A method and system for identifying brain tumors in brain magnetic resonance images

Abstract

This invention discloses a method and system for identifying brain tumors in brain MRI images. The method includes the following steps: S1, acquiring the original brain MRI image data of the object to be identified and performing preprocessing operations to obtain standardized multimodal MRI image data; S2, performing preliminary tumor region screening on the preprocessed image data using a tumor tissue probability decoupling identification method based on multi-echo relaxation differences. This invention relates to the field of medical image processing technology. This method and system for identifying brain tumors in brain MRI images employs a tumor tissue probability decoupling model based on multi-echo relaxation differences and cross-scale brain network topology perturbation analysis, which can accurately screen potential tumor regions. Furthermore, it improves identification accuracy through dual constraints of metabolism and morphology. The introduction of cross-scale topology analysis, combined with abnormal propagation detection of brain network structures, effectively identifies suspected tumor regions and improves identification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing, more particularly, to a brain tumor recognition method and system in brain magnetic resonance images. BACKGROUND

[0002] Early detection and accurate positioning of brain tumors are crucial for treatment. Traditional brain tumor detection methods mainly rely on single MRI images or CT scan images, which often have problems such as blurred tumor regions, difficult to distinguish, and cannot provide sufficient tumor feature information. Therefore, how to improve the accuracy of tumor recognition through multi-modal MRI images combined with advanced image processing and analysis techniques has become a research hotspot in the fields of medical imaging and tumor diagnosis.

[0003] In recent years, with the continuous development of medical imaging technology, multi-modal image analysis methods based on multi-echo MRI sequences, diffusion-weighted MRI and structural MRI have been gradually proposed. However, single-mode image processing methods often face limitations and are difficult to fully capture the relaxation characteristics, metabolic abnormalities and morphological features of tumors. Therefore, how to integrate multiple image data, use cross-scale analysis techniques, and combine metabolic and morphological information for comprehensive discrimination has become the key to improving brain tumor recognition results. SUMMARY

[0004] The purpose of the present application is to provide a brain tumor recognition method and system in brain magnetic resonance images, which solves the problem that single-mode image processing methods often face limitations and are difficult to fully capture the relaxation characteristics, metabolic abnormalities and morphological features of tumors, and cannot meet the use requirements.

[0005] The present application achieves the above-mentioned purpose through the following technical solutions: a brain tumor recognition method in brain magnetic resonance images, the method comprising the following steps: S1, obtaining brain magnetic resonance raw image data of an object to be recognized and performing a preprocessing operation to obtain standardized multi-modal MRI image data; S2, using a tumor tissue probability decoupling recognition method based on multi-echo relaxation differences to preliminarily screen the tumor regions of the preprocessed image data, and obtaining a set of potential tumor candidate regions through relaxation characteristic analysis and tissue signal decomposition; S3, using a tumor recognition method based on cross-scale brain network topology disturbance to verify the topological features of the set of potential tumor candidate regions, and combining cross-scale analysis of brain structural networks and abnormal propagation detection to screen out topologically abnormal tumor suspected regions; S4, using a brain tumor adaptive recognition method based on metabolic-morphological collaborative constraints to collaboratively discriminate and fuse the tumor suspected regions, and realizing accurate recognition and positioning of brain tumor regions through dual constraints of metabolic characteristics and morphological evolution.

[0006] Further, in the step S1, the acquired brain nuclear magnetic resonance raw image data includes: multi-echo MRI sequence data, structural MRI sequence data, diffusion weighted MRI sequence data; The preprocessing operation includes head region cropping, noise suppression, multi-modal and multi-echo image registration, and gray scale normalization in sequence, wherein the gray scale normalization maps the image pixel gray scale value to an interval, so as to realize the standardization processing of the multi-modal MRI image data.

[0007] Further, in the step S2, the tumor tissue probability decoupling identification method based on multi-echo relaxation difference includes: calculating the pixel point relaxation value based on the preprocessed multi-echo MRI sequence normalized data and constructing a relaxation difference response matrix, and calculating the relaxation gradient change field with the echo time as the dimension; constructing and training a tissue relaxation probability decoupling model, using the model to probabilistically decompose the mixed relaxation signal of the image pixel point, and obtaining the signal contribution weight of each type of tissue; determining a relaxation gradient threshold, combining the relaxation abnormal pixel labeling result and the tumor tissue signal contribution weight screening condition to obtain potential tumor pixels; performing region connectivity analysis and merging on the potential tumor pixels to form a potential tumor candidate region set.

[0008] Further, the tissue relaxation probability decoupling model is a Gaussian mixture model, the model mixture component number corresponds to four types of tissues including gray matter, white matter, cerebrospinal fluid and tumor tissue, the model is trained by EM algorithm, and after the training is completed, the mixed relaxation signal is probabilistically decomposed, and the signal contribution weight of each type of tissue in the pixel point mixed signal is output; The relaxation gradient threshold is adaptively adjusted by the percentile method combined with the brain image features of different age groups, and the tumor tissue signal contribution weight screening threshold is determined based on the maximum Youden index of the ROC curve analysis.

[0009] Further, in the step S3, the tumor identification method based on cross-scale brain network topology disturbance includes: constructing an undirected weighted brain structural connection network based on the preprocessed structural MRI sequence normalized data, taking the standardized brain region as the network node, and taking the structural similarity or spatial continuity between the brain regions as the connection edge weight; performing cross-scale topological feature analysis on the brain structural connection network in terms of local texture scale and brain region structure scale, and extracting core topological features; constructing and training a topological anomaly discrimination model, and calculating the topological feature deviation degree of each node in the brain network to be identified; The influence of the topological abnormal region on the adjacent brain area is calculated by using a graph propagation mechanism to obtain the deviation of the adjacent brain area after propagation; A topological abnormal threshold is determined, and the topological abnormal region is screened by combining the node topological feature deviation and the deviation of the adjacent brain area after propagation, and a tumor suspected region set is formed by integration.

[0010] Further, the cross-scale topological features include: The node degree, clustering coefficient of the local texture scale, and the characteristic path length, network efficiency of the brain region structure scale; The topological abnormality discrimination model is a random forest regression model, which is trained with the topological abnormality degree of the tumor region as the label, and is evaluated by 5-fold cross-validation; The graph propagation mechanism realizes abnormal information spatial propagation based on brain network connection weights, and the propagation weight is positively correlated with the connection weight between nodes and the source node topological feature deviation, and is normalized.

