial coordinates and resolutions;

[0094] The elastic configuration algorithm includes: In the low-resolution stage, the image data is downsampled, the coarse-grained global deformation field is calculated, and the overall structure alignment between modalities is adjusted; In the high-resolution stage, at a fine-grained resolution, the deformation field is locally optimized to correct the minor structural differences between modalities.

[0095] The brain regions extracted by the template-based segmentation method include:

[0096] Select a standard anatomical template containing information on the anatomical regions of the brain. The template is the MNI152 template or the ICBM152 template; perform spatial normalization on the brain images, and align the brain images with the template through an elastic registration algorithm so that the two are in the same anatomical space; based on the probability atlas in the template, perform pixel-level probability calculation on the aligned image data to obtain the probability values of each pixel belonging to different brain regions; apply a segmentation method based on the probability atlas, combined with a preset probability threshold, to segment the brain regions in the images and identify the gray matter, white matter, and cerebrospinal fluid regions; use a region growing algorithm to optimize the boundaries of the segmentation results, remove noise regions, and refine the structural boundaries; output a segmentation mask file, and the mask file annotates different brain regions and their corresponding anatomical structures.

[0097] Step S3, feature extraction and fusion: Extract multi-modal features from the preprocessed image data, including texture features, morphological features, and metabolic features; the texture features are extracted by the gray-level co-occurrence matrix method, the morphological features are used to extract the volume information of key brain structures from the structural images through a segmentation algorithm, and the metabolic features are obtained by statistically calculating the intensity indexes of metabolic hot regions in the metabolic images; perform standardization processing on the extracted multi-modal features, assign modal feature weights based on the entropy weight method, and fuse the multi-modal features through the principal component analysis method;

[0098] The feature extraction includes:

[0099] Texture feature extraction: For the preprocessed structural images, calculate the contrast, entropy, and homogeneity features using the gray-level co-occurrence matrix method; adopt a sliding window technique to extract local texture features on the two-dimensional slices of the structural images, and statistically summarize the feature values of all slices to generate a three-dimensional texture feature matrix;

[0100] Morphological feature extraction: Based on the segmentation results of the brain regions, calculate the volume features of the gray matter, white matter, and cerebrospinal fluid regions of the structural images; extract the shape features of the segmented regions, including surface area, compactness, and aspect ratio, to generate the morphological feature vectors corresponding to the regions;

[0101] Metabolic feature extraction: For the metabolic hot regions in the metabolic images, calculate the SUVmax, SUVmean, and standard deviation features; generate a metabolic intensity matrix for specific brain regions based on the distribution of metabolic hot regions in different intervals and normalize it into a standardized metabolic feature vector;

[0102] Group all the extracted features according to the modality and store them in matrix form, where each row represents the feature vector of a segmented region or brain region, and each column represents different feature types for subsequent feature fusion.

[0103] The fusion of the multi-modal features includes:

[0104] Feature standardization: Standardize the feature vectors extracted from different modality images using the z-score normalization method to transform each feature into a distribution with a mean of 0 and a standard deviation of 1.

[0105] Feature weight assignment: Calculate the importance weights of different modality features based on the entropy weight method, including constructing an information entropy model for each modality feature to evaluate the difference of feature data; assigning corresponding weights to each modality according to the information entropy value, and the weight values are used to adjust the contribution ratio of modality features in the fusion process.

[0106] Feature concatenation: Concatenate texture features, morphological features, and metabolic features in the column direction according to the feature type and modality to generate a unified high-dimensional feature matrix.

[0107] Dimensionality reduction processing: Use the principal component analysis method to perform dimensionality reduction on the high-dimensional feature matrix, extract the main components with a cumulative contribution rate exceeding 95%, and generate a low-dimensional fusion feature vector.

[0108] Store the dimensionality-reduced fusion feature vectors as a two-dimensional matrix according to patient individuals, where each row represents the features of a patient and each column represents different types of fusion features for subsequent comprehensive analysis.

[0109] Step S4, Brain network analysis: Based on functional images and structural images, construct a brain functional connection network and a brain structural connection network. Among them, the brain functional connection network is obtained by calculating the correlation of time series of different brain regions; the brain structural connection network is obtained by generating a white matter connection matrix through fiber tracking technology; extract the graph features of the brain functional connection network and the brain structural connection network, including node degree, clustering coefficient, and characteristic path length.

