Repeated transcranial magnetic stimulation image navigation system

Abstract

The invention relates to the technical field of repeated transcranial magnetic stimulation, and particularly discloses a repeated transcranial magnetic stimulation image navigation system which is characterized in that structural magnetic resonance data of autism patients are input into an image processing and semantic understanding algorithm based on artificial intelligence and deep learning to be processed; therefore, brain structure multi-dimensional joint perception feature information in the structural magnetic resonance data of the patient is captured, and semantic segmentation is carried out based on brain structure multi-dimensional joint perception features to determine an amygdala kernel region as a region of interest. Therefore, not only can a more personalized amygdaloid nucleus region-of-interest determination scheme be provided for different autism patients, but also the intelligent level of repeated transcranial magnetic stimulation image navigation can be improved, so that the specific requirements of different patients can be better met, and a more reliable basis is provided for subsequent functional connection analysis and rTMS treatment.

Description

[0001] Repetitive transcranial magnetic stimulation imaging navigation system Technical Field

[0002] The present application relates to the technical field of repetitive transcranial magnetic stimulation, and more specifically, to a repetitive transcranial magnetic stimulation image navigation system. Background Art

[0003] Transcranial magnetic stimulation (TMS) is a non-invasive brain stimulation technique that uses changing magnetic fields to affect the activity of neurons in the cerebral cortex, thereby changing physiological processes in the brain. Repetitive transcranial magnetic stimulation (rTMS) refers to the repeated application of this stimulation at a certain frequency and pattern, and has been used to treat a variety of neurological diseases and psychiatric disorders, such as depression, anxiety, and autism. However, a challenge of traditional rTMS is the precise positioning of the stimulation target, especially for those disorders associated with complex neural networks. In autism spectrum disorder (ASD), the amygdala is an important research focus because it plays a key role in emotional processing and social behavior, which are common areas of difficulty for people with autism. Deep brain stimulation (DBS), as a treatment for autism that targets the amygdala, requires effective imaging navigation during repetitive transcranial magnetic stimulation to solve the problem that conventional rTMS cannot reach the amygdala target and DBS is invasive.

[0004] Chinese patent CN108355250B proposes a method of repetitive transcranial magnetic stimulation image navigation based on the functional circuit of the amygdala, which can use magnetic resonance data to determine the DLPFC area associated with the amygdala circuit, and use rTMS to target the DLPFC to indirectly act on the difficult-to-reach amygdala area. However, in this process, the scheme finds the amygdala area by aligning the structural automatic labeling template to the patient's structural magnetic resonance data, and uses it as the region of interest. This method relies on existing anatomical knowledge and standard templates to locate specific brain areas, but the use of standard templates for alignment may not fully take into account the anatomical differences between individuals. This is because the brain structure and functional connections of different people may vary greatly, which means that even the amygdala in the same anatomical position may have different functions in different people. Therefore, simply applying a universal template to all patients may lead to inaccurate positioning of the amygdala.

[0005] Therefore, an optimized repetitive transcranial magnetic stimulation image navigation system is desired. Summary of the invention

[0006] The present application provides a repetitive transcranial magnetic stimulation image navigation system, which can provide more personalized amygdala region of interest determination solutions for different autistic patients, and can also improve the intelligence level of repetitive transcranial magnetic stimulation image navigation, so as to better meet the specific needs of different patients and provide a more reliable foundation for subsequent functional connectivity analysis and rTMS treatment.

[0007] According to one aspect of the present application, a repetitive transcranial magnetic stimulation imaging navigation system is provided, comprising: a magnetic resonance data acquisition module, used to obtain structural magnetic resonance data and functional magnetic resonance data of autistic patients; an amygdala region of interest determination module, used to find the amygdala region in the structural magnetic resonance data and use it as the region of interest; a functional connectivity map generation module, used to pre-process the functional magnetic resonance data, and calculate the functional connectivity based on the pre-processed functional magnetic resonance data and the region of interest to obtain a functional connectivity map; a target region determination module, used to intersect the functional connectivity map with the DLPFC brain template to obtain an intersection region, and determine the target region based on the intersection region; Wherein, the amygdala region of interest determination module includes: A brain structure feature extraction unit, used for performing brain structure multi-scale feature extraction on the structural magnetic resonance data of the autistic patient to obtain shallow context coding features of the patient's brain structure and deep context coding features of the patient's brain structure; A brain structure multi-dimensional feature joint perception unit, used for performing feature joint perception under prior guidance on the shallow context coding features of the patient's brain structure and the deep context coding features of the patient's brain structure to obtain the shallow-deep joint perception features of the patient's brain structure; The amygdala region semantic segmentation unit is used to perform semantic segmentation on the shallow-deep joint perception features of the patient's brain structure to obtain a semantic segmentation result, and determine the amygdala region based on the semantic segmentation result and use it as a region of interest.

