Tinnitus diagnosis and classification system
By preprocessing and extracting features from tinnitus MRI data, constructing a multi-map connection matrix and fusing node features, accurate classification of tinnitus is achieved, solving the problem of low detection rate of objective tinnitus and providing an effective treatment plan.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-04-07
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The low clinical detection rate of objective tinnitus in current technologies prevents patients from receiving targeted and effective treatment.
By preprocessing tinnitus MRI data, connection matrices under multiple preset brain maps are constructed, node feature vectors are extracted and optimized, and feature vectors from multiple maps are fused to ultimately achieve tinnitus classification.
It has improved the clinical detection rate of objective tinnitus, provided targeted treatment options, and reduced patient suffering.
Smart Images
Figure CN121808477A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of functional nuclear magnetic resonance signal processing technology, specifically a tinnitus diagnostic classification system. Background Technology
[0002] Tinnitus is a common symptom where patients perceive sound even without an external sound source. It has a high global incidence rate and causes significant distress to patients' lives and mental well-being. Clinically, tinnitus is classified into subjective and objective tinnitus based on whether the sound source can be perceived by others. Objective tinnitus currently has better clinical intervention outcomes; however, the current clinical detection rate for objective tinnitus is low, resulting in patients not receiving targeted and effective treatment and enduring unnecessary suffering.
[0003] Therefore, providing a system that can accurately classify tinnitus is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] In view of this, embodiments of this application provide a tinnitus diagnosis and classification system, which solves the technical problem that the existing clinical detection rate of objective tinnitus is low, resulting in patients being unable to receive targeted and effective treatment.
[0005] The first aspect of this application provides a tinnitus diagnostic classification system, including: The preprocessing module is used to preprocess the tinnitus MRI data to be classified, and obtain preprocessed MRI data. A module is constructed to use N preset brain maps to construct a first connection matrix under each preset brain map through the preprocessed MRI data, where N is a non-zero natural number. A node feature extraction module is used to extract the node features of the first node in each of the first connection matrices to obtain the first node feature vector corresponding to the first node, where the first node is any node in the first connection matrix. A node feature optimization module is used to perform graph-level feature extraction on the node feature vector of the first node to obtain the second node feature vector corresponding to the important node, wherein the important node is the key node in the first node; The fusion module is used to fuse all the feature vectors of the second node to obtain the fused multi-map feature vector; A classification module is used to classify the multi-spectral feature vectors to obtain the tinnitus classification corresponding to the tinnitus MRI data.
[0006] Furthermore, the construction module includes: The construction module includes: The extraction module is used to extract the signal time series of each node under each preset brain map based on the preprocessed MRI data and according to N preset brain maps; The calculation module is used to calculate the correlation coefficient between nodes based on the signal time series of the nodes under the same preset brain map. The first transformation module is used to perform Fisher Z-transform on the correlation coefficients under each of the preset brain maps to obtain transformed correlation coefficients; A construction module is used to construct a first connection matrix under each preset brain map based on the transformation correlation coefficients under each preset brain map.
[0007] Furthermore, N is 3; The three preset brain maps are the AAL90 map, the DOS160 map, and the Power264 map.
[0008] Furthermore, the node feature extraction module includes: The aggregation module is used to perform mean aggregation on the node features of the first node and its adjacent nodes in each of the first connection matrices to obtain the aggregated features corresponding to the first node. The mapping module is used to map the aggregated features of the first node to a preset feature space to obtain the first node feature vector corresponding to the first node.
[0009] Furthermore, the mapping module includes: The transformation unit is used to perform a linear transformation on the aggregated features of the first node to obtain the corresponding first feature; The mapping unit is used to map the first feature to a preset feature space through a nonlinear activation function to obtain the first node feature vector corresponding to the first node.