[0110] The brain network analysis includes:

[0111] Construction of the brain functional connection network: Based on functional images, extract time series for the whole brain partition; calculate the time series correlation between every two brain regions to construct a functional connection matrix, where the elements of the matrix represent the correlation coefficients between brain regions.

[0112] Construction of the brain structural connection network: Based on structural images, use fiber tracking technology to generate a brain white matter fiber connection map; construct a structural connection matrix, where the elements of the matrix represent the number of fibers or fiber density between brain regions.

[0113] Extraction of brain network features: Extract graph features from the brain functional connection network and the brain structural connection network, including node degree: representing the number of connections of each brain region with other brain regions; clustering coefficient: representing the tightness of the local network of each brain region; characteristic path length: representing the average length of the shortest path between any two nodes in the network.

[0114] Store the graph features of the brain functional connection network and the brain structural connection network in vector format, where each row represents the features of a brain region and each column represents different types of graph features for subsequent comprehensive analysis.

[0115] Step S5, Comprehensive analysis: Input the fused multi-modal features and graph features into the comprehensive analysis model to output the classification of brain diseases and related evaluation results; the comprehensive analysis model adopts a dual-branch structure and realizes the deep fusion of the two types of features through a cross-modal attention mechanism.

[0116] The structure of the comprehensive analysis model includes:

[0117] Input layer: Used to receive multi-modal fusion features and graph features. Among them, the multi-modal fusion features are a two-dimensional matrix, with rows representing each sample and columns representing multi-modal features; the graph features are graph structure data, including a functional connection matrix and a structural connection matrix.

[0118] Modal feature processing branch:

[0119] For the multi-modal fusion features, use MLP to process different modal features respectively. Each modal MLP consists of two fully connected layers. The first layer uses the ReLU activation function, and the second layer outputs the modal feature embeddings after Batch-ormalization normalization; splice the modal embedding features and input them into the feature interaction module.

[0120] For the brain network graph features, process the functional connection matrix and the structural connection matrix through a graph neural network to extract the local and global features of the graph structure; use a graph convolutional layer in the graph neural network to aggregate node information layer by layer and calculate the high-dimensional feature embeddings of the nodes; fuse the embedding features of the functional graph and the structural graph into a unified graph feature vector.

[0121] Feature interaction module: For the modal features and graph features, adopt a cross-modal attention mechanism to construct query, key, and value matrices for the modal features and graph features respectively; calculate the attention weight matrix between the modal features and the graph features, and generate an interaction-enhanced feature representation through a weighting mechanism.

[0122] Feature fusion module: Input the interaction-enhanced feature representation into a multi-layer stacked fully connected network to extract the comprehensive features related to the disease layer by layer; each layer of the fully connected network uses the ReLU activation function and adds a Dropout mechanism to avoid overfitting.

[0123] Task branch module: Classification branch, classify brain diseases for the comprehensive features through a Softmax classification layer; regression branch, calculate the lesion size, disease severity, or progression prediction score through a fully connected layer.

[0124] Data processing flow: After the multi-modal fusion features and brain network graph features are processed through branches respectively, embedded feature vectors are generated; the two types of features are fused in the feature interaction module to capture the interaction relationship between the modality and graph features; the comprehensive features are further dimension-reduced and extracted through the fusion module, and finally the classification results and relevant evaluation indicators are output.

[0125] The training of the comprehensive analysis model includes:

[0126] Collect multi-modal imaging data and brain network data, and extract multi-modal fusion features and brain network graph features as inputs; divide the data set into a training set, a validation set and a test set, with a ratio of 8:1:1;

[0127] Define the objective function of the classification task as the cross-entropy loss function, and the calculation formula is:

[0128] where N is the number of samples, K is the number of categories, y ik is the true label of sample i, is the probability value predicted by the model;

[0129] Define the objective function of the regression task as the mean square error loss function, and the calculation formula is: where z i and are the true value and the value predicted by the model respectively;

[0130] The comprehensive loss function is the multi-task weighted loss: L = αL cls + βL reg , where α and β are the loss weight coefficients of the classification and regression tasks;

[0131] Perform model training, initialize the model parameters, and use the stochastic gradient descent method for optimization; set the learning rate and update it using the cosine annealing learning rate adjustment strategy, and the formula is: where η t represents the learning rate of the t-th iteration, T is the total number of training rounds, η min and η max are the minimum and maximum values of the learning rate during training respectively;

[0132] In each round of training iteration, input the training set data into the model, calculate the prediction results of the classification and regression tasks; calculate the loss function; calculate the gradient through the backpropagation algorithm; finally update the model parameters: where θ represents the parameters of the model;

[0133] After each training round ends, use the validation set to evaluate the model performance, record the classification accuracy, recall rate and regression mean square error; adjust the hyperparameters according to the validation results;

[0134] Set an early stopping mechanism to terminate the training when the validation loss does not decrease for several consecutive rounds; save the model parameters with the best performance on the validation set as the final trained model;

[0135] Use the test set to conduct a final evaluation of the trained model, calculate the accuracy, AUC value for the classification task, and the mean squared error for the regression task, and output the evaluation results to verify the performance of the model in practical applications.