[0008] The present application provides a repetitive transcranial magnetic stimulation image navigation system, which realizes the precise positioning of the amygdala region by introducing semantic segmentation technology based on artificial intelligence and image processing, which is helpful for the subsequent generation of functional connection maps, and guides the location selection of rTMS stimulation according to the functional connection maps, that is, the target area, thereby realizing personalized and intelligent repetitive transcranial magnetic stimulation image navigation. In this way, by inputting the structural magnetic resonance data of autistic patients into the image processing and semantic understanding algorithm based on artificial intelligence and deep learning for processing, the multidimensional joint perception feature information of the brain structure in the patient's structural magnetic resonance data is captured, and semantic segmentation is performed based on the multidimensional joint perception feature of the brain structure to determine the amygdala region as the region of interest. In this way, not only can a more personalized amygdala region of interest determination scheme be provided for different autistic patients, but also the intelligence level of repetitive transcranial magnetic stimulation image navigation can be improved, so as to better meet the specific needs of different patients, and provide a more reliable basis for subsequent functional connection analysis and rTMS treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application are briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present application, and are not intended to limit the present application.

[0010] Figure 1 Schematic block diagram of a repetitive transcranial magnetic stimulation imaging navigation system according to an embodiment of the present application.

[0011] Figure 2 This is a schematic block diagram of an amygdala region of interest determination module in the repetitive transcranial magnetic stimulation image navigation system according to an embodiment of the present application.

[0012] Figure 3 This is a schematic diagram of data flow in the amygdala region of interest determination module in the repetitive transcranial magnetic stimulation image navigation system of an embodiment of the present application.

[0013] Figure 4 This is a schematic block diagram of a brain structure multi-dimensional feature joint perception unit in a repetitive transcranial magnetic stimulation imaging navigation system according to an embodiment of the present application.

[0014] Figure 5 This is a schematic block diagram of the amygdala region semantic segmentation unit in the repetitive transcranial magnetic stimulation image navigation system of an embodiment of the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work also fall within the scope of protection of the present application.

[0016] In view of the above technical problems, in the technical solution of this application, a repetitive transcranial magnetic stimulation imaging navigation system is proposed, such as Figure 1 As shown, the repetitive transcranial magnetic stimulation imaging navigation system includes: a magnetic resonance data acquisition module 10, used to obtain structural magnetic resonance data and functional magnetic resonance data of autistic patients; an amygdala region of interest determination module 20, used to find the amygdala region in the structural magnetic resonance data and use it as the region of interest; a functional connectivity map generation module 30, used to pre-process the functional magnetic resonance data, and calculate the functional connectivity based on the pre-processed functional magnetic resonance data and the region of interest to obtain a functional connectivity map; a target area determination module 40, used to intersect the functional connectivity map with the DLPFC brain template to obtain an intersection area, and determine the target area based on the intersection area.

[0017] Exemplarily, in the magnetic resonance data acquisition module 10, structural magnetic resonance data and functional magnetic resonance data of autistic patients are obtained. It should be understood that obtaining structural magnetic resonance data and functional magnetic resonance data of autistic patients provides a basis for the subsequent determination of treatment targets. In this process, a detailed medical assessment of the patient is first required to ensure that they are suitable for this type of neuroimaging examination. Once confirmed to be suitable, the patient will be arranged to a medical institution equipped with magnetic resonance imaging (MRI) equipment. The acquisition of structural magnetic resonance data is intended to capture anatomical information inside the brain, which is crucial for identifying specific brain regions such as the amygdala. At this stage, the patient will be asked to lie still in the MRI scanner so that the machine can generate high-quality cross-sectional images of the brain. These static images can reveal the location and morphological characteristics of important structures such as the cerebral cortex, white matter, and deep gray matter nuclei, which are of great significance for the subsequent analysis of the amygdala and its surrounding environment. Next, in order to understand the functional connection between different brain regions, functional magnetic resonance data needs to be collected. Unlike structural MRI, functional MRI (fMRI) focuses on brain activity patterns, especially the dynamic interactions between various brain regions when individuals perform certain tasks or are at rest. To obtain such data, the patient also needs to stay in the MRI scanner for a period of time, during which time he or she may be asked to complete some simple cognitive or sensorimotor tasks, or just relax and let the thoughts flow freely. In this way, the present application can obtain a series of time-varying blood oxygen level dependent (BOLD) signal images, which reflect the local blood flow changes caused by neuronal activity, thereby indirectly indicating the functional connectivity of the brain.

[0018] Exemplarily, in the amygdala region of interest determination module 20, the amygdala region is searched in the structural magnetic resonance data and used as a region of interest. It should be understood that the amygdala region is searched in the structural magnetic resonance data and used as a region of interest (ROI) because the amygdala plays a vital role in the emotional processing and social behavior of patients with autism. For neurodevelopmental disorders such as autism, the dysfunction of the amygdala is closely related to the symptoms such as social disorders and difficulty in emotion recognition shown by the patient. Therefore, accurately locating the amygdala is of great significance for determining the target area. Specifically, the amygdala is located in the medial temporal lobe of the brain and is part of the limbic system. It is mainly responsible for processing emotional reactions, especially negative emotions such as fear and anxiety. It is also involved in the interpretation of facial expressions, memory formation, and emotional communication in social interactions. Studies have shown that there may be structural or functional differences in the amygdala of patients with autism, which may cause them to have problems in perceiving and responding to social information. However, since the amygdala is located deep in the brain, traditional transcranial magnetic stimulation (TMS) technology is difficult to directly act on this area, and although deep brain stimulation (DBS) can achieve this goal, it is an invasive treatment method and has certain risks. In order to solve the above problems, the present application searches for the amygdala region in the structural magnetic resonance data and uses it as the region of interest.