[0010] Furthermore, the node feature optimization module includes: The convolution module is used to aggregate and transform the node feature vectors of the first node and its neighboring nodes through graph convolution to generate the node update features of the first node. The filtering module is used to filter the first nodes under each preset brain map based on the node importance score to obtain important nodes, and to obtain the subgraph adjacency matrix based on the node update features corresponding to each important node. The structure learning module is used to reconstruct the adjacency matrix of each subgraph through a sparse attention mechanism to obtain the second node feature vector corresponding to each important node.
[0011] Furthermore, the fusion module includes: The first fusion module is used to fuse the feature vectors of the second nodes under each preset brain map based on in-map attention to obtain the overall map representation of the preset brain map; The second fusion module is used to fuse the overall representation of all the graphs based on inter-graph attention to obtain a multi-graph feature vector.
[0012] Furthermore, the first fusion module includes: The first allocation unit is used to allocate a first weight to the important nodes corresponding to the feature vectors of each second node. The summation unit is used to obtain the overall representation of the preset brain map by using the first weight and the second node feature vector under each preset brain map.
[0013] Furthermore, the second fusion module includes: The second allocation unit is used to allocate a second weight to the overall representation of the map under each preset brain map; The fusion unit is used to fuse all the overall representations of the graphs based on the second weight to obtain a multi-graph feature vector.
[0014] Furthermore, the classification module includes: The second transformation module is used to perform a linear transformation on the multi-spectral feature vector based on a preset weight matrix to obtain a first intermediate vector; The bias processing module is used to bias the first intermediate vector to obtain the tinnitus classification corresponding to the tinnitus MRI data.
[0015] One of the above technical solutions has the following advantages and effects: One of the tinnitus diagnosis and classification systems described above includes: a preprocessing module for preprocessing the tinnitus MRI data to be classified to obtain preprocessed MRI data; a construction module for constructing a first connection matrix under each of the N preset brain maps using the preprocessed MRI data, where N is a non-zero natural number; a node feature extraction module for extracting node features from the first nodes in each first connection matrix to obtain the first node feature vector corresponding to the first node; a node feature optimization module for filtering and optimizing the node feature vectors of the first nodes to obtain the second node feature vectors corresponding to important nodes, where important nodes are the key nodes in the first nodes; a fusion module for fusing all the second node feature vectors to obtain the fused multi-map feature vector; and a classification module for classifying the multi-map feature vector to obtain the tinnitus classification corresponding to the tinnitus MRI data.
[0016] As described above, the tinnitus diagnosis and classification system first preprocesses the tinnitus MRI data to be classified using a preprocessing module, obtaining preprocessed MRI data. Next, a construction module constructs a first connection matrix under each preset brain atlas using the preprocessed MRI data. Then, a node feature extraction module obtains the first node feature vector of the first node in each first connection matrix. Next, a node feature optimization module filters and optimizes the first node feature vector of the first node to obtain the second node feature vector corresponding to important nodes. Then, a fusion module fuses all the second node feature vectors to obtain a fused multi-atlas feature vector. Finally, a classification module classifies the multi-atlas feature vector to obtain the tinnitus classification corresponding to the tinnitus MRI data. This application, through a series of processing and analysis steps using the resting-state functional MRI data of the user's brain (i.e., tinnitus MRI data), combined with the central mechanisms of tinnitus, achieves tinnitus classification, solving the technical problem of low clinical detection rate of objective tinnitus, which prevents patients from receiving targeted and effective treatment. Attached Figure Description
[0017] 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 schematic diagram of the structure of an embodiment of a tinnitus diagnosis and classification system provided in this application. Figure 2 This is a schematic diagram of a second embodiment of a tinnitus diagnosis and classification system provided in this application. Detailed Implementation
[0018] This application provides a tinnitus diagnosis and classification system that solves the technical problem of low clinical detection rate of objective tinnitus, which prevents patients from receiving targeted and effective treatment.
[0019] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this application.