[0136] Step S6, lesion detection: Based on the comprehensive analysis results, use the region growing algorithm to detect and segment the lesion area, quantify the volume, morphology, and metabolic characteristics of the lesion, and generate an auxiliary diagnosis report.

[0137] The lesion detection includes:

[0138] Receive the segmentation mask data of the gray matter, white matter, and cerebrospinal fluid regions extracted by the template segmentation method; combine the classification results of the comprehensive analysis model to determine the target regions with high disease risks and generate the initial mask of the target regions;

[0139] In the classification results of the comprehensive analysis model, identify the initial seed points within the target regions by locating the high-risk regions that contribute the most to the classification results (represented by the heat map generated by Grad-CAM);

[0140] The selection of seed points is based on the voxels with significantly abnormal multimodal features, specifically including the statistical threshold conditions of texture, morphology, or metabolic features;

[0141] Starting from the seed points, use the region growing algorithm to expand the lesion area. Specifically, check whether the eigenvalue of the neighboring voxels around the seed points meets the expansion conditions, including intensity value proximity (the difference from the seed point is less than the preset threshold) and spatial connectivity; the voxels that meet the conditions are included in the lesion area, and the newly added voxels are marked as new seed points and continue to expand; repeat the above steps until the lesion area stops expanding;

[0142] For the finally expanded lesion area, calculate its morphological features, including volume, surface area, aspect ratio, and compactness; for the metabolic image, calculate the metabolic feature values of the lesion area, including SUVmax, SUVmean, and the standard deviation of metabolic intensity; superimpose the lesion area on the original image to generate a three-dimensional visualized lesion image; output an auxiliary diagnosis report containing the lesion location, volume, metabolic features, and morphological features for further clinical diagnosis and evaluation.

[0143] Embodiment 2, a brain image analysis system based on multimodal fusion, as shown in Figure 1 shown, includes the following modules:

[0144] Data acquisition module: used to acquire image data containing multi-modal brain images through medical imaging devices or image databases;

[0145] Data preprocessing module: used to denoise, perform intensity normalization and inter-modal registration on the brain image data, and extract brain regions based on a template-based segmentation method;

[0146] Feature extraction and fusion module: used to extract multi-modal features from the preprocessed image data, perform standardization processing on the extracted multi-modal features, and fuse the multi-modal features through principal component analysis;

[0147] Brain network analysis module: used to construct a brain functional connectivity network and a brain structural connectivity network based on functional images and structural images, and extract graph features of the brain functional connectivity network and the brain structural connectivity network;

[0148] Comprehensive analysis module: used to input the fused multi-modal features and graph features into a comprehensive analysis model, and output brain disease classification and related evaluation results;

[0149] Lesion detection module: used to detect and segment lesion regions based on the comprehensive analysis results, quantify the volume, morphology and metabolic characteristics of the lesions, and generate an auxiliary diagnosis report.

[0150] The specific embodiments described above further elaborate on the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A brain image analysis method based on multimodal fusion, characterized in that: The following steps are involved: Step S1, data acquisition: acquiring image data including multimodal brain images through medical imaging equipment or an image database, wherein the image data includes structural images, functional images, and metabolic images; Step S2, data preprocessing: performing denoising, intensity normalization and inter-modality registration on the brain image data, wherein the inter-modality registration is implemented by an elastic registration algorithm, and the brain region is extracted by a template-based segmentation method; Step S3, feature extraction and fusion: extracting multimodal features from the preprocessed image data, including texture features, morphological features and metabolic features; The texture features are extracted by the gray-level co-occurrence matrix method, the morphological features are extracted from the structural images by the segmentation algorithm to extract the volume information of the key brain structures, and the metabolic features are obtained by statistically analyzing the intensity index of the metabolic hotspot areas in the metabolic images; the extracted multimodal features are standardized, the modal feature weights are assigned based on the entropy weight method, and the multimodal features are fused by the principal component analysis method; Step S4, brain network analysis: constructing a brain functional connection network and a brain structural connection network based on functional images and structural images, wherein the brain functional connection network is obtained by calculating the correlation of time series of different brain regions; the brain structural connection network is obtained by generating a white matter connection matrix through fiber tracking technology; extracting graph features of the brain functional connection network and the brain structural connection network, including node degree, clustering coefficient and characteristic path length; Step S5, comprehensive analysis: input the fused multimodal features and graph features into a comprehensive analysis model, and output brain disease classification and related evaluation results; the comprehensive analysis model adopts a dual-branch structure and realizes deep fusion of two types of features through a cross-modal attention mechanism; Step S6, lesion detection: Based on the comprehensive analysis results, the lesion area is detected and segmented using a region growing algorithm to quantify the volume, morphology and metabolic characteristics of the lesion, and an auxiliary diagnosis report is generated.