[0019] Exemplarily, in the functional connectivity map generation module 30, after the functional magnetic resonance data is preprocessed, the functional connectivity is calculated based on the preprocessed functional magnetic resonance data and the region of interest to obtain a functional connectivity map. It should be understood that after the functional magnetic resonance data is preprocessed, the functional connectivity is calculated based on the preprocessed functional magnetic resonance data and the region of interest to obtain a functional connectivity map in order to reveal the functional connection between different brain regions by analyzing the synchronous activity patterns between them, thereby providing a scientific basis for determining the most effective rTMS treatment target. First, after obtaining the functional magnetic resonance data of autistic patients, a series of strict preprocessing steps are required. These steps include deleting the first few time points to avoid unstable signals generated when the device is started, time correction to ensure that the time series data of all voxels are consistent, head motion correction to reduce image artifacts caused by slight head movement, spatial standardization to map the individual's brain structure into a standard spatial coordinate system, smoothing to increase the signal-to-noise ratio and reduce the impact of local fluctuations, removing linear drift to eliminate possible systematic changes during long-term scanning, and bandpass filtering to retain physiological signals within a specific frequency band. After such preprocessing, the quality and reliability of subsequent data analysis can be significantly improved. Next, using the preprocessed functional magnetic resonance data and the previously determined amygdala region of interest (ROI), this application will calculate the functional connection between the two. Specifically, the time series of each voxel is extracted from the preprocessed data set, and the Pearson correlation coefficient between these sequences and the average time series in the amygdala ROI is calculated. The Pearson correlation coefficient measures the strength of the linear relationship between two variables, and is used here to evaluate the degree of synchronization of neural activity between various locations in the whole brain and the amygdala. For each voxel, the correlation value between it and the amygdala ROI is calculated and assigned to the voxel, eventually forming a functional connection map covering the entire brain. This functional connection map intuitively shows which brain regions are most closely connected to the amygdala in function, especially those that show strong functional connections in the dorsolateral prefrontal cortex (DLPFC) region. Because the amygdala plays a key role in emotional processing and social behavior, which are common areas of difficulty for people with autism, it is particularly important to find the DLPFC region that is highly correlated with the amygdala as a potential target for rTMS. Functional connectivity maps not only help identify the best therapeutic targets, but also help understand the complex network interaction mechanisms within the brain, providing valuable information for further exploring the pathophysiology behind neurodevelopmental disorders such as autism.

[0020] Exemplarily, in the target area determination module 40, the functional connection map is intersected with the DLPFC brain template to obtain an intersection area, and the target area is determined based on the intersection area. It should be understood that after completing the preprocessing of the functional magnetic resonance data of autistic patients and calculating the functional connection, the present application will obtain a map showing the functional connection strength between various locations in the whole brain and the amygdala, that is, a functional connection map. This map intuitively reflects which brain regions are most closely connected to the amygdala in function, especially those that show strong functional connections in the dorsolateral prefrontal cortex (DLPFC) region. However, in order to further refine potential therapeutic targets, it is necessary to compare this functional connection map with a standard DLPFC brain template. Specifically, the DLPFC brain template is a standard model that contains detailed anatomical information of the region, which defines the spatial range of the DLPFC and its internal subdivision structure. By intersecting the functional connection map with this template, specific areas that are both located in the DLPFC and have strong functional connections with the amygdala can be screened out. This intersection not only takes into account functional associations, but also combines accurate anatomical locations to ensure that the final selected targets are both consistent with the structural characteristics of the brain and have significant functional significance. Once the intersection area is obtained, the next step is to determine the specific targets that are most suitable for rTMS treatment. This step relies on evaluating the functional connectivity strength of each voxel in the intersection area and selecting those locations that show the strongest functional connectivity signals as priority objects.

[0021] In particular, when searching for the amygdala region as a region of interest, this solution achieves precise positioning of the amygdala region by introducing semantic segmentation technology based on artificial intelligence and image processing, which helps to generate a functional connection map in the future, and guides the location selection of rTMS stimulation, i.e., the target area, according to the functional connection map, thereby realizing personalized and intelligent repetitive transcranial magnetic stimulation image navigation. Specifically, the technical concept of this application is to capture the multi-dimensional joint perception feature information of the brain structure in the patient's structural magnetic resonance data by inputting the structural magnetic resonance data of autistic patients into an image processing and semantic understanding algorithm based on artificial intelligence and deep learning, and to determine the amygdala region as a region of interest based on semantic segmentation of the multi-dimensional joint perception features of the brain structure. In this way, not only can a more personalized amygdala region of interest determination scheme be provided for different autistic patients, but also the intelligence level of repetitive transcranial magnetic stimulation image navigation can be improved, so as to better meet the specific needs of different patients, and provide a more reliable basis for subsequent functional connection analysis and rTMS treatment.