[0020] For easier understanding, please refer to Figure 1 , Figure 1 This is a schematic diagram of the structure of a tinnitus diagnosis and classification system according to an embodiment of this application. Figure 1 As shown, the tinnitus diagnosis and classification system in this embodiment specifically includes: The preprocessing module is used to preprocess the tinnitus MRI data to be classified, and obtain preprocessed MRI data. A module is constructed to use N preset brain maps and preprocess MRI data to construct the first connection matrix under each preset brain map, where N is a non-zero natural number. The node feature extraction module is used to extract the node features of the first node in each first connection matrix to obtain the first node feature vector corresponding to the first node, where the first node is any node in the first connection matrix. The node feature optimization module is used to perform graph-level feature extraction on the node feature vector of the first node to obtain the feature vector of the second node corresponding to the important node. The important node is the key node in the first node. The fusion module is used to fuse all the feature vectors of the second node to obtain the fused multi-map feature vector; The classification module is used to classify multi-spectral feature vectors to obtain the tinnitus classification corresponding to the tinnitus MRI data.
[0021] It should be noted that, in one embodiment, the steps for preprocessing tinnitus MRI data using a preprocessing module are as follows: 1. Remove initial frames: Discard the first 5 frames of tinnitus MRI data to eliminate the magnetic saturation effect during machine initialization.
[0022] 2. Time-level correction: Since the tinnitus MRI data was acquired using a slice-by-slice scanning strategy with an acquisition time interval of approximately TR / 2, where TR stands for Repetition Time, which is the time interval between two adjacent radiofrequency excitation pulses, time-level correction was applied to the tinnitus MRI data after removing the initial frame. By using sinc interpolation, the tinnitus MRI data acquired from different slices were resampled to a unified time point to obtain time-aligned tinnitus MRI data, thereby ensuring the synchronicity and comparability of signals from different brain regions.
[0023] 3. Head movement correction: Align the reference MRI data of the user to be classified with the time-aligned tinnitus MRI data to obtain time series data with minimal translation and rotation effects, thereby minimizing the interference of head movement on the signal and achieving head movement correction on the individual time axis, thus achieving partial head movement correction effect.
[0024] 4. Spatial Standardization: The head-motion-corrected tinnitus MRI data is mapped to a standard brain template space using a linear transformation method, and the images are resampled to obtain spatially standardized tinnitus MRI data. The specific processing flow is as follows: First, the T1 structural image of the head-motion-corrected tinnitus MRI data is aligned with the functional image using a registration method to obtain T1_1; then, T1_1 is segmented and linearly registered to the EPI template (Echo Planar Imaging Template, a standard functional MRI image template constructed based on echo planar imaging sequences to provide a spatial reference consistent with the functional image for subsequent spatial normalization and registration processing), resulting in T1_2 aligned with the standard brain template space; finally, the transformation parameters of T1_1 segmentation and registration to T1_2 are applied to the filtered tinnitus MRI data to obtain tinnitus MRI data in the standard brain template space.
[0025] 5. Regression of Confounding Variables: For the spatially standardized tinnitus MRI data, linear regression was used to regress confounding variables to obtain the net signal after removing non-neural signal interference. Specific processing included: regressing six head movement parameters (translation and rotation) and their first-order time derivatives and squared terms; regressing ventricular signals and their first derivatives; and regressing white matter signals and their first derivatives, thereby minimizing the impact of head movement and non-gray matter signals on functional connectivity analysis.
[0026] 6. Spatial smoothing: Spatial smoothing is applied to the spatially standardized tinnitus MRI data. A three-dimensional Gaussian kernel with a full width at half maximum (FWHM) of 6 mm is used for convolution to obtain spatially smoothed tinnitus MRI data, thereby reducing high-frequency noise and enhancing signal consistency between neighboring voxels, which facilitates subsequent statistical analysis.
[0027] 7. Temporal filtering: Temporal filtering is applied to the spatially smoothed tinnitus MRI data. First, linear detrending is performed to remove slow drift, and then bandpass filtering of 0.01–0.08 Hz is applied to the signal to obtain a net signal that retains only low-frequency fluctuations, thereby enhancing the time series features related to neural activity and facilitating subsequent functional connectivity analysis.
