Abnormal region boundary determination method

By extracting and fitting the neural signals obtained by array electrodes, combined with the image segmentation model of tomography images, the problem of not being able to accurately identify lesion boundaries in the existing technology is solved, and the precise identification and image segmentation model of lesion boundaries are realized. optimization of accuracy.

CN118941578APending Publication Date: 2024-11-12WUHAN NEURACOM TECH DEV CO LTD
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202410980808.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The prior art cannot accurately identify the boundaries of the lesion, and can only lock the part where the lesion is located, and cannot accurately identify it.

Method used

By extracting and fitting the neural signals obtained from array electrodes, combining the image segmentation model of tomography images, the loss function is constructed and optimized by gradient descent method to obtain the boundary point coordinates of the abnormal regions of the modified neural signal, and iteratively optimized Image segmentation model to determine the boundaries of abnormal areas.

Benefits of technology

It realizes accurate identification of lesion boundaries, improves the spatial accuracy of neural signal, optimizes the accuracy of image segmentation model, and has important medical diagnostic significance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941578A_ABST
    Figure CN118941578A_ABST
Patent Text Reader

Abstract

The invention provides an abnormal region boundary determination method, and the method comprises the steps: carrying out the analysis and processing of a neural signal, obtained by an array electrode, of a target object, and obtaining the boundary point coordinates of a neural signal abnormal region; carrying out segmentation through a tomography image of the target object and a deep learning image segmentation model to obtain boundary point coordinates of an image abnormal region; according to the coordinate of the boundary point of the image abnormal region, optimizing the boundary determination parameter of the neural signal abnormal region, and continuing to optimize the image segmentation model; and comparing a pixel point region with the highest confidence coefficient of the re-optimized image segmentation model with a neuron clustering result, stopping iteration when a comparison result error is smaller than a preset expected threshold value, and determining a boundary of an abnormal region of the target object through image segmentation of the image segmentation model. According to the method, through mutual optimization and iteration of two boundary determination ideas of neural signals and a tomography image + image segmentation model, accurate recognition of the abnormal region boundary can be realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of medical assistance technology, and more specifically, to a method for determining the boundary of an abnormal area. Background Art

[0002] A lesion is a part of a body tissue or organ where abnormal changes or pathological changes occur. When a part of the body is damaged by bacteria or viruses, a nest-like area is formed, where pathogens live. These pathogens invade the tissue, causing the corresponding tissue to die and the lesion to appear. This lesion is called a lesion.

[0003] The existing technology includes technologies that collect, identify and analyze the target's neural consciousness through neural signal collection, thereby assisting the target to complete its conscious behavior, and there are also related technologies for lesion identification based on neural signals.

[0004] Currently, it is possible to diagnose epilepsy / Parkinson's disease based on EEG or to collect abnormal neural signals based on SEEG, but such methods can only lock the location of the lesion and cannot accurately identify the boundaries of the lesion itself. Summary of the invention

[0005] In response to at least one defect or improvement need of the prior art, the present application provides a method for determining the boundary of an abnormal region, which is used to at least accurately identify the boundary of the lesion itself.

[0006] To achieve the above objectives, the present application provides a method for determining the boundary of an abnormal area, comprising:

[0007] Perform feature extraction and feature fitting on the neural signals of the target object acquired by the array electrodes, and perform smoothing on the acquired abnormal recognition results, eliminate channels without abnormal intervals, and obtain the average confidence of all the retained abnormal channels;

[0008] Performing straight line fitting processing on the average confidence of the abnormal channel and the corresponding channel point coordinates, and obtaining the boundary points of the abnormal area based on a preset confidence threshold;

[0009] Performing interpolation processing on the boundary points of the abnormal area to obtain corrected coordinates of the boundary points of the abnormal area of ​​the neural signal;

[0010] Acquire a tomographic image of the target object, perform landmark segmentation through an image segmentation model to obtain corresponding coordinates of boundary points of abnormal image regions, and construct a loss function based on the coordinates of boundary points of abnormal image regions and the coordinates of boundary points of abnormal neural signal regions;

[0011] Iteratively optimizing the loss function using a gradient descent method to obtain optimized neural signal abnormal region boundary determination parameters related to the smoothing process, straight line fitting process, and interpolation process;