[0022] In one embodiment, Figure 2 and Figure 3As shown, the amygdala region of interest determination module 20 includes: a brain structure feature extraction unit 21, which is used to perform brain structure multi-scale feature extraction on the structural magnetic resonance data of the autistic patient to obtain the patient's brain structure shallow context coding features and the patient's brain structure deep context coding features; a brain structure multi-dimensional feature joint perception unit 22, which is used to perform feature joint perception under prior guidance on the patient's brain structure shallow context coding features and the patient's brain structure deep context coding features to obtain the patient's brain structure shallow-deep joint perception features; an amygdala region semantic segmentation unit 23, which is used to perform semantic segmentation on the patient's brain structure shallow-deep joint perception features to obtain a semantic segmentation result, and determine the amygdala region based on the semantic segmentation result and use it as the region of interest.

[0023] Exemplarily, in the brain structure feature extraction unit 21, the structural magnetic resonance data of the autistic patient is subjected to brain structure multi-scale feature extraction to obtain shallow context coding features of the patient's brain structure and deep context coding features of the patient's brain structure. It should be understood that since the brain of an autistic patient is a highly complex organ, its structural information shows diversity at different scales. Through multi-scale feature extraction, it is possible to work at a higher spatial resolution, thereby improving the accuracy of positioning specific brain regions (such as the amygdala). This is particularly important for overcoming the challenges brought about by anatomical differences between individuals, because the brain structure of each patient may be different, and simple template matching may cause positioning errors. In one embodiment, the brain structure feature extraction unit is used to: input the structural magnetic resonance data of the autistic patient into a brain structure multi-scale feature extractor based on the FPT (Frozen Pretrained Transformer, frozen pre-trained transformer) model to obtain a shallow context coding feature map of the patient's brain structure as the shallow context coding feature of the patient's brain structure and a deep context coding feature map of the patient's brain structure as the deep context coding feature of the patient's brain structure. Here, the application of the multi-scale feature extractor of brain structure based on the FPT model can use context coding to focus on the relationship between adjacent regions in a larger range on the basis of capturing the shallow features and deep features of the brain structure of autistic patients respectively. This approach is very important for understanding complex structures (such as the amygdala and its surroundings) because it can help identify those parts that are closely connected to the target area but not easy to observe directly. In this way, the shallow context coding feature map of the patient's brain structure can capture more macroscopic brain structure feature information, such as the approximate location and morphology of the brain area; while the deep context coding feature map of the patient's brain structure can reveal more subtle local features of the brain structure, including cell arrangement, nerve fiber direction, etc. Combining these two feature maps can provide a more complete description of the amygdala and other related areas, and provide a basis for the subsequent determination of the region of interest.

[0024] Exemplarily, in the brain structure multidimensional feature joint perception unit 22, the shallow context coding features of the patient's brain structure and the deep context coding features of the patient's brain structure are subjected to feature joint perception under prior guidance to obtain the shallow-deep joint perception features of the patient's brain structure. It should be understood that the shallow context coding feature map of the patient's brain structure and the deep context coding feature map of the patient's brain structure respectively contain the shallow context semantic features and deep context semantic features of the brain structure of autistic patients. The shallow context coding features of the patient's brain structure can provide a broader overview of the brain structure, while the deep context coding features of the patient's brain structure can reveal more detailed local features of the brain structure. Combining the features of these two different aspects of brain structure context information can provide a more comprehensive and detailed description of the patient's brain structure, which is helpful to discover changes and brain structure feature information that are difficult to detect from any level alone. However, in the actual process of feature fusion, since there may be some irrelevant interference information in the shallow context coding feature map of the patient's brain structure and the deep context coding feature map of the patient's brain structure, these interference information will affect the interaction and feature expression between the shallow semantics and deep semantics of the patient's brain structure. Based on this, in the technical solution of the present application, the shallow context coding features of the patient's brain structure and the deep context coding features of the patient's brain structure are further perceived jointly under prior guidance to obtain the shallow-deep joint perception features of the patient's brain structure. Through the processing method of feature joint perception under prior guidance, the prior information in the medical field knowledge or historical data of the patient's brain structure can be integrated to enhance the model's understanding and expression capabilities of the shallow context coding features of the patient's brain structure and the deep context coding features of the patient's brain structure, and on this basis, a converter structure is introduced to perform a two-way attention interaction of the two features to generate a comprehensive feature representation that can more accurately reflect the intrinsic structure and semantic information between the shallow context coding feature map of the patient's brain structure and the deep context coding feature map of the patient's brain structure. This can establish rich two-way interactions between the shallow and deep features of the patient's brain structure, which means that even features at two locations far apart may be associated, thereby capturing spatial dependencies across multiple scales. This is particularly useful for identifying complex neural network patterns and brain structures because they often involve coordinated activities between widely distributed brain regions.

[0025] In one embodiment, Figure 4As shown, the brain structure multi-dimensional feature joint perception unit 22 includes: a shape reshaping prior guided modulation processing subunit 221, which is used to perform shape reshaping prior guided modulation processing on the patient's brain structure shallow context coding feature map and the patient's brain structure deep context coding feature map to obtain a priori modulated patient brain structure shallow context coding feature shape reshaping matrix and a priori modulated patient brain structure deep context coding feature shape reshaping matrix; a shape reshaping matrix interactive encoding subunit 222, which is used to perform shape reshaping processing on the priori modulated patient brain structure shallow context coding feature shape reshaping matrix and the priori modulated patient brain structure deep context coding feature shape reshaping matrix. Modulating the shape reshaping matrix of the deep context coding feature of the patient's brain structure for interactive coding to obtain the patient's brain structure shallow-deep context coding feature interactive coding matrix and the patient's brain structure deep-shallow context coding feature interactive coding matrix; the shape reshaping feature fusion subunit 223 is used to perform shape reshaping feature fusion on the patient's brain structure shallow-deep context coding feature interactive coding matrix and the patient's brain structure deep-shallow context coding feature interactive coding matrix to obtain the patient's brain structure shallow-deep joint perception feature map as the patient's brain structure shallow-deep joint perception feature.