2. A brain image analysis method based on multimodal fusion according to claim 1, characterized in that: The brain image data is subjected to denoising, intensity normalization and inter-modality registration, including: Denoising: Using the non-local mean denoising algorithm, for each pixel of the image, the grayscale similarity between it and the surrounding pixel blocks is calculated, and the pixel values ​​are weighted averaged based on the similarity weight to remove high-frequency noise while retaining texture features and edge information; Intensity normalization: Linear normalization: for structural imaging data, the maximum and minimum values ​​are calculated, and all pixel values ​​are linearly mapped to normalize the corresponding pixel intensity to the range of [0,1] to eliminate the influence of equipment or scanning parameters during data acquisition. For functional imaging, based on the z-score normalization method, the mean and standard deviation of the whole brain functional signal are calculated, and the signal value corresponding to each pixel is standardized to a distribution with a mean of 0 and a standard deviation of 1. For metabolic hotspot areas of metabolic imaging, piecewise linear normalization is used to map the hotspot intensity values ​​in the high dynamic range to the range of 0.8-1, and the low-intensity background values ​​to the range of 0-0.2 to retain metabolic characteristics. Inter-modality registration: preliminary registration, using structural images as the reference modality, using a rigid registration algorithm based on mutual information, and performing preliminary alignment of functional images and metabolic images by optimizing rotation and translation parameters; nonlinear registration, based on rigid registration, using an elastic registration algorithm based on a multi-resolution strategy to optimize detail differences layer by layer; resampling and standardization, using cubic B-spline interpolation to resample the registered images, interpolating the registered image data to a uniform resolution, aligning all modality images to a standard anatomical template to ensure that all modality images have consistent spatial coordinates and resolution; The elastic configuration algorithm includes: in the low-resolution stage, the image data is downsampled, the coarse-grained global deformation field is calculated, and the alignment of the overall structure between the modalities is adjusted; in the high-resolution stage, the deformation field is locally optimized at a fine-grained resolution to correct the tiny structural differences between the modalities.

3. The brain image analysis method based on multimodal fusion according to claim 1, characterized in that: The template-based segmentation method for extracting brain regions includes: A standard anatomical template containing brain anatomical structure partition information is selected, wherein the template is an MNI152 template or an ICBM152 template; the brain image is spatially normalized, and the brain image is aligned with the template through an elastic registration algorithm so that the two are located in the same anatomical space; based on the probability map in the template, the aligned image data is subjected to pixel-level probability calculation to obtain the probability value of each pixel belonging to a different brain region; a segmentation method based on the probability map is applied, combined with a preset probability threshold, to segment the brain region in the image, and identify the gray matter, white matter and cerebrospinal fluid regions; a region growing algorithm is used to optimize the boundary of the segmentation result, remove the noise region and refine the structure boundary; and a segmentation mask file is output, wherein the mask file labels different brain regions and their corresponding anatomical structures.

4. The brain image analysis method based on multimodal fusion according to claim 1, characterized in that: The feature extraction comprises: Texture feature extraction: For the preprocessed structural images, the gray-level co-occurrence matrix method is used to calculate the contrast, entropy and homogeneity features; the sliding window technology is used to extract local texture features on the two-dimensional slices of the structural images, and the eigenvalues ​​of all slices are statistically summarized to generate a three-dimensional texture feature matrix; Morphological feature extraction: Based on the segmentation results of the brain region, the volume features of the gray matter, white matter and cerebrospinal fluid regions of the structural image are calculated; the shape features of the segmented region, including surface area, compactness and aspect ratio, are extracted to generate the morphological feature vector corresponding to the region; Metabolic feature extraction: Calculate SUVmax, SUVmean and standard deviation features for metabolic hotspots in metabolic images; Generate a metabolic intensity matrix for a specific brain region based on the distribution of metabolic hotspots in different intervals and normalize it into a standardized metabolic feature vector; All extracted features are grouped according to modality and stored in matrix form, where each row represents a feature vector of a segmented region or brain area and each column represents a different feature type for subsequent feature fusion.