[0012] Performing threshold method spike segment extraction processing on the neural signal of the target object acquired by the array electrode, clustering the spike segments to obtain the coordinates of each neuron, adding the coordinates of the neuron to the corresponding coordinates of the boundary points of the abnormal area of ​​the image to form a new training set image, and retraining and optimizing the image segmentation model based on the new training set image;

[0013] Based on the segmentation boundary of the optimized image segmentation model, the loss function is re-iteratively optimized, and the image segmentation model is continuously re-optimized, and the pixel point area with the highest confidence in the re-optimized image segmentation model is compared with the neuron clustering result. When the comparison result error is less than the preset expected value threshold, the iteration is stopped, and the boundary of the abnormal area of ​​the target object is determined by image segmentation of the image segmentation model.

[0014] Furthermore, the feature extraction of the neural signal of the target object acquired by the array electrode includes:

[0015] Cut each channel on each probe of the array electrode into segments based on a sliding window, and obtain a plurality of segments for each channel;

[0016] Perform time domain feature extraction and frequency domain feature extraction on the frequency domain signal of each segment of each channel;

[0017] Fourier transform is performed on the time domain signal of each segment of each channel to obtain a time-frequency image, and image features are extracted from the time-frequency image.

[0018] Further, performing the feature fitting includes:

[0019] The features extracted from the frequency domain signal and the time domain signal are sent as training data to a machine learning model for model training to obtain an optimal model for abnormality recognition and the selected features;

[0020] Each channel on each probe of the array electrode is segmented, feature extraction is performed according to the selected features, the trained best abnormality recognition model is loaded, and the abnormality recognition result of the current segment and the confidence of the model output are obtained.

[0021] Furthermore, the obtained abnormal recognition results are smoothed, channels without abnormal intervals are eliminated, and the average confidence of all the retained abnormal channels is obtained, including:

[0022] In the abnormal identification result, if the identification result of at least the second consecutive data segments among the first consecutive data segments is abnormal, then mark the first consecutive data segments as abnormal intervals;

[0023] The smoothing mechanism continues until at least the second consecutive data segments in the first consecutive data segments have normal recognition results, then this interval result is overturned as a normal interval, otherwise it continues to be an abnormal interval, and the start and end time of each channel abnormal interval is recorded; the first data is greater than the second data;

[0024] For each channel on each probe, the mean interval of the start and end time of each abnormal interval is calculated respectively. If the overlap rate between the start and end time of the abnormal interval and the mean interval of the start and end time is less than a preset overlap rate threshold, the abnormal interval is reset to a normal interval.

[0025] Eliminate channels without abnormal intervals, obtain all the retained abnormal channels and obtain the corresponding average confidence.

[0026] Furthermore, performing straight line fitting processing on the average confidence of the abnormal channel and the corresponding channel point coordinates, and obtaining the boundary points of the abnormal area based on a preset confidence threshold comprises:

[0027] Performing a straight line fitting on the average confidence of the abnormal channel and the corresponding channel point coordinates by the least square method to obtain a corresponding straight line fitting equation;

[0028] The confidence threshold is substituted into the straight line fitting equation as a dependent variable to obtain boundary points on all transverse channels and longitudinal channels of the abnormal area.

[0029] Furthermore, the interpolation processing of the boundary points of the abnormal region to obtain the corrected coordinates of the boundary points of the abnormal region of the neural signal includes: one or more correction methods of singular value correction, whole line missing correction and boundary point density increase correction;

[0030] Singular value correction:

[0031] For every three adjacent boundary points A, B and C on the horizontal or vertical channels, a group of judgments is made. If the corresponding horizontal or vertical coordinates x A 、x B and x C A value in When the deviation exceeds the preset boundary point correction threshold β, the coordinate is marked as a singular coordinate and needs to be corrected;

[0032] Draw a straight line through points A and C, and intersect the extension of the straight line of point B at point O. Take a point B' on the line connecting BO. When , point B' is the updated coordinate position of point B;

[0033] Whole line missing fix:

[0034] After the singular value correction is completed, take the midpoint F of the line segment AG, and draw a straight line through point F parallel to the straight line where point A is located; the point G is a boundary point adjacent to point A in a direction different from that of point C;

[0035] Boundary point density increase correction:

[0036] After the entire row missing is corrected, the midpoint of line segment B'C is taken as a newly added boundary point.