[0011] Further, in step S4, the brain tumor adaptive recognition method based on metabolic-morphological collaborative constraint includes: The apparent diffusion coefficient and fractional anisotropy of the pixel are calculated based on the preprocessed diffusion weighted MRI sequence normalized data, a tissue water migration model is constructed and trained, and the metabolic abnormality probability map is obtained by using the model; A tumor morphological evolution constraint model is constructed and trained, a tumor morphological diffusion potential function is defined, and a morphological diffusion potential value of the pixel is obtained; A metabolic-morphological collaborative discrimination model is constructed and trained, the metabolic abnormality probability feature and the morphological diffusion potential feature are weighted and fused to obtain a collaborative discrimination value of the pixel; A collaborative discrimination threshold is determined, and the tumor pixels are screened according to the threshold, and the spatial connectivity analysis, contour extraction and region merging are performed to realize the accurate recognition and positioning of the brain tumor region, and the spatial coordinates, contour boundary and area / volume information of the tumor region are output.

[0012] Further, the tissue water migration model is a fitting model based on multivariate linear regression, which is trained by gradient descent method combined with Adam optimizer, and the model output is mapped to The metabolic abnormality probability value in the interval by using the Sigmoid activation function; The tumor morphological diffusion potential function value decays with the increase of the distance from the pixel to the center of the tumor suspected region, and the diffusion scale parameter is adaptively adjusted according to the tumor morphological statistical results and the area of the tumor suspected region; The metabolic-morphological collaborative discrimination model optimizes the weight coefficients of the metabolic feature and the morphological feature by cross-validation method, the optimal value of the weight coefficient is determined based on the maximum value of the F1 score, and the collaborative discrimination threshold is determined by the maximum value of the Youden index of the ROC curve analysis.

[0013] Furthermore, it also includes: The steps for quantitative analysis of tumor characteristics in precisely identified brain tumor regions include extracting core morphological and functional characteristic indicators of the tumor. The morphological characteristic indicators include the two-dimensional area and three-dimensional volume of the tumor, and the functional characteristic indicators include the average relaxation difference value and the average metabolic abnormality probability value of the tumor region. All quantitative indicators provide objective quantitative basis for the graded diagnosis and disease assessment of brain tumors. The three-dimensional volume of the tumor is calculated based on the actual planar area of ​​the MRI image pixels, the scanning slice thickness, and the pixel set of the tumor region.

[0014] A brain tumor identification system in brain magnetic resonance imaging (MRI) images, applied to the above-described brain tumor identification method in brain MRI images, the system comprising: Brain MRI image preprocessing unit, multi-echo tumor preliminary screening unit, cross-scale tumor topology verification unit, and metabolic morphology-assisted tumor identification unit; The brain MRI image preprocessing unit is used to acquire the raw brain MRI image data of the object to be identified and perform preprocessing operations to obtain standardized multimodal MRI image data. The multi-echo tumor preliminary screening unit is used to perform preliminary tumor region screening on preprocessed image data based on the tumor tissue probability decoupling identification method based on multi-echo relaxation differences, and obtain a set of potential tumor candidate regions through relaxation characteristic analysis and tissue signal decomposition. The cross-scale tumor topology verification unit is used to verify the topological features of the potential tumor candidate region set by the tumor identification method based on cross-scale brain network topology perturbation, and to screen out the suspected tumor regions with topological anomalous by combining cross-scale analysis of brain structural networks and abnormal propagation detection. The metabolic morphology-coordinated tumor identification unit is used to perform collaborative discrimination and fusion of the suspected tumor region based on the adaptive brain tumor identification method of metabolic-morphological co-constraint, and to achieve accurate identification and localization of brain tumor region through the dual constraints of metabolic features and morphological evolution.

[0015] The beneficial effects of this invention are as follows: 1. By combining multiple imaging data such as multi-echo MRI, structural MRI, and diffusion-weighted MRI, the different characteristics of brain tumors can be comprehensively captured, improving the accuracy of tumor identification.

[0016] 2. By employing a probabilistic decoupling model of tumor tissue based on multi-echo relaxation differences and cross-scale brain network topological perturbation analysis, potential tumor regions can be accurately screened, and the identification accuracy can be further improved through dual constraints of metabolism and morphology.

[0017] 3. By introducing cross-scale topological analysis and combining it with the detection of abnormal propagation in brain network structures, we can effectively identify suspected tumor regions and overcome the shortcomings of traditional methods in accurately locating tumors.

[0018] 4. Through the metabolic-morphological co-constraint model, adaptive identification and precise localization of tumor regions can be achieved. The discrimination threshold can be automatically adjusted according to the metabolic and morphological evolution characteristics of the tumor, thereby improving the ability to identify different types of tumors.

[0019] 5. The present invention also includes quantitative analysis of the identified brain tumor regions, extracting morphological characteristics of the tumor, such as area and volume, and functional characteristics, such as relaxation differences and metabolic abnormality probability, to provide objective basis for clinical graded diagnosis and disease assessment. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a flowchart of the multi-echo relaxation recognition sub-process of the present invention; Figure 3 This is a flowchart of the metabolism-morphology collaborative recognition sub-process of the present invention; Figure 4 This is a system block diagram of the present invention. Detailed Implementation