5. The brain image analysis method based on multimodal fusion according to claim 1, characterized in that: The fusing of multimodal features includes: Feature standardization: The feature vectors extracted from different modal images are standardized, and the z-score normalization method is used to convert each feature into a distribution with a mean of 0 and a standard deviation of 1; Feature weight allocation: Calculate the importance weights of different modal features based on the entropy weight method, including building an information entropy model for each modal feature and evaluating the differences in feature data; assign corresponding weights to each modality based on the information entropy value, and the weight value is used to adjust the contribution ratio of the modal feature in the fusion process; Feature splicing: According to the feature type and mode, texture features, morphological features and metabolic features are spliced ​​in the column direction to generate a unified high-dimensional feature matrix; Dimensionality reduction: Use the principal component analysis method to reduce the dimension of the high-dimensional feature matrix, extract the main components with a cumulative contribution rate of more than 95%, and generate a low-dimensional fusion feature vector; The fused feature vectors after dimensionality reduction are stored as a two-dimensional matrix according to individual patients, with each row representing the feature of a patient and each column representing a different type of fused features for subsequent comprehensive analysis.

6. The brain image analysis method based on multimodal fusion according to claim 1, characterized in that: The brain network analysis includes: Construction of brain functional connection network: Based on functional imaging, time series extraction of whole brain partitions is performed; the time series correlation between every two brain regions is calculated, and a functional connection matrix is ​​constructed, in which the elements of the matrix represent the correlation coefficients between brain regions; Construction of brain structural connection network: Based on structural imaging, fiber tracking technology is used to generate brain white matter fiber connection map; a structural connection matrix is ​​constructed, in which the elements of the matrix represent the number of fibers or fiber density between brain regions; Extraction of brain network features: Extract graph features from brain functional connection networks and brain structural connection networks, including: node degree: indicates the number of connections between each brain region and other brain regions; clustering coefficient: indicates the compactness of the local network of each brain region; characteristic path length: indicates the average length of the shortest path between any two nodes in the network; The graph features of the brain functional connection network and the brain structural connection network are stored in vector format, with each row representing the features of a brain region and each column representing a different graph feature type for subsequent comprehensive analysis.

7. The brain image analysis method based on multimodal fusion according to claim 1, characterized in that: The structure of the comprehensive analysis model includes: Input layer: used to receive multimodal fusion features and graph features, where the multimodal fusion features are two-dimensional matrices, with rows representing each sample and columns representing multimodal features; graph features are graph structure data, including functional connection matrices and structural connection matrices; Modal feature processing branch: For multimodal fusion features, MLP is used to process different modal features separately. The MLP of each modality consists of two fully connected layers. The first layer uses the ReLU activation function, and the second layer outputs the modal feature embedding after Batch-ormalization standardization. The embedded features of each modality are spliced ​​and input into the feature interaction module. Aiming at the features of brain network graphs, the functional connection matrix and structural connection matrix are processed by graph neural network to extract local and global features of graph structure. The graph convolution layer is used in the graph neural network to aggregate node information layer by layer and calculate the high-dimensional feature embedding of nodes. The embedded features of functional graph and structural graph are fused into a unified graph feature vector. Feature interaction module: For modal features and graph features, a cross-modal attention mechanism is used to construct query, key, and value matrices for modal features and graph features respectively; the attention weight matrix between modal features and graph features is calculated, and interactively enhanced feature representations are generated through a weighted mechanism; Feature fusion module: The interactively enhanced feature representation is input into a multi-layer stacked fully connected network to extract comprehensive features related to the disease layer by layer. Each layer of the fully connected network uses the ReLU activation function and adds the Dropout mechanism to avoid overfitting. Task branch module: Classification branch, which classifies brain diseases based on comprehensive features through the Softmax classification layer; regression branch, which calculates lesion size, disease severity or progression prediction score through the fully connected layer; Data processing flow: After the multimodal fusion features and brain network graph features are processed by branches respectively, embedded feature vectors are generated; the two types of features are fused in the feature interaction module to capture the interactive relationship between the modality and graph features; the comprehensive features are further reduced in dimension and extracted through the fusion module, and finally the classification results and related evaluation indicators are output.