[0028] It should be noted that the first connection matrix constructed by the module is essentially a functional connection matrix. This functional connection matrix is mainly used to describe the strength or correlation of functional connections between different regions of the brain, reflecting the functional network organization pattern of the brain. The second node feature vector extracted by the node feature optimization module is actually a high-fidelity representation of important nodes in the graph convolution space. After completing the node selection, this module completes the local topology after pooling through sparse attention structure learning, so that the node features corresponding to important nodes retain complete functional semantic associations while reducing dimensionality, thereby effectively suppressing the feature confounding problem caused by noise, redundant connections, or weakly correlated nodes in the original graph. The second node feature vector generated in this way not only has stronger discriminative ability and stability, but also strengthens the potential collaborative patterns between key nodes while maintaining the consistency of local topology, making it a more expressive and robust basic representation in subsequent cross-graph fusion and classification tasks. The fusion module completes the aggregation of nodes to the graph. The feature vector of a single node only contains local information (such as its own attributes or neighborhood relationships). By aggregating the features of all nodes (or important nodes) in the graph, the local information is integrated into a global representation. This global representation can capture the structural patterns, functional connectivity patterns, or semantic information of the entire graph.
[0029] Specifically, existing methods for tinnitus grading based on functional connectivity constructed using electroencephalography (EEG) rely on inferences from scalp electrical signals, which are easily affected by volume conduction, electromyographic interference, and electrode placement limitations, making it difficult to distinguish the true synchronicity of different brain regions. In contrast, the functional magnetic resonance imaging (fMRI) method used in this embodiment acquires low-frequency fluctuations across the entire brain using blood oxygenation level-dependent signals. This allows for the identification of cortical and deep structural activity patterns at the millimeter scale, resulting in more precise localization and higher resolution of network connectivity. Furthermore, fMRI connectivity measurements are unaffected by electrical signal aliasing, exhibiting better stability, and can be combined with structural imaging or metabolic information to verify the anatomical basis of connectivity. Therefore, fMRI can more reliably reflect the coordinated changes in multiple brain regions and networks involved in tinnitus, making it more suitable for constructing tinnitus-related functional connectivity features and presenting the linkages between tinnitus-related brain regions with higher spatial accuracy.
[0030] In this embodiment, the tinnitus diagnosis and classification system first preprocesses the tinnitus MRI data to be classified using a preprocessing module to obtain preprocessed MRI data. Next, a construction module constructs a first connection matrix under each preset brain atlas using the preprocessed MRI data. Then, a node feature extraction module obtains the first node feature vector of the first node in each first connection matrix. Next, a node feature optimization module filters and optimizes the first node feature vector of the first node to obtain the second node feature vector corresponding to important nodes. Then, a fusion module fuses all the second node feature vectors to obtain a fused multi-atlas feature vector. Finally, a classification module classifies the multi-atlas feature vector to obtain the tinnitus classification corresponding to the tinnitus MRI data. This application uses the resting-state functional MRI data of the user's brain (i.e., tinnitus MRI data) to classify tinnitus, through a series of processing and analysis, combined with the central mechanisms of tinnitus, to achieve tinnitus classification. This solves the technical problem of low clinical detection rate of objective tinnitus, which prevents patients from receiving targeted and effective treatment.
[0031] The above is an embodiment of a tinnitus diagnosis and classification system provided in this application. The following is an embodiment of a tinnitus diagnosis and classification system provided in this application.