[0037] Furthermore, the tomographic image of the target object is obtained, and the landmark segmentation is performed through the image segmentation model to obtain the corresponding coordinates of the boundary points of the abnormal image area, and the loss function is constructed based on the coordinates of the boundary points of the abnormal image area and the coordinates of the boundary points of the abnormal neural signal area, including:

[0038] Slicing the three-dimensional magnetic resonance image according to the layer where the vertical channel of the array electrode is located to obtain a plurality of two-dimensional slices;

[0039] Using the image segmentation model, the two-dimensional slices are respectively segmented for the markers, and the coordinates of the boundary points of the image abnormal area corresponding to the boundary of the markers in each slice layer are respectively obtained;

[0040] Gridding the cross-sections of the array electrodes in the vertical, horizontal and longitudinal directions, respectively obtaining the Euclidean distances between the boundary points of each abnormal image region and the center points of each grid, and performing statistics to obtain the grid where the boundary points of each abnormal image region are located;

[0041] The loss function is constructed, and its specific formula is as follows:

[0042]

[0043] Where Num represents the number of layers of the two-dimensional slice; Count represents the number of grids in the current slice layer; K1 represents the number of coordinates obtained by image segmentation in the current grid; K2 represents the number of coordinates obtained by neural signals in the current grid; x fj_mean and fj_mean They respectively represent the average values ​​of the horizontal coordinates and the vertical coordinates of the boundary points of the abnormal image region corresponding to the boundary of the landmark obtained by image segmentation by the image segmentation model in the j-th layer cross section; j_mean and j_mean They respectively represent the average abscissa and average ordinate of the corresponding corrected boundary points of the abnormal neural signal region obtained through electrode signal analysis in the j-th layer cross section.

[0044] Furthermore, the gradient descent method is used to iteratively optimize the loss function to obtain optimized neural signal abnormal region boundary determination parameters related to the smoothing process, straight line fitting process and interpolation process, including:

[0045] Define the objective function, loss = 0;

[0046] Adopting a back-propagation control method, taking one or more of the first data, the second data, the coincidence rate threshold, the intercept and slope of the straight line fitting equation, the confidence threshold and the boundary point correction threshold as weight parameters;

[0047] The gradient descent method is used as the optimization function. When loss is less than 0.001, the iterative optimization can be terminated to obtain the corresponding boundary determination parameters of all the abnormal neural signal regions at this time.

[0048] Furthermore, the neural signals of the target object acquired by the array electrodes are subjected to a threshold method spike segment extraction process, and the spike segments are clustered to obtain the coordinates of each neuron, and the coordinates of the neuron are added to the corresponding coordinates of the boundary points of the abnormal image area to form a new training set image, and the image segmentation model is retrained and optimized based on the new training set image, including:

[0049] Extract spike fragments from the target object's neural signals acquired by the array electrodes using the threshold method, obtain the equivalent coordinates of each spike fragment, perform neuron clustering on the spike fragments, and then calculate the average of the coordinates of all spike fragments of the same neuron to obtain the coordinates of each neuron;

[0050] Add the coordinate point of each neuron vertically to the coordinate position corresponding to the slice layer of the tomographic image to obtain a new training set image;

[0051] The image segmentation model is retrained and optimized based on the new training set images.

[0052] Further, the continuing to optimize the image segmentation model includes:

[0053] Based on the iteratively optimized loss function, optimizing the abnormal neural signal region recognition results and neuron clustering results;

[0054] The image segmentation model is further optimized based on the optimized neural signal abnormal area recognition results and neuron clustering results.