[0026] In one embodiment, the shape reshaping prior guided modulation processing subunit 221 is used to: perform feature shape reshaping on the shallow context coding feature map of the patient's brain structure and the deep context coding feature map of the patient's brain structure to obtain a shallow context coding feature shape reshaping matrix of the patient's brain structure and a deep context coding feature shape reshaping matrix of the patient's brain structure. Specifically, the process can be expressed by the formula:

[0027]

[0028] in, and are respectively a shallow context encoding feature map of the patient's brain structure and a deep context encoding feature map of the patient's brain structure, Reshape the feature shape. and They are the shape reshaping matrix of the shallow context encoding features of the patient's brain structure and the shape reshaping matrix of the deep context encoding features of the patient's brain structure, respectively.

[0029] The shallow context encoding feature shape reshaping matrix of the patient's brain structure and the deep context encoding feature shape reshaping matrix of the patient's brain structure are input into the prior guided modulation module to obtain the prior modulated shallow context encoding feature shape reshaping matrix of the patient's brain structure and the prior modulated deep context encoding feature shape reshaping matrix of the patient's brain structure. Specifically, the process can be expressed as follows:

[0030]

[0031] in, and denote the first learnable memory parameter matrix and the second learnable memory parameter matrix respectively, represents the normalization function, and They are the shape reshaping matrix of shallow context encoding features of brain structure in patients with prior modulation and the shape reshaping matrix of deep context encoding features of brain structure in patients with prior modulation, respectively.

[0032] In one embodiment, the shape reshaping matrix interaction encoding subunit 222 is used to: input the prior modulated patient brain structure shallow context encoding feature shape reshaping matrix and the prior modulated patient brain structure deep context encoding feature shape reshaping matrix into the attention interaction encoding module based on the transformer-like structure to obtain the patient brain structure shallow-deep context encoding feature interaction encoding matrix and the patient brain structure deep-shallow context encoding feature interaction encoding matrix. Specifically, the process can be expressed as:

[0033]

[0034] in, The scale of the matrix is ​​reshaped to reshape the feature shape of the deep context encoding of the patient's brain structure for the prior modulation, that is, the length of the matrix is ​​multiplied by the width, is matrix multiplication, for function, and They are respectively the interactive coding matrix of the shallow-deep contextual coding features of the patient's brain structure and the interactive coding matrix of the deep-shallow contextual coding features of the patient's brain structure.

[0035] In one embodiment, the shape reshaping feature fusion subunit 223 includes: a feature shape reshaping secondary subunit, which is used to perform feature shape reshaping on the patient's brain structure shallow-deep context coding feature interaction coding matrix and the patient's brain structure deep-shallow context coding feature interaction coding matrix to obtain the patient's brain structure shallow-deep context coding feature interaction coding feature map and the patient's brain structure deep-shallow context coding feature interaction coding feature map. Specifically, the process can be expressed by the formula:

[0036]

[0037] in, and They are respectively the interactive coding feature map of the shallow-deep contextual coding features of the patient's brain structure and the interactive coding feature map of the deep-shallow contextual coding features of the patient's brain structure.

[0038] The patient's brain structure feature joint secondary subunit is used to fuse the patient's brain structure shallow-deep contextual coding feature interactive coding feature map and the patient's brain structure deep-shallow contextual coding feature interactive coding feature map to obtain the patient's brain structure shallow-deep joint perceptual feature map.

[0039] In one embodiment, the patient brain structure feature joint secondary subunit is used to calculate the weighted sum between the patient brain structure shallow-deep contextual coding feature interactive coding feature map and the patient brain structure deep-shallow contextual coding feature interactive coding feature map to obtain the patient brain structure shallow-deep joint perceptual feature map. Specifically, the process can be expressed as follows:

[0040] in, and are the first weight coefficient and the second weight coefficient respectively, It is a joint perceptual feature map of the superficial and deep layers of the patient's brain structure.