[0032] Please see Figure 2 , Figure 2 This is a schematic diagram of a second embodiment of a tinnitus diagnosis and classification system according to this application. Figure 2 As shown, the tinnitus diagnosis and classification system in this embodiment specifically includes: The preprocessing module is used to preprocess the tinnitus MRI data to be classified, and obtain preprocessed MRI data. A module is constructed to use N preset brain maps and preprocess MRI data to construct the first connection matrix under each preset brain map, where N is a non-zero natural number. The node feature extraction module is used to extract the node features of the first node in each first connection matrix to obtain the first node feature vector corresponding to the first node, where the first node is any node in the first connection matrix. The node feature optimization module is used to perform graph-level feature extraction on the node feature vector of the first node to obtain the feature vector of the second node corresponding to the important node. The important node is the key node in the first node. The fusion module is used to fuse all the feature vectors of the second node to obtain the fused multi-map feature vector; The classification module is used to classify multi-spectral feature vectors to obtain the tinnitus classification corresponding to the tinnitus MRI data.
[0033] When constructing the first connectivity matrix under various preset brain atlases using the construction module, the preprocessed MRI data are first divided according to the number of brain regions under different preset brain atlases. Then, the correlation coefficients between different brain regions under the same preset brain atlas are calculated. Next, the correlation coefficients are subjected to Fisher Z-transformation to obtain transformed correlation coefficients. Finally, the transformed correlation coefficients are used to construct the first connectivity matrix under different preset brain atlases. Specifically, in one optional implementation, the construction module includes: The extraction module is used to extract the signal time series of each node under each preset brain map based on preprocessed MRI data and N preset brain maps. The calculation module is used to calculate the correlation coefficient between nodes based on the signal time series of the nodes under the same preset brain map. The first transformation module is used to perform a Fisher Z-transform on the correlation coefficients under each preset brain map to obtain the transformed correlation coefficients. It can be understood that the purpose of the Fisher Z-transform here is to make the correlation distribution approximate a normal distribution. The module is used to construct the first connection matrix under each preset brain map based on the transformation correlation coefficients under each preset brain map.
[0034] It should be noted that in the above description, nodes are brain regions, each representing a region of interest under a different preset brain map, node features are brain region features, and node feature vectors are brain region feature vectors.
[0035] Furthermore, N is 3; The three preset brain maps are the AAL90 map, the DOS160 map, and the Power264 map.
[0036] AAL90 Atlas: Based on anatomical partitioning, the brain is divided into 90 brain regions, that is, the preprocessed MRI data is divided into 90 segmented signal time series. These 90 brain regions cover the main cortex and subcortex areas and are clearly structured.
[0037] The DOS160 map, constructed by Dosenbach et al., divides the brain into 160 regions. This involves dividing preprocessed MRI data into 160 segmented signal time series. These 160 brain regions include task-oriented functional networks and resting-state network nodes. It covers cognitive control, attention, and sensorimotor functional networks, taking into account both task-related and resting-state brain regions.
[0038] Power264 Atlas: Constructed based on large-scale resting-state fMRI data, the Power264 atlas divides the brain into 264 brain regions. This involves dividing preprocessed MRI data into 264 segmented signal time series, covering default mode networks, attention networks, and sensorimotor networks. It is entirely based on resting-state networks and offers high spatial resolution.
[0039] Understandably, when the preset brain atlases are AAL90, DOS160, and Power264, the corresponding first connectivity matrices for each brain region are obtained: AAL90 atlas yields a 90x90 first connectivity matrix, DOS160 atlas yields a 160x160 first connectivity matrix, and Power264 atlas yields a 264x264 first connectivity matrix. It should be noted that the initial features of each node in these three first connectivity matrices are the connection vectors between that node and all other nodes in the brain.
[0040] After obtaining the first connection matrix under different preset brain maps by constructing the module, the node features of each first connection matrix are then extracted by the node feature extraction module. In one specific implementation, the node feature extraction module includes: The aggregation module is used to perform mean aggregation on the node features of the first node and its adjacent nodes in each first connection matrix to obtain the aggregated features corresponding to the first node. The mapping module is used to map the aggregated features of the first node to a preset feature space to obtain the first node feature vector corresponding to the first node.