[0055] In general, the above technical solutions conceived by the present application can achieve the following beneficial effects compared with the prior art:

[0056] The present application proposes a method for determining the boundary of an abnormal region, which analyzes and processes the neural signals of the target object obtained by the array electrode to obtain the coordinates of the boundary points of the modified neural signal abnormal region; then obtains the coordinates of the boundary points of the image abnormal region by segmenting the tomographic image of the target object + deep learning image segmentation model; according to the coordinates of the boundary points of the image abnormal region, optimizes the parameters for determining the boundary of the neural signal abnormal region, and continues to optimize the image segmentation model; compares the pixel point area with the highest confidence of the re-optimized image segmentation model with the neuron clustering result, stops iteration when the comparison result error is less than the preset expected value threshold, and determines the boundary of the abnormal region of the target object by image segmentation of the image segmentation model. The method can realize accurate identification of the boundary of the abnormal region through mutual optimization and iteration of the two abnormal region boundary determination ideas of neural signal and tomographic image + image segmentation model. The spatial accuracy of neural signals is at the micron level, while the accuracy of MRI tomographic images is at the millimeter level. The present application optimizes the accuracy of the image algorithm segmentation model again with the spatial accuracy of neural signals, which is of great significance for the accurate identification of the boundary of the body lesion area. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0058] Figure 1 A flowchart of a method for determining an abnormal area boundary provided in an embodiment of the present application;

[0059] Figure 2 A schematic diagram of the size parameters of the array electrodes provided in the embodiment of the present application;

[0060] Figure 3 A smoothing logic schematic diagram provided for an embodiment of the present application;

[0061] Figure 4 A schematic diagram of boundary determination provided in an embodiment of the present application;

[0062] Figure 5 A schematic diagram of boundary coordinate correction provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0064] The terms "first", "second" or "nth" in the specification, claims or the above-mentioned drawings of this application are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices.

[0065] As described in the background technology section of the specification, currently, it is possible to diagnose epilepsy / Parkinson's disease based on EEG or collect abnormal neural signals based on SEEG, but such methods can only lock the location of the lesion and cannot accurately identify the boundary of the lesion itself. In view of this, the present application proposes a method for determining the boundary of an abnormal region, which is used to at least accurately identify the boundary of the lesion itself (i.e., the boundary of the abnormal region).

[0066] refer to Figure 1 One embodiment of the present application provides a method for determining the boundary of an abnormal region. The method can be used to identify whether there are markers (markers that will produce abnormal discharge signals, such as epileptogenic foci of epilepsy) in the brain (or other parts of the body) by analyzing the neural signals collected by the array electrodes. The method is a method for identifying the boundary of the marker based on the abnormal gradient line of the neural signal; the method can also be combined with the diagnostic results of other methods (such as pathological / MRI image analysis results) to modify the above judgment logic. The method can specifically include the following steps.

[0067] Step 1: extract and fit the features of the neural signals of the target object acquired by the array electrodes, smooth the acquired abnormal recognition results, remove the channels without abnormal intervals, and obtain the average confidence of all the abnormal channels retained. In some embodiments, step 1 may also specifically include:

[0068] Step 1.1: Cut each channel into segments based on the sliding window, and obtain Q segments for each channel. The array electrode is M×N×K, where M is the number of probe rows, N is the number of probe columns, and K is the number of channels on each probe (the number of channels is the number of sensors, Figure 2 The tip on the left is a probe. Figure 2 The picture on the right shows a probe, and each small dot on it is a sensor, or a channel).

[0069] Step 1.2: Extract the time domain features of the frequency domain signal of each segment of each channel. tpq , such as maximum value, mean value, standard deviation, margin index and peak index, where p represents the pth channel and q represents the qth segment.

[0070] Step 1.3: Extract frequency domain features of the frequency domain signal of each segment of each channel. fpq , such as center frequency, frequency variance, spectral distance and wavelet standard deviation.

[0071] Step 1.4: Perform Fourier transform on the time domain signal of each segment of each channel to obtain the time-frequency image, and perform image feature extraction label on the time-frequency image gpq , such as image color histogram, color principal components, etc.

[0072] Step 1.5: Based on the existing data sets, train the anomaly recognition model in advance: label each segment in step 1.1 with labels 0-normal and 1-abnormal, and then send all the extracted features in steps 1.2 to 1.4 to machine learning models such as SVM / DT / RF for model training to obtain the best anomaly recognition model (such as RF) and the selected features (such as time domain mean, time domain peak index, frequency domain frequency variance, time-frequency diagram color principal component, etc.).