[0021] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0022] Example 1: Please see Figures 1-3 This invention provides a technical solution: a method for identifying brain tumors in brain MRI images, the method comprising: S1. Acquire the raw brain MRI image data of the object to be identified and preprocess it to obtain standardized multimodal MRI image data; The subjects to be identified refer to individuals who need to undergo brain tumor identification examinations, and may be patients suspected of having brain tumors. Raw brain MRI image data refers to the original image data of brain structure directly acquired through MRI technology, containing signal information of different brain tissues under the influence of a magnetic field. However, this raw data may have issues such as noise and differences in parameters from different devices, requiring further processing. Preprocessing involves a series of operations on the raw image data to eliminate noise, correct geometric distortions, and standardize image size and resolution, making the image data more standardized and accurate for subsequent analysis. The goal is to obtain standardized multimodal MRI image data. Standardized multimodal MRI image data, after preprocessing, has a unified standard format and parameters, and integrates multiple imaging modes, such as T1-weighted images, T2-weighted images, and FLAIR images. Different modalities of images can reflect the characteristics of brain tissue from different angles, and the combination of multimodal data can provide richer information for tumor identification. S2. A probabilistic decoupling identification method for tumor tissue based on multi-echo relaxation differences is used to perform preliminary tumor region screening on preprocessed image data, and a set of potential tumor candidate regions is obtained through relaxation characteristic analysis and tissue signal decomposition. Among them, the multi-echo relaxation difference tumor tissue probabilistic decoupling identification method is a method for identifying tumor tissue based on multi-echo technology and relaxation characteristic differences in MRI. Multi-echo refers to acquiring signals with multiple different echo times during MRI scanning; relaxation is the process by which atomic nuclei recover from an excited state to an equilibrium state in MRI, and different tissues have different relaxation time characteristics; probabilistic decoupling is achieved by analyzing the relaxation characteristics of different tissues, separating the signals of tumor tissue from those of other normal tissues, and calculating the probability of tumor tissue occurrence, thereby achieving preliminary identification of tumor tissue; relaxation characteristic analysis and tissue signal decomposition analyze the relaxation characteristics of different tissues in multimodal MRI image data to understand the signal change patterns of various tissues, such as normal brain tissue and tumor tissue, during the relaxation process, and then decompose the mixed signal in the image into signal components corresponding to different tissues based on these patterns, so as to further identify tumor tissue; potential tumor candidate region set is a set of regions that may contain tumor tissue obtained after the above relaxation characteristic analysis and tissue signal decomposition. These regions are only the places where tumors may exist in the preliminary screening and need to be further verified and confirmed. S3. A tumor identification method based on cross-scale brain network topological perturbation verifies the topological features of potential tumor candidate regions and screens out suspected tumor regions with topological anomalous by combining cross-scale analysis of brain structural networks and abnormal propagation detection. Among them, the tumor identification method based on cross-scale brain network topological perturbation identifies tumors based on changes in brain network topology at different scales. The brain network is a complex network formed by neural connections between different regions of the brain. These different scales can be from microscopic (e.g., neuron level), mesoscopic (e.g., brain region level), to macroscopic (e.g., whole brain level). Topological perturbation refers to the disruption of the normal brain network topology caused by the presence of a tumor, leading to changes in network connection patterns and node characteristics. Tumors are identified by analyzing these cross-scale topological changes. The method also involves cross-scale analysis of brain structural networks, analyzing brain structural networks at multiple scales—microscopic, mesoscopic, and macroscopic—to study... Structural features of brain networks at different scales, such as the connection strength of nodes and brain regions, the clustering coefficient of the network, and path length, are used to comprehensively understand the organizational structure of brain networks. Anomaly propagation detection involves detecting the propagation of abnormal signals or connection changes caused by the presence of tumors in brain structural networks. Tumors may affect signal transduction and network connectivity in surrounding normal tissues; detecting this abnormal propagation pattern can help determine the location and extent of the tumor. Topologically anomalous tumor-suspected regions are identified through cross-scale brain network topological perturbation analysis, revealing regions with abnormal brain network topology that are suspected to be tumor locations. S4. A brain tumor adaptive identification method based on metabolic-morphological co-constraint performs co-discrimination and fusion on suspected tumor regions, and achieves accurate identification and localization of brain tumor regions through the dual constraints of metabolic features and morphological evolution. Among them, the metabolic-morphological co-constraint adaptive brain tumor identification method combines the metabolic and morphological evolution characteristics of brain tumors. Through the co-constraint of these two characteristics, it achieves more accurate brain tumor identification. Metabolic characteristics refer to the properties of tumor tissue that differ from normal tissue during metabolism, such as changes in the concentration of certain metabolites and abnormalities in metabolic pathways, which can be detected using techniques such as functional magnetic resonance imaging (fMRI). Morphological evolution refers to the changes in the morphology of the tumor during growth, such as changes in tumor size, shape, and boundaries, which can be observed using techniques such as structural magnetic resonance imaging (sMRI). "Adaptive" indicates that this method can identify brain tumors based on different tumor characteristics and individual... The system automatically adjusts its identification strategy based on body differences, improving the accuracy and adaptability of identification. The dual constraints of metabolic characteristics and morphological evolution consider both aspects when identifying brain tumor regions, combining and mutually restricting each other. Only regions that simultaneously meet the criteria of abnormal metabolic characteristics and abnormal morphological evolution are more likely to be identified as true tumor regions, thus improving identification accuracy. Precise identification and localization of brain tumor regions, through the metabolic-morphological co-constraint method, ultimately pinpoints the actual tumor location in the brain and accurately marks its position, size, and boundaries, providing crucial information for subsequent diagnosis and treatment.

[0023] It should be noted that, during the process, S1 acquires and preprocesses raw data to eliminate noise, standardize data, provide a high-quality foundation for subsequent analysis, and reduce interference factors. S2 performs preliminary screening based on multi-echo relaxation differences, utilizing the differences in relaxation characteristics of different tissues to quickly locate potential tumor areas, improve screening efficiency, and narrow the scope of subsequent analysis. S3 verifies through cross-scale brain network topological perturbation, analyzing changes in brain network structure and abnormal propagation at multiple scales, which can more accurately screen suspected tumor areas and avoid misjudgment. S4 adopts metabolic-morphological synergistic constraints, combining dual information of tumor metabolism and morphological characteristics to achieve accurate identification and localization, improving diagnostic accuracy. The overall method is progressive and multi-dimensional, giving full play to the advantages of different technologies, effectively improving the efficiency and accuracy of brain tumor identification, and providing a reliable basis for clinical diagnosis and treatment.