[0041] It should be noted that the node feature extraction module uses a graph neural network—GraphSage—for node feature extraction. The extraction process essentially involves updating the node features in the first connection matrix, combining the node's own features with those of neighboring nodes to generate new node features. GraphSage has three layers, with each layer having a fixed node feature dimension of 64. The ReLU activation function enhances non-linear expressiveness while maintaining gradient stability. During forward propagation at each layer, each node first averages its own features with those of its neighbors, then maps them to a new feature space using a linear transformation and a non-linear activation function, obtaining the current node's first node feature vector. After forward propagation, each node obtains a 64-dimensional first node feature vector. This vector not only preserves the node's own whole-brain connectivity information but also incorporates features from neighboring nodes, comprehensively reflecting the node's local and global functional role in the brain network.
[0042] Understandably, when the brain atlas is preset to AAL90, DOS160, and Power264, the resulting node feature vectors are 1*64 dimensional node feature vectors corresponding to the number of brain regions. That is, AAL90 atlas yields 90 1*64 dimensional node feature vectors, DOS160 atlas yields 160 1*64 dimensional node feature vectors, and Power264 atlas yields 264 1*64 dimensional node feature vectors.
[0043] As can be seen from the above, in one optional implementation, the mapping module includes: The transformation unit is used to perform a linear transformation on the aggregated features of the first node to obtain the corresponding first feature; The mapping unit is used to map the first feature to a preset feature space through a nonlinear activation function to obtain the first node feature vector corresponding to the first node.
[0044] It is understandable that after obtaining the node feature vectors, a node feature optimization module performs graph-level feature extraction on the node feature vectors. Simultaneously, nodes are filtered to retain important nodes. Therefore, after processing by the node feature optimization module, the second node feature vectors of the important nodes are obtained. Specifically, in one optional implementation, the node feature optimization module includes: The convolution module is used to aggregate and transform the node feature vectors of the first node and its neighboring nodes through graph convolution to generate the node update features of the first node. The filtering module is used to filter the first node under each preset brain map based on the node importance score to obtain important nodes, and obtain the subgraph adjacency matrix based on the node update features corresponding to each important node. The structure learning module is used to reconstruct the adjacency matrix of each subgraph through a sparse attention mechanism to obtain the second node feature vector corresponding to each important node.
[0045] It should be noted that the convolution module learns the functional connectivity patterns between different brain regions (nodes) through graph convolution, the filtering module filters key nodes based on node importance scores through graph pooling to obtain important nodes, and selects the most discriminative important nodes, thus retaining relevant features for classification while reducing dimensionality; the structure learning module reconstructs the adjacency matrix of the subgraph after graph pooling through a sparse attention mechanism to avoid network breaks caused by node loss and enhance the potential functional associations across networks.
[0046] Specifically, when the preset brain atlas is the AAL90 atlas, DOS160 atlas, and Power264 atlas, the number of second node feature vectors obtained by the node feature optimization module is the same as the number of important nodes. For example, when there are 48 important nodes selected under the AAL90 atlas, 48 1*64 second node feature vectors are obtained; when there are 135 important nodes selected under the DOS160 atlas, 135 1*64 second node feature vectors are obtained; and when there are 218 important nodes selected under the Power264 atlas, 218 1*64 second node feature vectors are obtained.
[0047] After the aforementioned processing, due to differences in brain region division and network structure among different brain atlases (AAL90 atlas, DOS160 atlas, and Power264 atlas), the information they contain is heterogeneous and complementary. Simply concatenating the feature vectors of the second nodes of the three atlases would introduce redundancy or even noise. Therefore, this embodiment designs a hierarchical attention mechanism to complete multi-atlas fusion. Thus, in one illustrative implementation, the fusion module includes: The first fusion module is used to fuse the feature vectors of the second nodes under each preset brain map based on in-map attention, so as to obtain the overall map representation of the preset brain map; The second fusion module is used to fuse the overall representation of all graphs based on inter-graph attention to obtain multi-graph feature vectors.