[0041] Specifically, the processing of feature joint perception under prior guidance is to reshape the shallow context encoding feature map of the patient's brain structure and the deep context encoding feature map of the patient's brain structure respectively and then perform prior guidance modulation, so as to implement targeted selection and enhancement at the feature level by introducing prior information, improve the model's ability to capture key shallow feature information of the patient's brain structure and deep feature information of the patient's brain structure, thereby improving the subsequent semantic segmentation accuracy. Further, by dynamically capturing the complex interaction between the shallow context encoding features of the patient's brain structure and the deep context encoding features of the patient's brain structure with the help of the self-attention mechanism, the semantic segmentation and amygdala region identification schemes have stronger generalization ability and sensitivity to subtle brain structure changes. This helps to more comprehensively and accurately reflect the true situation of the patient's brain structure, including those subtle but important brain structure variations and personalized features. In particular, it is worth mentioning that as deep learning models become more and more complex, it becomes increasingly difficult to understand their decision-making process. By explicitly incorporating professional knowledge or historical data in the medical field, the transparency of model behavior is increased, which helps to better understand and optimize the feature expression of the patient's brain structure. For example, we know that certain brain regions (such as the amygdala) are closely related to social behavior and may show abnormalities in autism. Therefore, corresponding prior rules can be set to emphasize the importance of these regions, ensuring that the model can pay more attention to these key parts, thereby providing a basis for subsequent semantic segmentation tasks and region of interest determination tasks.

[0042] Exemplarily, in the amygdala region semantic segmentation unit 23, the shallow-deep joint perception features of the patient's brain structure are semantically segmented to obtain a semantic segmentation result, and the amygdala region is determined based on the semantic segmentation result and used as a region of interest. It should be understood that semantic segmentation is an image processing method that can assign each pixel in an image to a specific category or label, such as different brain tissue types. For MRI data of brain structures, semantic segmentation can help distinguish gray matter, white matter, and specific anatomical structures such as the amygdala. Compared with traditional template matching methods, semantic segmentation can better adapt to anatomical differences between individuals because it does not rely on fixed standard templates to locate specific brain regions, but rather performs personalized segmentation based on the specific imaging data of each patient. Therefore, the amygdala region is automatically identified and marked by semantic segmentation of the shallow-deep joint perception features.

[0043] In one embodiment, Figure 5As shown, the amygdala region semantic segmentation unit 23 includes: a semantic segmentation result determination subunit 231, which is used to convert the shallow-deep joint perception feature map of the patient's brain structure into a feature matrix and then perform semantic segmentation to obtain the semantic segmentation result; a region of interest determination subunit 232, which is used to determine the amygdala region based on the semantic segmentation result and use it as the region of interest.

[0044] In one embodiment, the semantic segmentation result determination subunit 231 is used to: perform maximum pooling along the channel dimension on the shallow-deep joint perceptual feature map of the patient's brain structure to obtain a shallow-deep joint significant perceptual feature matrix of the patient's brain structure; input the shallow-deep joint significant perceptual feature matrix of the patient's brain structure into a semantic segmenter based on a softmax classification function to obtain the semantic segmentation result.

[0045] It should be understood that the maximum pooling process along the channel dimension means extracting the strongest activation pattern from the shallow-deep joint perceptual feature map of the patient's brain structure in multiple channels. These activation patterns are usually closely related to the functional or anatomical characteristics of specific brain regions, so they can be used to highlight information that is crucial for identifying the amygdala. Then, the feature matrix after maximum pooling is input into the semantic segmenter based on the softmax classification function, with the aim of classifying different types of brain tissue or functional areas at the pixel level. The softmax function can convert the output into a probability distribution, so that the possibility of each pixel belonging to the amygdala region can be quantified. Then, based on the semantic segmentation results, the amygdala region is determined and used as the region of interest. In this way, not only can a more personalized amygdala region of interest determination scheme be provided for different autistic patients, but also the intelligence level of repeated transcranial magnetic stimulation image navigation can be improved, so as to better meet the specific needs of different patients, and provide a more reliable basis for subsequent functional connectivity analysis and rTMS treatment.

[0046] Preferably, inputting the shallow-deep joint significant perceptual feature matrix of the patient's brain structure into a semantic segmenter based on a softmax classification function to obtain a semantic segmentation result includes: The absolute value sum of all eigenvalues ​​of the shallow-deep joint significant perception feature matrix of the patient's brain structure is calculated to obtain the shallow-deep joint significant perception spatial structure value of the first patient's brain structure, and the square root of the sum of squares is calculated to obtain the shallow-deep joint significant perception spatial structure value of the second patient's brain structure, that is:

[0047]