[0024] In one embodiment, acquiring raw MRI image data of the brain of the object to be identified and performing preprocessing includes: Acquire raw multimodal MRI images of the brain of the subject to be identified. The data should include at least multi-echo MRI sequence data, structural MRI sequence data, and diffusion-weighted MRI sequence data. The multi-echo MRI sequence data is denoted as:

[0025] in, For the number of valid echoes, For the first The two-dimensional nuclear magnetic resonance image matrix corresponding to the second echo has a matrix dimension that matches the image pixel resolution. ; The acquired raw image data undergoes a series of preprocessing operations. First, head region cropping is performed using a head mask to remove background and non-brain tissue areas. Then, Gaussian filtering is used for noise suppression to eliminate random noise interference during imaging. Next, a rigid body registration algorithm is used to register multimodal and multi-echo images, ensuring spatial consistency between different image sequences. Finally, the registered images are normalized to map pixel grayscale values ​​to... The gray-level normalization expression for the interval is:

[0026] in, For the first Spatial coordinates in the secondary echo image The original pixel grayscale value at that location, , The first The minimum and maximum gray values ​​of all pixels in the secondary echo image. This is the normalized grayscale value of the pixel. These are the two-dimensional spatial coordinates of the image pixels.

[0027] This design involves acquiring and preprocessing multimodal MRI raw image data. Multimodal data, including multi-echo, structural, and diffusion-weighted MRI sequences, provides more comprehensive information about the brain. Head masking removes irrelevant areas and reduces interference; Gaussian filtering suppresses noise and improves image quality; rigid body registration ensures spatial consistency across multiple image sequences, facilitating subsequent analysis; and grayscale normalization maps pixel values ​​to a unified range, eliminating grayscale differences between different images and making the data more comparable. This lays the foundation for accurate tumor identification and improves the stability and reliability of the entire identification process.

[0028] In one embodiment, a tumor tissue probability decoupling identification method based on multi-echo relaxation differences is used to perform preliminary tumor region screening on preprocessed image data to obtain a set of potential tumor candidate regions, including: Based on the preprocessed normalized data of multi-echo MRI sequences, the T2 relaxation value of each pixel in the image was calculated using a single exponential relaxation fitting algorithm. and T2 relaxation value A two-dimensional T2 relaxation difference response matrix is ​​constructed based on the relaxation value difference of each pixel. The spatial dimension of the matrix is ​​consistent with that of the preprocessed image, and each element in the matrix corresponds to the relaxation difference value of the image pixels, i.e.:

[0029] Using echo time as the dimension, calculate the relaxation gradient change field between adjacent echoes. The relaxation gradient change field is a three-dimensional data field, with dimensions consisting of the two-dimensional spatial dimension of the image pixels superimposed with the dimension of the echo sequence. For the th... and the The relaxation gradient of the next adjacent echo is calculated as follows:

[0030] in, The fixed echo time interval between two adjacent echoes is determined by the acquisition parameters of the magnetic resonance imaging (MRI) equipment. , The first , The relaxation difference response matrix corresponding to the second echo in spatial coordinates The element value at that position, For this coordinate, the position is at the th The relaxation gradient value of each echo interval; The specific steps for constructing and training an organizational relaxation probability decoupling model are as follows: Construct a training dataset and select no less than 200 labeled samples, including no less than 100 normal brain samples and no less than 100 samples of different types of brain tumors. Extract relaxation feature data of gray matter, white matter, cerebrospinal fluid and tumor tissue regions from all samples. The sample size for each tissue type is no less than 5,000 pixels. Initialize the Gaussian mixture model parameters, setting the mixture factor to 4, corresponding to the four tissue types. The initial mean vector is based on the relaxation characteristic mean of normal tissues, the initial covariance matrix is ​​the identity matrix, and the initial mixture coefficients are... ; The EM algorithm is used to train the model. The number of iterations is set to 100, and the convergence threshold is set to 1e-5. Each iteration includes E steps to calculate the posterior probability of the latent variables and M steps to maximize the likelihood function to update the model parameters. Training stops when the difference between the likelihood functions of two adjacent iterations is less than the convergence threshold or the maximum number of iterations is reached. Model validation involves dividing the training dataset into a training set and a validation set in an 8:2 ratio. If the model's classification accuracy on the validation set is not less than 90%, the model is considered to have completed training; otherwise, the sample size is increased and the model is retrained. The mixed relaxation signal of each pixel in the image is obtained by using the trained tissue relaxation probability decoupling model. Probability decomposition is performed, and the decomposition formula is:

[0031] in, The weights for gray matter, white matter, cerebrospinal fluid, and tumor tissue in the mixed signal at this pixel are respectively assigned, with weight values ​​ranging from [value range missing]. The weights sum to 1; The weight value is directly output by the fitting probability of the Gaussian mixture model to the relaxation features of the four types of tissues. The fitting process uses the relaxation feature data of normal brain tissue and labeled tumor tissue as training set, and solves the model parameters through the maximum likelihood estimation method to determine the signal contribution weight of each tissue. Determine the threshold for abnormal relaxation gradients , Determined using the percentile method, at least 100 normal brain multi-echo MRI data were selected, and the global value of their relaxation gradient change field was calculated. The relaxation gradient value corresponding to the quantile is used as If the data to be identified is a child's brain image, adjust it to... Quantities, if the image is of an elderly brain, should be adjusted to... quantiles; The relaxed gradient changing field satisfies Pixels with abnormal relaxation are marked as such. Then, combined with the tissue relaxation probability decoupling results, pixels with abnormal relaxation and significant signal contribution from tumor tissue are selected. The pixels are used as potential tumor pixels, and the contribution of tumor tissue signals is weighted by a threshold. To determine the threshold based on ROC curve analysis, tumor samples were identified as positive examples and normal brain tissue as negative examples. The weight value corresponding to the maximum value of the Youden index was selected as the threshold. By performing region connectivity analysis and merging on all spatially contiguous potential tumor pixels, a set of potential tumor candidate regions is obtained. This region set contains the spatial extent and pixel set of all suspected tumor regions identified in the initial screening.

[0032] This design, based on multi-echo relaxation differences, performs preliminary screening. Utilizing the differences in relaxation characteristics among different tissues, it accurately captures potential tumor signal changes by constructing a response matrix and gradient change field. A tissue relaxation probability decoupling model is built and trained, supported by a large number of samples, to accurately decompose the signal contribution weights of individual pixels and determine the thresholds for abnormal relaxation gradients and tissue signal contribution weights. This effectively filters out potential tumor pixels. After merging through regional connectivity analysis, a candidate region set is obtained. This method extracts key information from multi-echo data, improving the accuracy and efficiency of preliminary screening and reducing the workload of subsequent analysis.