8. The brain image analysis method based on multimodal fusion according to claim 7, characterized in that: The training of the comprehensive analysis model includes: Collect multimodal imaging data and brain network data, extract multimodal fusion features and brain network map features as input; divide the data set into training set, validation set and test set with a ratio of 8:1:1; The objective function of the classification task is defined as the cross entropy loss function, and the calculation formula is: Where N is the number of samples, K is the number of categories, and y ik is the true label of sample i, is the probability value predicted by the model; The objective function of the regression task is defined as the mean square error loss function, and the calculation formula is: where z i and are the true value and the model predicted value respectively; The comprehensive loss function is a multi-task weighted loss: L = αL cls +βL reg , where α and β are the loss weight coefficients for classification and regression tasks; Perform model training, initialize model parameters, and use stochastic gradient descent method for optimization; set the learning rate, and use cosine annealing learning rate adjustment strategy update. The formula is: where η t represents the learning rate of the tth iteration, T is the total number of training rounds, η min and η max They are the minimum and maximum values ​​of the learning rate during the training process; In each training iteration, the training set data is input into the model, the prediction results of the classification and regression tasks are calculated; the loss function is calculated; the gradient is calculated through the backpropagation algorithm; and finally the model parameters are updated: Where θ represents the parameters of the model; After each training round, use the validation set to evaluate the model performance, record the classification accuracy, recall rate and regression mean square error; adjust the hyperparameters based on the validation results; Set up an early stopping mechanism to terminate training when the validation loss does not decrease within several consecutive rounds; save the model parameters with the best performance on the validation set as the final training model; Use the test set to perform a final evaluation of the trained model, calculate the accuracy and AUC value of the classification task, and the mean square error of the regression task, and output the evaluation results to verify the performance of the model in practical applications.

9. The brain image analysis method based on multimodal fusion according to claim 1, characterized in that: The lesion detection includes: Receive segmentation mask data of gray matter, white matter and cerebrospinal fluid regions extracted based on the template segmentation method; determine the target region with high disease risk in combination with the classification results of the comprehensive analysis model, and generate an initial mask of the target region; In the classification results of the comprehensive analysis model, the initial seed points within the target area are identified by locating the high-risk areas that contribute most to the classification results; The selection of seed points is based on voxels with significant multimodal features, including statistical threshold conditions for texture, morphology, or metabolic features; Taking the seed point as the starting point, the lesion area is expanded using the region growing algorithm, which specifically includes checking whether the characteristic values ​​of the neighborhood voxels around the seed point meet the expansion conditions, including the proximity of intensity values ​​and spatial connectivity; the voxels that meet the conditions are included in the lesion area, and the newly added voxels are marked as new seed points to continue expanding; the above steps are repeated until the lesion area stops expanding; For the final expanded lesion area, its morphological characteristics are calculated, including volume, surface area, aspect ratio and compactness; for metabolic images, the metabolic characteristic values ​​of the lesion area are calculated, including SUVmax, SUVmean and standard deviation of metabolic intensity; the lesion area is superimposed on the original image to generate a three-dimensional visual lesion image; an auxiliary diagnosis report containing the lesion location, volume, metabolic characteristics and morphological characteristics is output for further clinical diagnosis and evaluation.

10. A brain image analysis system based on multimodal fusion, characterized in that: The system applies a brain image analysis method based on multimodal fusion according to any one of claims 1 to 9, and comprises the following modules: Data acquisition module: used to acquire image data including multimodal brain images through medical imaging equipment or image database; Data preprocessing module: used for denoising, intensity normalization and inter-modality registration of the brain image data, and extracting brain regions based on a template-based segmentation method; Feature extraction and fusion module: used to extract multimodal features from preprocessed image data, standardize the extracted multimodal features, and fuse the multimodal features through principal component analysis method; Brain network analysis module: used to construct brain functional connection network and brain structural connection network based on functional images and structural images, and extract graph features of brain functional connection network and brain structural connection network; Comprehensive analysis module: used to input the fused multimodal features and graph features into the comprehensive analysis model, and output brain disease classification and related evaluation results; Lesion detection module: Based on the comprehensive analysis results, it uses the region growing algorithm to detect and segment the lesion area, quantify the volume, morphology and metabolic characteristics of the lesion, and generate an auxiliary diagnosis report.

Images