[0048] in, The first one represents the joint significant perceptual feature matrix of the shallow-deep layer of the patient's brain structure eigenvalues, represents the total number of eigenvalues ​​of the shallow-deep joint significant perceptual feature matrix of the patient's brain structure, It represents the joint significant perceived spatial structure value of the superficial and deep layers of the brain structure of the first patient. Indicates the second patient's brain structure superficial-deep joint significant perceived spatial structure value; For each eigenvalue of the shallow-deep joint significant perception feature matrix of the patient's brain structure, the shallow-deep joint significant perception spatial structure value of the first patient's brain structure minus the product of the eigenvalue and the total number of eigenvalues ​​of the shallow-deep joint significant perception feature matrix of the patient's brain structure is calculated to obtain the shallow-deep joint significant perception long-range dependency value of the first patient's brain structure ,in, It represents the joint significant perceived spatial structure value of the superficial and deep layers of the brain structure of the first patient. The first one represents the joint significant perceptual feature matrix of the shallow-deep layer of the patient's brain structure eigenvalues, represents the total number of eigenvalues ​​of the shallow-deep joint significant perceptual feature matrix of the patient's brain structure, Indicates the long-range dependency value of the combined significant perception of the superficial and deep layers of the brain structure of the first patient; Calculate the square root of the total number of eigenvalues ​​multiplied by the product of the eigenvalues ​​minus the second patient's brain structure shallow-deep joint significant perception spatial structure value to obtain the second patient's brain structure shallow-deep joint significant perception long-range dependency value ,in, It represents the joint significant perceived spatial structure value of the superficial and deep layers of the brain structure of the second patient. The first one represents the joint significant perceptual feature matrix of the shallow-deep layer of the patient's brain structure eigenvalues, represents the total number of eigenvalues ​​of the shallow-deep joint significant perceptual feature matrix of the patient's brain structure, Indicates the second patient's brain structure shallow-deep joint significant perception long-range dependency value; The index value calculated by taking the shallow-deep joint significant perceptual long-range dependency value of the first patient's brain structure as the exponent of the natural constant and the inverse of the shallow-deep joint significant perceptual long-range dependency value of the second patient's brain structure are weighted summed to obtain the optimized eigenvalue corresponding to each eigenvalue ,in, Indicates the long-range dependency value of the shallow-deep joint perception of the first patient's brain structure, Indicates the second patient's brain structure shallow-deep joint significant perception long-range dependency value, represents a natural constant, and represents the weighted hyperparameter, Indicates the optimized eigenvalue corresponding to each eigenvalue; The optimized shallow-deep joint significant perceptual feature matrix of the patient's brain structure composed of the optimized feature values ​​is input into a semantic segmenter based on a softmax classification function to obtain a semantic segmentation result.

[0049] Here, when the shallow context encoding feature map of the patient's brain structure and the deep context encoding feature map of the patient's brain structure respectively represent image semantic features of different depths and scales of the structural magnetic resonance data of autistic patients, after performing feature significant joint perception based on prior guidance, the shallow-deep joint perception feature map of the patient's brain structure is subjected to global maximum pooling along the channel dimension to obtain a shallow-deep joint significant perception feature matrix of the patient's brain structure. This will also have interactive fusion distribution spatial structure differences due to the misalignment of cross-scale and cross-depth image semantic space distributions, affecting the semantic segmentation convergence consistency, thereby affecting the accuracy of the semantic segmentation results obtained by the semantic segmentator based on the softmax classification function.

[0050] Based on this, in order to solve the problem that the feature set of the shallow-deep joint salient perceptual feature matrix of the patient's brain structure may have spatial structure loss in the high-dimensional space, which leads to the convergence inconsistency of the image semantic segmentation based on the implicit inference of spatial structure information based on the features, a long-distance feature dependency relationship is established based on the overall feature scale of the shallow-deep joint salient perceptual feature matrix of the patient's brain structure relative to the spatial structure representation of the shallow-deep joint salient perceptual feature matrix of the patient's brain structure, so as to establish the feature local connectivity of the shallow-deep joint salient perceptual feature matrix of the patient's brain structure, and capture the spatial ambiguous information of the object feature value through the unstructured eigenvalue point prediction of the shallow-deep joint salient perceptual feature matrix of the patient's brain structure, thereby improving the spatial inductive bias perception ability of the feature set of the shallow-deep joint salient perceptual feature matrix of the patient's brain structure, improving the convergence consistency of semantic segmentation, and improving the accuracy of the semantic segmentation result obtained by inputting the shallow-deep joint salient perceptual feature matrix of the patient's brain structure into a semantic segmentor based on the softmax classification function. In this way, not only can a more personalized amygdala region of interest determination plan be provided for different autistic patients, but the intelligence level of repetitive transcranial magnetic stimulation image navigation can also be improved, so as to better meet the specific needs of different patients.

[0051] In summary, the repetitive transcranial magnetic stimulation imaging navigation system according to the embodiment of the present application is explained, which captures the multi-dimensional joint perception feature information of the brain structure in the patient's structural magnetic resonance data by inputting the structural magnetic resonance data of the autistic patient into the image processing and semantic understanding algorithm based on artificial intelligence and deep learning for processing, and performs semantic segmentation based on the multi-dimensional joint perception feature of the brain structure to determine the amygdala region as the region of interest. In this way, not only can a more personalized amygdala region of interest determination scheme be provided for different autistic patients, but also the intelligence level of repetitive transcranial magnetic stimulation imaging navigation can be improved, so as to better meet the specific needs of different patients, and provide a more reliable foundation for subsequent functional connectivity analysis and rTMS treatment.

[0052] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0053] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0054] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the present invention and the scope of protection of the present invention, all of which are within the protection of the present invention.

Claims

1. A repetitive transcranial magnetic stimulation imaging navigation system, characterized in that: include: A magnetic resonance data acquisition module, used to obtain structural magnetic resonance data and functional magnetic resonance data of autistic patients; an amygdala region of interest determination module, used to find the amygdala region in the structural MRI data and use it as the region of interest; a functional connectivity map generation module, used to pre-process the functional MRI data, calculate the functional connectivity based on the pre-processed functional MRI data and the region of interest to obtain a functional connectivity map; a target region determination module, used to intersect the functional connectivity map with the DLPFC brain template to obtain an intersection region, and determine the target region based on the intersection region; Wherein, the amygdala region of interest determination module includes: A brain structure feature extraction unit, used for performing brain structure multi-scale feature extraction on the structural magnetic resonance data of the autistic patient to obtain shallow context coding features of the patient's brain structure and deep context coding features of the patient's brain structure; A brain structure multi-dimensional feature joint perception unit, used for performing feature joint perception under prior guidance on the shallow context coding features of the patient's brain structure and the deep context coding features of the patient's brain structure to obtain the shallow-deep joint perception features of the patient's brain structure; The amygdala region semantic segmentation unit is used to perform semantic segmentation on the shallow-deep joint perception features of the patient's brain structure to obtain a semantic segmentation result, and determine the amygdala region based on the semantic segmentation result and use it as a region of interest.

2. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 1, characterized in that: The brain structure feature extraction unit is used to: input the structural magnetic resonance data of the autistic patient into a brain structure multi-scale feature extractor based on the FPT model to obtain a shallow context coding feature map of the patient's brain structure as the shallow context coding feature of the patient's brain structure and a deep context coding feature map of the patient's brain structure as the deep context coding feature of the patient's brain structure.

3. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 2, characterized in that: The brain structure multi-dimensional feature joint perception unit includes: A shape reshaping prior-guided modulation processing subunit is used to perform shape reshaping prior-guided modulation processing on the patient's brain structure shallow context coding feature map and the patient's brain structure deep context coding feature map to obtain a priori modulated patient brain structure shallow context coding feature shape reshaping matrix and a priori modulated patient brain structure deep context coding feature shape reshaping matrix; A shape reshaping matrix interactive coding subunit is used to interactively code the shape reshaping matrix of the prior modulated patient's brain structure shallow context coding feature and the shape reshaping matrix of the prior modulated patient's brain structure deep context coding feature to obtain a patient's brain structure shallow-deep context coding feature interactive coding matrix and a patient's brain structure deep-shallow context coding feature interactive coding matrix; The shape reshaping feature fusion subunit is used to perform shape reshaping feature fusion on the patient's brain structure shallow-deep contextual coding feature interaction coding matrix and the patient's brain structure deep-shallow contextual coding feature interaction coding matrix to obtain the patient's brain structure shallow-deep joint perception feature map as the patient's brain structure shallow-deep joint perception feature.

4. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 3, characterized in that: The shape reshaping prior-guided modulation processing subunit is used to: Performing feature shape reshaping on the shallow context coding feature map of the patient's brain structure and the deep context coding feature map of the patient's brain structure to obtain a shallow context coding feature shape reshaping matrix of the patient's brain structure and a deep context coding feature shape reshaping matrix of the patient's brain structure; The shallow context encoding feature shape reshaping matrix of the patient's brain structure and the deep context encoding feature shape reshaping matrix of the patient's brain structure are input into the prior guided modulation module to obtain the prior modulated shallow context encoding feature shape reshaping matrix of the patient's brain structure and the prior modulated deep context encoding feature shape reshaping matrix of the patient's brain structure.

5. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 4, characterized in that: The shape reshaping matrix interaction coding subunit is used to: input the prior modulated patient brain structure shallow context coding feature shape reshaping matrix and the prior modulated patient brain structure deep context coding feature shape reshaping matrix into the attention interaction coding module based on the transformer-like structure to obtain the patient brain structure shallow-deep context coding feature interaction coding matrix and the patient brain structure deep-shallow context coding feature interaction coding matrix.

6. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 5, characterized in that: The shape reshaping feature fusion subunit comprises: The feature shape reshaping secondary subunit is used to perform feature shape reshaping on the patient's brain structure shallow-deep context coding feature interaction coding matrix and the patient's brain structure deep-shallow context coding feature interaction coding matrix to obtain a patient's brain structure shallow-deep context coding feature interaction coding feature map and a patient's brain structure deep-shallow context coding feature interaction coding feature map; The patient's brain structure feature joint secondary subunit is used to fuse the patient's brain structure shallow-deep contextual coding feature interactive coding feature map and the patient's brain structure deep-shallow contextual coding feature interactive coding feature map to obtain the patient's brain structure shallow-deep joint perceptual feature map.

7. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 6, characterized in that: The patient's brain structure feature joint secondary subunit is used to calculate the weighted sum between the patient's brain structure shallow-deep contextual coding feature interactive coding feature map and the patient's brain structure deep-shallow contextual coding feature interactive coding feature map to obtain the patient's brain structure shallow-deep joint perceptual feature map.

8. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 7, characterized in that: The amygdala region semantic segmentation unit comprises: A semantic segmentation result determination subunit, used for converting the shallow-deep joint perceptual feature map of the patient's brain structure into a feature matrix and then performing semantic segmentation to obtain the semantic segmentation result; The region of interest determination subunit is used to determine the amygdala region based on the semantic segmentation result and use it as the region of interest.

9. The repetitive transcranial magnetic stimulation imaging navigation system according to claim 8, characterized in that: The semantic segmentation result determination subunit is used to: Performing maximum pooling along the channel dimension on the shallow-deep joint perceptual feature map of the patient's brain structure to obtain a shallow-deep joint significant perceptual feature matrix of the patient's brain structure; The shallow-deep joint significant perceptual feature matrix of the patient's brain structure is input into a semantic segmenter based on a softmax classification function to obtain the semantic segmentation result.

Images