[0033] In one embodiment, a tumor identification method based on cross-scale brain network topological perturbation verifies the topological features of a set of potential tumor candidate regions and filters out suspected tumor regions with topological anomalies, including: Based on the preprocessed normalized structural MRI sequence data, an automatic brain atlas segmentation algorithm was used to perform standardized brain region segmentation. Each segmented independent brain region was considered as a node in a brain structural connectivity network, and the node set was denoted as:

[0034] in, To standardize the number of brain regions, an undirected weighted brain structural connectivity network atlas is constructed by using the structural similarity or spatial continuity between different brain regions as the weights of the connection edges:

[0035] Among them, edge set Middle elements Represents a node With nodes The connection weights and structural similarity between brain regions are calculated using the gray-level correlation between brain regions, and the formula is as follows:

[0036] in, For preprocessed structural MRI images, normalized data, brain region Spatial coordinates of the inner pixel, brain region Spatial coordinates of the inner pixel; Cross-scale topological feature analysis was performed on the constructed brain structural connectivity network, extracting core topological features at both the local texture scale and the brain region structural scale. The local texture scale focused on the pixel-texture connectivity characteristics within a single brain region, and the extracted topological features included node degree and clustering coefficient. The brain region structural scale focused on the inter-brain connectivity characteristics across the entire brain, and the extracted topological features included feature path length and network efficiency. The node degree at the local texture scale was also analyzed. The number of connections between a single brain region node and other brain region nodes is represented by the following formula:

[0037] Among them, the connection weight threshold The connection weights of normal brain networks were determined by statistical characteristics. Brain networks were constructed using MRI data from at least 80 normal brain structures, and the mean of all connection weights was calculated. with standard deviation ,Pick As a threshold for connection weight; This is an indicator function; it takes the value 1 when the condition within the parentheses is true and 0 when the condition is false; the feature path length at the brain region structural scale. The average shortest path between nodes in a brain network, reflecting the network's global connectivity, is calculated as follows:

[0038] in, For nodes With nodes The shortest path length between the two is obtained by solving the connection weight matrix of the brain network using Dijkstra's algorithm. The algorithm's iteration step size is 1, and the maximum depth of the shortest path search is [missing value]. ; A topological anomaly detection model was constructed and trained to quantify the degree of topological anomalies in brain region nodes. The specific steps are as follows: Construct a topological feature dataset by selecting no fewer than 150 samples, 100 normal samples, and 50 tumor samples. Extract four types of topological features from all nodes in the brain network of each sample: node degree, clustering coefficient, feature path length, and network efficiency. Form a feature matrix with the dimension of number of samples × number of nodes × 4. The topological features are standardized, and each feature value is mapped to... The interval, the processing formula is:

[0039] The anomaly calculation model was trained using a random forest regression model with 100 decision trees, a maximum depth of 10, a minimum number of sample splits of 5, a learning rate of 0.1, and 50 training epochs. The topological anomalies of the labeled tumor regions were used as labels, and a 5-fold cross-validation model was used to evaluate the model. The model was considered to be trained when the mean squared error (MSE) on the validation set was no higher than 0.01. Calculate the topological feature deviation of each node in the brain network to be identified using the trained model. The deviation is the mean relative error between each topological feature value and the mean of the corresponding topological feature value in a normal brain network, and is calculated as follows:

[0040] in, For nodes The One topological feature value, For a large number of normal brain networks The statistical mean of each topological feature. These correspond to node degree, clustering coefficient, feature path length, and network efficiency, respectively. The impact of topologically anomalous regions on the structural stability of adjacent brain regions is calculated using a graph propagation mechanism. This mechanism leverages the connection weights of brain networks to achieve spatial propagation of anomalous information. Connection weights between nodes Positive correlation, deviation from the topological features of the source node For positive correlation, the propagation weights are first normalized, and then the deviation of neighboring brain regions after propagation is calculated. The calculation formula is:

[0041] in, For nodes The set of directly adjacent nodes, i.e., nodes All nodes connected by edges exist; propagation weights. The normalized relative weights are obtained by summing and normalizing the products of the connection weights and deviations of the source node and its neighboring nodes, ensuring that the sum of the propagation weights of all neighbors of a single node is 1. Determine the topology anomaly threshold , Based on the topological feature deviation statistics of labeled tumor samples, at least 50 labeled brain tumor samples were selected, and the topological feature deviation of the brain region where the tumor was located and the adjacent brain regions was calculated. The median of all deviations was taken as the mean. Set up potential tumor candidate regions In the context of brain regions, the deviation of topological features from the nodes in the relevant brain region is considered. The deviation of the region or its adjacent brain regions after map propagation. Regions identified as topologically anomalous areas were further filtered and integrated to obtain a set of suspected tumor regions with topological anomalous features. .

[0042] This design, based on cross-scale brain network topological perturbation verification, constructs a brain structural connectivity network map, extracts core topological features from local texture and brain region structural scales, comprehensively reflects the connectivity of brain structures, trains a topological anomaly discrimination model, and, based on a large amount of sample data, can accurately quantify the degree of node topological anomalies. It utilizes the map propagation mechanism to consider the impact of abnormal regions on neighboring brain regions, determines the topological anomaly threshold, and screens out suspected tumor regions. From the perspective of brain network topology, combined with multi-scale feature analysis, it can more deeply discover brain structural changes caused by tumors, improve the accuracy of tumor identification, and avoid missed diagnoses and misdiagnoses.