[0048] Furthermore, the first fusion module includes: The first allocation unit is used to assign first weights to the important nodes corresponding to the feature vectors of each second node. The summation unit is used to obtain the overall representation of the brain map under each preset brain map by utilizing the first weight and the second node feature vector.
[0049] Furthermore, the second fusion module includes: The second allocation unit is used to assign a second weight to the overall representation of the map under each preset brain map. The fusion unit is used to fuse all the overall representations of the graphs based on the second weight to obtain a multi-graph feature vector.
[0050] Specifically, the first fusion module mentioned above can be understood as an in-graph attention module. After completing graph convolution, graph pooling, and structure learning of Multi-GCN through the node feature optimization module, each brain map retains a set of selected important nodes and their corresponding second node feature vectors. The in-graph attention mechanism assigns attention weights to each important node, with higher weights indicating that the node is more discriminative for the classification task. Subsequently, the second node feature vectors of the important nodes are weighted and summed to obtain the overall graph representation under that map, realizing information aggregation from the node level to the graph level. That is, the AAL90 map, DOS160 map, and Power264 map each obtain one overall graph representation of that map.
[0051] It is understandable that the second fusion module mentioned above can be understood as an inter-graph attention module, which takes as input the overall graph representations of the three graphs. Weights are then distributed among the three overall graph representations, and the three overall graph representations are summed according to their weights to obtain the final unified, low-dimensional, and compact fusion representation (i.e., a single fused multi-graph feature vector).
[0052] Furthermore, in one alternative implementation, the classification module includes: The second transformation module is used to perform a linear transformation on the feature vectors of the multi-spectral maps based on a preset weight matrix to obtain the first intermediate vector; The bias processing module is used to bias the first intermediate vector to obtain the tinnitus classification corresponding to the tinnitus MRI data.
[0053] Understandably, the aforementioned classification module can be viewed as a classifier. The fused multi-spectral feature vector is input into the classifier and first passes through a fully connected layer. In this layer, the multi-spectral feature vector undergoes a linear transformation through a weight matrix and is then biased, mapping it to the class space. Subsequently, a softmax activation function is used to convert the output vector into a probability distribution, where each element represents the probability that a sample belongs to that category (subjective or objective tinnitus). Finally, by selecting the category with the highest probability, the model identifies the tinnitus type from the tinnitus MRI data, achieving automatic diagnosis of subjective and objective tinnitus.
[0054] In this embodiment, the tinnitus diagnosis and classification system first preprocesses the tinnitus MRI data to be classified using a preprocessing module to obtain preprocessed MRI data. Next, a construction module constructs a first connection matrix under each preset brain atlas using the preprocessed MRI data. Then, a node feature extraction module obtains the first node feature vector of the first node in each first connection matrix. Next, a node feature optimization module filters and optimizes the first node feature vector of the first node to obtain the second node feature vector corresponding to important nodes. Then, a fusion module fuses all the second node feature vectors to obtain a fused multi-atlas feature vector. Finally, a classification module classifies the multi-atlas feature vector to obtain the tinnitus classification corresponding to the tinnitus MRI data. This application uses the resting-state functional MRI data of the user's brain (i.e., tinnitus MRI data) to classify tinnitus, through a series of processing and analysis, combined with the central mechanisms of tinnitus, to achieve tinnitus classification. This solves the technical problem of low clinical detection rate of objective tinnitus, which prevents patients from receiving targeted and effective treatment.
[0055] Those skilled in the art will readily understand that the terms "first," "second," "third," "fourth," etc. (if present), in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0056] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.
[0057] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.