[0043] In one embodiment, a brain tumor adaptive identification method based on metabolic-morphological co-constraints performs co-discriminatory fusion on suspected tumor regions to achieve accurate identification and localization of brain tumor regions, including: Based on preprocessed diffusion-weighted MRI sequence normalized data, the apparent diffusion coefficient of each pixel in the image was calculated using the diffusion tensor imaging algorithm. and fractional anisotropy ; The specific steps for constructing and training a tissue water migration model are as follows: A training dataset was constructed by selecting no fewer than 180 normal brain diffusion-weighted MRI samples, covering different age and gender groups. The ADC value, FA value, and corresponding measured values ​​of tissue water migration of each pixel in each sample were extracted and obtained through a metabolic detection device. Initialize the parameters of the multiple linear regression model, initial coefficients Set to 0, constant term Set to 0, and use the mean squared error (MSE) as the loss function; The model was trained using gradient descent with a learning rate of 0.001, 1000 iterations, and a batch size of 32. The convergence condition was that the loss function value was less than 0.005 or the maximum number of iterations was reached. Model validation involves dividing the dataset into training and validation sets in a 7:3 ratio, and then calculating the coefficient of determination on the validation set. If the value is not lower than 0.85, the model is considered to have completed training; otherwise, additional samples are provided and the model is retrained. Tissue water transport model completed using training By inferring the mapping of metabolic-related signals, a two-dimensional probability map of metabolic abnormalities is obtained. The spatial dimension of this probability map is consistent with that of the preprocessed image. Each pixel corresponds to a metabolic abnormality probability value. The calculation formula is:

[0044] in, The statistical mean of the output values ​​of the water migration model for normal brain tissue. This represents the statistical standard deviation of the output values ​​from the normal brain tissue water migration model. The Sigmoid activation function is used to map the calculation result to... The interval serves as the probability value for metabolic abnormalities; The specific steps for constructing and training a tumor morphological evolution constraint model are as follows: A tumor morphology dataset was constructed, and no fewer than 120 labeled samples of brain tumors of different stages and types were selected. The spatial coordinates and morphological parameters of the tumor core region and invasion region of each sample were extracted. Initialize morphological diffusion model parameters and diffusion scale parameters. The initial value is set to the average radial diffusion distance of the tumor, and the initial value of the diffusion attenuation coefficient is set to 0.1; The model was trained using the maximum likelihood estimation method. The objective function was to maximize the likelihood probability of tumor morphological spread. The number of iterations was set to 80, and the convergence threshold was set to 1e-4. Model validation shows that the morphological prediction accuracy on the test set is no less than [percentage missing]. At that time, it is determined that the model training is complete; Define the tumor morphological diffusion potential function using the trained model. The value of this function decreases as the distance from the pixel to the center of the suspected tumor area increases, and the calculation formula is:

[0045] in, Spatial coordinates Pixels at the location to the suspected tumor area Euclidean distances between the centers of each region, morphological diffusion scale parameters Tumor morphology statistics were used to identify at least 60 labeled brain tumor samples. The maximum radial diffusion distance of each tumor region was calculated, and the mean of all distances was taken as the mean. The initial value is then adaptively adjusted based on the size of the suspected tumor area, multiplied by 1.2 when the area is greater than 100 pixels and multiplied by 0.8 when the area is less than 20 pixels. The specific steps for constructing and training a metabolism-morphology co-discrimination model are as follows: Construct a fusion feature dataset, select no less than 200 labeled brain tumor samples, and extract the metabolic abnormality probability features and morphological diffusion potential features of each sample, as well as the corresponding tumor annotation labels; Initialize fusion model parameters, metabolic feature weights The initial value is set to 0.5, and the morphological feature weight is... The initial value is set to 0.5; We optimized the weight parameters using cross-validation, with 5-fold cross-validation, 50 iterations per fold, and a learning rate of 0.01. Model evaluation uses F1 score as the evaluation metric. The parameter corresponding to the maximum F1 score is the optimal parameter. The training is considered complete when the model's F1 score on the test set is not lower than 0.9. The trained co-discriminative model is used to map metabolic abnormality probabilities. Reflecting tissue metabolic characteristics and morphological diffusion potential function The spatial morphological characteristics of the tumor are weighted and fused to obtain the co-discriminative value of each pixel. The integrated type is:

[0046] in, The weighting coefficients for metabolic features satisfy... , The weighting coefficients for morphological features; Determine the collaborative discrimination threshold , ROC curve analysis was used to determine the codiscriminative discriminant value. Using pixels labeled with tumor as positive examples and normal pixels within the suspected tumor area as negative examples, an ROC curve was plotted, and the codiscriminative discriminant value corresponding to the maximum Youden's exponent was selected as the codiscriminative discriminant value. Grouping suspected tumor areas China satisfies The pixels are identified as tumor pixels. Spatial connectivity analysis, contour extraction and region merging are performed on all tumor pixels to obtain the accurately identified brain tumor region. At the same time, the two-dimensional / three-dimensional spatial coordinate range, contour boundary and region area / volume information of the region are output to complete the accurate identification and localization of brain tumor.

[0047] This design, based on metabolic-morphological co-constraint for precise identification, constructs a tissue water migration model and a tumor morphological evolution constraint model. It mines tumor features from metabolic and morphological perspectives respectively, trains a metabolic-morphological co-discrimination model, and weights and fuses metabolic and morphological features. It judges tumor pixels by integrating multiple aspects of information, determines the co-discrimination threshold, and accurately identifies brain tumor regions. By combining metabolic and morphological features, it fully leverages the advantages of both and overcomes the limitations of single features, enabling more accurate and comprehensive identification of brain tumors. At the same time, it outputs detailed information, providing strong support for clinical diagnosis and treatment.

[0048] In one embodiment, a tissue water migration model This is a fitting model based on multiple linear regression, trained using a large amount of diffusion-weighted MRI data from normal brain tissue. The model expression is as follows:

[0049] in, These are the apparent diffusion coefficients. Fractional anisotropy The fitting coefficient, For model constants, All of these are numerical constants obtained by fitting using the least squares method; during model training, the Adam optimizer is used for gradient descent, and the weight decay coefficient is set to 0.0001 to prevent overfitting.

[0050] This design employs multiple linear regression with the Adam optimizer to model tissue water migration. Multiple linear regression can easily and effectively fit the relationship between ADC, FA, and tissue water migration in diffusion-weighted MRI data of normal brain tissue. The Adam optimizer can dynamically adjust the learning rate of each parameter based on the first and second moment estimates of the gradient, adapting to different parameter conditions. The weight decay coefficient prevents the model from overfitting, making the features learned by the model on the training set more generalizable and better applicable to real data. This improves the accuracy of calculating the metabolic abnormality probability map and provides a reliable basis for subsequent precise tumor identification.

[0051] In one embodiment, metabolic feature weighting coefficient The determination process is as follows: A set of brain tumor MRI samples labeled by professional physicians is selected, and the sample set is randomly divided into a training set and a validation set in a 7:3 ratio. The search scope is The search step size is On the training set, for different The collaborative discriminant model with different values ​​is trained, and different values ​​are calculated on the validation set. The tumor identification accuracy, precision, and recall of the model were evaluated using the F1 score as a comprehensive metric. The F1 score is calculated as follows:

[0052] The maximum F1 score corresponds to As the optimal weight coefficient, it is fixed in the collaborative discrimination model; if there are multiple For the same highest F1 score, the value with the smallest weight of metabolic features is selected as the optimal weight coefficient; an early stopping strategy is adopted during model training. When the F1 score on the validation set no longer improves for 5 consecutive rounds, training is stopped to avoid overfitting.

[0053] This design, through sample set partitioning and different... The weight coefficients of metabolic features are determined through training and validation. The sample set is divided into training and validation sets, and the values ​​are set accordingly. Search range and step size, in different The model is trained and evaluation metrics are calculated on the validation set. The model with the highest F1 score is selected as the optimal one. The early stop strategy avoids overfitting. It uses a systematic approach to find the optimal weights, comprehensively considering the accuracy, precision, and recall of tumor identification, so that metabolic and morphological features play the most reasonable role in collaborative discrimination, thereby improving the accuracy and stability of tumor identification and providing more reliable results for clinical diagnosis.

[0054] In one embodiment, the method further includes performing quantitative analysis of tumor characteristics on the precisely identified brain tumor region, extracting core morphological and functional characteristics of the tumor. The morphological characteristics include the tumor's two-dimensional area and three-dimensional volume, while the functional characteristics include the average relaxation difference value and average metabolic abnormality probability value of the tumor region. All quantitative indicators provide objective quantitative basis for the graded diagnosis and disease assessment of brain tumors, including the tumor's three-dimensional volume. The formula for calculation is:

[0055] in, A pixel set for precise identification of brain tumor regions. The actual planar area of ​​a single pixel in an MRI image is determined by the image's scan resolution. The slice thickness of the MRI image is determined by the acquisition parameters of the magnetic resonance imaging equipment.

[0056] This design allows for the quantitative analysis of precisely identified brain tumor regions, extracting core morphological features, two-dimensional area, three-dimensional volume, functional indicators, average relaxation difference value, and average metabolic abnormality probability value. This provides objective quantitative evidence for the graded diagnosis and disease assessment of brain tumors. The three-dimensional tumor volume calculation formula takes into account the actual planar area and layer thickness of image pixels, making the calculation results more accurately reflect the actual size of the tumor. These quantitative indicators help doctors to have a more comprehensive understanding of the tumor, formulate more reasonable treatment plans, track disease progression, and improve the scientificity and accuracy of brain tumor diagnosis and treatment.

[0057] Example 2: Please see Figure 4 A brain tumor identification system in brain magnetic resonance imaging (MRI) images, applied to the above-described brain tumor identification method in brain MRI images, the system comprising: Brain MRI image preprocessing unit, multi-echo tumor preliminary screening unit, cross-scale tumor topology verification unit, and metabolic morphology-assisted tumor identification unit; The brain MRI image preprocessing unit is used to acquire the raw brain MRI image data of the object to be identified and perform preprocessing operations to obtain standardized multimodal MRI image data; The multi-echo tumor preliminary screening unit is used to perform preliminary tumor region screening on preprocessed image data based on the tumor tissue probability decoupling identification method based on multi-echo relaxation differences. The potential tumor candidate region set is obtained through relaxation characteristic analysis and tissue signal decomposition. The cross-scale tumor topology verification unit is used to verify the topological features of potential tumor candidate regions by a tumor identification method based on cross-scale brain network topology perturbation. It combines cross-scale analysis of brain structural networks and abnormal propagation detection to screen out suspected tumor regions with topological anomalies. The metabolic-morphological co-organized tumor identification unit is used to perform co-discrimination and fusion of suspected tumor regions in an adaptive brain tumor identification method based on metabolic-morphological co-constraints. It achieves accurate identification and localization of brain tumor regions through the dual constraints of metabolic features and morphological evolution.

[0058] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying brain tumors in brain magnetic resonance imaging (MRI) images, characterized in that, The method includes the following steps: S1. Acquire the raw MRI image data of the brain of the object to be identified and perform preprocessing operations to obtain standardized multimodal MRI image data; S2. A probabilistic decoupling identification method for tumor tissue based on multi-echo relaxation differences is used to perform preliminary tumor region screening on preprocessed image data, and a set of potential tumor candidate regions is obtained through relaxation characteristic analysis and tissue signal decomposition. S3. A tumor identification method based on cross-scale brain network topological perturbation verifies the topological features of the potential tumor candidate region set, and combines cross-scale analysis of brain structural networks and abnormal propagation detection to screen out suspected tumor regions with topological abnormalities. S4. The adaptive brain tumor identification method based on metabolic-morphological co-constraint performs co-discrimination and fusion on the suspected tumor region, and achieves accurate identification and localization of brain tumor region through the dual constraints of metabolic features and morphological evolution.

2. The method for identifying brain tumors in brain magnetic resonance images according to claim 1, characterized in that, In step S1, the raw brain MRI image data obtained includes: Multi-echo MRI sequence data, structural MRI sequence data, diffusion-weighted MRI sequence data; The preprocessing operations sequentially include head region cropping, noise suppression, multimodal and multi-echo image registration, and grayscale normalization. Grayscale normalization maps the image pixel grayscale values ​​to... This allows for the standardized processing of multimodal MRI image data within a given range.

3. The method for identifying brain tumors in brain magnetic resonance images according to claim 1, characterized in that, In step S2, the tumor tissue probabilistic decoupling identification method based on multi-echo relaxation differences includes: The relaxation values ​​of pixels were calculated and the relaxation difference response matrix was constructed based on the normalized data of the preprocessed multi-echo MRI sequence. The relaxation gradient change field was calculated with echo time as the dimension. A tissue relaxation probability decoupling model was constructed and trained. This model was then used to perform probability decomposition on the mixed relaxation signals of image pixels to obtain the signal contribution weights of various tissues. The relaxation gradient threshold is determined, and potential tumor pixels are obtained by combining the results of relaxed abnormal pixel labeling with the tumor tissue signal contribution weight screening conditions. Regional connectivity analysis and merging are performed on potential tumor pixels to form a set of potential tumor candidate regions.