[0058] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0059] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0060] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A tinnitus diagnostic classification system, characterized in that, The tinnitus diagnostic classification system includes: The preprocessing module is used to preprocess the tinnitus MRI data to be classified, and obtain preprocessed MRI data. A module is constructed to use N preset brain maps to construct a first connection matrix under each preset brain map through the preprocessed MRI data, where N is a non-zero natural number. A node feature extraction module is used to extract the node features of the first node in each of the first connection matrices to obtain the first node feature vector corresponding to the first node, where the first node is any node in the first connection matrix. A node feature optimization module is used to perform node feature filtering and optimization on the node feature vector of the first node to obtain a second node feature vector corresponding to an important node, wherein the important node is the key node in the first node. The fusion module is used to fuse all the feature vectors of the second node to obtain the fused multi-map feature vector; A classification module is used to classify the multi-spectral feature vectors to obtain the tinnitus classification corresponding to the tinnitus MRI data.
2. The tinnitus diagnostic classification system according to claim 1, characterized in that, The construction module includes: The extraction module is used to extract the signal time series of each node under each preset brain map based on the preprocessed MRI data and according to N preset brain maps; The calculation module is used to calculate the correlation coefficient between nodes based on the signal time series of the nodes under the same preset brain map. The first transformation module is used to perform Fisher Z-transform on the correlation coefficients under each of the preset brain maps to obtain transformed correlation coefficients; A construction module is used to construct a first connection matrix under each preset brain map based on the transformation correlation coefficients under each preset brain map.
3. The tinnitus diagnostic classification system according to claim 1 or 2, characterized in that, The value of N is 3; The three preset brain maps are the AAL90 map, the DOS160 map, and the Power264 map.
4. The tinnitus diagnostic classification system according to claim 1, characterized in that, The node feature extraction module includes: The aggregation module is used to perform mean aggregation on the node features of the first node and its adjacent nodes in each of the first connection matrices to obtain the aggregated features corresponding to the first node. The mapping module is used to map the aggregated features of the first node to a preset feature space to obtain the first node feature vector corresponding to the first node.
5. The tinnitus diagnostic classification system according to claim 4, characterized in that, The mapping module includes: The transformation unit is used to perform a linear transformation on the aggregated features of the first node to obtain the corresponding first feature; The mapping unit is used to map the first feature to a preset feature space through a nonlinear activation function to obtain the first node feature vector corresponding to the first node.
6. The tinnitus diagnostic classification system according to claim 1, characterized in that, The node feature optimization module includes: The convolution module is used to aggregate and transform the node feature vectors of the first node and its neighboring nodes through graph convolution to generate the node update features of the first node. The filtering module is used to filter the first nodes under each preset brain map based on the node importance score to obtain important nodes, and to obtain the subgraph adjacency matrix based on the node update features corresponding to each important node. The structure learning module is used to reconstruct the adjacency matrix of each subgraph through a sparse attention mechanism to obtain the second node feature vector corresponding to each important node.
7. The tinnitus diagnostic classification system according to claim 1, characterized in that, The fusion module includes: The first fusion module is used to fuse the feature vectors of the second nodes under each preset brain map based on in-map attention to obtain the overall map representation of the preset brain map; The second fusion module is used to fuse the overall representation of all the graphs based on inter-graph attention to obtain a multi-graph feature vector.
8. The tinnitus diagnostic classification system according to claim 7, characterized in that, The first fusion module includes: The first allocation unit is used to assign a first weight to the important nodes corresponding to the feature vectors of each second node. The summation unit is used to obtain the overall representation of the preset brain map by using the first weight and the second node feature vector under each preset brain map.
9. The tinnitus diagnostic classification system according to claim 7, characterized in that, The second fusion module includes: The second allocation unit is used to allocate a second weight to the overall representation of the map under each preset brain map; The fusion unit is used to fuse all the overall representations of the graphs based on the second weight to obtain a multi-graph feature vector.
10. The tinnitus diagnostic classification system according to claim 1, characterized in that, The classification module includes: The second transformation module is used to perform a linear transformation on the multi-spectral feature vector based on a preset weight matrix to obtain a first intermediate vector; The bias processing module is used to bias the first intermediate vector to obtain the tinnitus classification corresponding to the tinnitus MRI data.