4. The method for identifying brain tumors in brain magnetic resonance images according to claim 3, characterized in that: The tissue relaxation probability decoupling model is a Gaussian mixture model. The model mixture components correspond to four types of tissues: gray matter, white matter, cerebrospinal fluid, and tumor tissue. The model is trained using the EM algorithm. After training, the mixed relaxation signal is probabilistically decomposed to output the signal contribution weight of each type of tissue in the pixel mixed signal. The relaxation gradient threshold is adaptively adjusted by combining the percentile method with brain image features of different age groups, and the tumor tissue signal contribution weight screening threshold is determined based on the maximum value of the Youden index from ROC curve analysis.

5. The method for identifying brain tumors in brain magnetic resonance images according to claim 1, characterized in that, In step S3, the tumor identification method based on cross-scale brain network topological perturbation includes: An undirected weighted brain structure connectivity network was constructed based on preprocessed normalized structural MRI sequence data, with standardized brain regions as network nodes and structural similarity or spatial continuity between brain regions as connection edge weights. Cross-scale topological feature analysis of brain structural connectivity networks at both the local texture scale and the brain region structural scale was performed to extract core topological features; Construct and train a topological anomaly discrimination model, and calculate the topological feature deviation of each node in the brain network to be identified; The influence of topologically anomalous regions on neighboring brain regions was calculated using the atlas propagation mechanism, and the deviation of neighboring brain regions after propagation was obtained. A threshold for topological anomalies is determined, and topologically anomalous regions are screened by combining the deviation of node topological features with the deviation of neighboring brain regions after propagation, and integrated to form a set of suspected tumor regions.

6. The method for identifying brain tumors in brain magnetic resonance images according to claim 5, characterized in that, The cross-scale topological features include: Node degree and clustering coefficient at the local texture scale, and feature path length and network efficiency at the brain region structure scale; The topological anomaly discrimination model is a random forest regression model. The model is trained with the topological anomaly degree of the tumor region as the label and evaluated by 5-fold cross-validation. The graph propagation mechanism is based on the brain network connection weights to achieve spatial propagation of abnormal information. The propagation weights are positively correlated with the connection weights between nodes and the deviation of the source node topological features, and are normalized.

7. The method for identifying brain tumors in brain magnetic resonance images according to claim 1, characterized in that, In step S4, the adaptive brain tumor identification method based on metabolic-morphological co-constraint includes: Based on the preprocessed diffusion-weighted MRI sequence normalized data, the apparent diffusion coefficient and fractional anisotropy of the pixels were calculated, a tissue water migration model was constructed and trained, and a metabolic abnormality probability map was obtained using the model. A tumor morphology evolution constraint model is constructed and trained, a tumor morphology diffusion potential function is defined, and the morphology diffusion potential value of each pixel is obtained. A metabolism-morphology co-discrimination model was constructed and trained, and the metabolic abnormality probability features and morphological diffusion potential features were weighted and fused to obtain the co-discrimination value of the pixel. A collaborative discrimination threshold is determined, and tumor pixels are filtered based on the threshold. Through spatial connectivity analysis, contour extraction, and region merging, accurate identification and localization of brain tumor regions are achieved, while outputting the spatial coordinates, contour boundaries, and area / volume information of the tumor regions.

8. The method for identifying brain tumors in brain magnetic resonance images according to claim 7, characterized in that: The tissue water migration model is a fitting model based on multiple linear regression, trained using gradient descent combined with the Adam optimizer, and the model output is mapped to a Sigmoid activation function. The probability value of metabolic abnormalities within the interval; The value of the tumor morphology diffusion potential function decreases as the distance from the pixel to the center of the suspected tumor region increases, and its diffusion scale parameter is adaptively adjusted according to the tumor morphology statistical results and the area of ​​the suspected tumor region. The metabolism-morphology co-discrimination model optimizes the weighting coefficients of metabolic and morphological features through cross-validation. The optimal value of the weighting coefficients is determined based on the maximum F1 score, and the co-discrimination threshold is determined by the maximum Youden index obtained from ROC curve analysis.

9. The method for identifying brain tumors in brain magnetic resonance images according to claim 1, characterized in that, Also includes: The steps for quantitative analysis of tumor characteristics in precisely identified brain tumor regions include extracting core morphological and functional characteristic indicators of the tumor. The morphological characteristic indicators include the two-dimensional area and three-dimensional volume of the tumor, and the functional characteristic indicators include the average relaxation difference value and the average metabolic abnormality probability value of the tumor region. All quantitative indicators provide objective quantitative basis for the graded diagnosis and disease assessment of brain tumors. The three-dimensional volume of the tumor is calculated based on the actual planar area of ​​the MRI image pixels, the scanning slice thickness, and the pixel set of the tumor region.

10. A brain tumor identification system in brain magnetic resonance imaging, characterized in that, A system for identifying brain tumors in brain magnetic resonance images according to any one of claims 1-9, the system comprising: Brain MRI image preprocessing unit, multi-echo tumor preliminary screening unit, cross-scale tumor topology verification unit, and metabolic morphology-assisted tumor identification unit; The brain MRI image preprocessing unit is used to acquire the raw brain MRI image data of the object to be identified and perform preprocessing operations to obtain standardized multimodal MRI image data. The multi-echo tumor preliminary screening unit is used to perform preliminary tumor region screening on preprocessed image data based on the tumor tissue probability decoupling identification method based on multi-echo relaxation differences, and obtain a set of potential tumor candidate regions through relaxation characteristic analysis and tissue signal decomposition. The cross-scale tumor topology verification unit is used to verify the topological features of the potential tumor candidate region set by the tumor identification method based on cross-scale brain network topology perturbation, and to screen out the suspected tumor regions with topological anomalous by combining cross-scale analysis of brain structural networks and abnormal propagation detection. The metabolic morphology-coordinated tumor identification unit is used to perform collaborative discrimination and fusion of the suspected tumor region based on the adaptive brain tumor identification method of metabolic-morphological co-constraint, and to achieve accurate identification and localization of brain tumor region through the dual constraints of metabolic features and morphological evolution.

Images