[0073] Step 1.6: In actual use, cut each channel into segments according to step 1.1, extract features according to the features selected in step 1.5, load the best anomaly recognition model trained in step 1.5 to obtain the anomaly recognition result of the current segment. pq And the model output confidence score of the current segment pq .

[0074] Step 1.7: Smooth the abnormal recognition results of each channel and each segment, such as Figure 3 As shown in the figure, if at least w consecutive segments among v consecutive segments have abnormal recognition results, then it is assumed that these v consecutive segments are all abnormal segments, and the smoothing mechanism continues until ≥w segments among v segments have normal results, then the result of this interval is overturned as a normal interval, otherwise it continues to be an abnormal interval, and the start and end time of each channel abnormal interval [t pu0 ,t pu1], where p represents the p-th channel, u represents the u-th abnormal interval, 0 represents the start mark of the abnormal interval, and 1 represents the end mark of the abnormal interval.

[0075] Step 1.8: For each channel on each probe, calculate the mean start time and end time of each abnormal interval to obtain [t0, t1], and calculate the time overlap rate τ between the start and end time of each abnormal interval and [t0, t1] pu , if τ pu <α, then the interval is a false abnormal interval, and the reset result is a normal interval.

[0076] Step 1.9: Based on step 1.8, remove channels without an abnormal interval (i.e., remove all normal channels and retain those that are not all normal). For all the abnormal channels that are retained, calculate the average confidence of the channel model.

[0077] Step 2: Perform straight line fitting processing on the average confidence of the abnormal channel and the corresponding channel point coordinates, and obtain the boundary points of the abnormal area based on a preset confidence threshold. In some embodiments, step 2 may further include:

[0078] Step 2.1, Reference Figure 4 , along a certain vertical height h i The position of the j-th layer longitudinal channel is taken as an example. The channel model confidence obtained in step 1.9 is Where i is the i-th vertical channel, j is the j-th vertical channel, and k is the k-th horizontal channel. The corresponding channel point coordinates are x ijk .

[0079] Step 2.2: Use the least squares method to fit an approximate straight line l1:y=k1·x+b1, where the channel point coordinates correspond to the independent variable x and the confidence corresponds to the dependent variable y.

[0080] Step 2.3, set the boundary conditions of the lateral channel. When the confidence is lower than s, it is the boundary point of the current fitting line. Substitute s into l1 and find the boundary coordinate x ijs1 and x ijs2 .

[0081] Step 2.4: Refer to steps 2.1 to 2.3 to find the boundary points on each transverse channel.

[0082] Step 2.5: Refer to Step 2.1 to Step 2.4 to obtain the boundary points on all longitudinal channels.

[0083] Step 3: interpolate the boundary points of the abnormal area to obtain the corrected coordinates of the boundary points of the abnormal area of ​​the neural signal. In some embodiments, step 3 may further include:

[0084] Taking the horizontal boundary point correction as an example, it is divided into three cases: singular value correction, whole row missing correction and boundary point density increase correction.

[0085] Step 3.1: Singular value correction. Figure 5 When the coordinate value of the boundary point is a singular value, a group of judgments is made for each three adjacent boundary points in the abnormal area, such as points A / B / C. When the corresponding coordinates x A / x B / x C A value in When the deviation exceeds β, the coordinate is considered to be a singular coordinate and needs to be corrected. Connect AC and intersect the extension line of point B at point O. Take a point B' on the line connecting BO. , then B' is the updated coordinate of point B.

[0086] Step 3.2: Correct the missing line. After step 3.1 is completed, connect points G and A, and draw a straight line through the midpoint F of GA parallel to the straight line where point G is located or the straight line where point A is located.

[0087] Step 3.3, increase the density of boundary points. After step 3.2 is completed, add a boundary point B'C' between point B' and point C, that is, take the midpoint of the line connecting point B' and point C as the newly added point B'C'.

[0088] Step 3.4: The longitudinal boundary point correction refers to the transverse boundary point correction, and finally the current vertical h is obtained. i All boundary points (x i ,y i ).

[0089] Step 4, obtaining a tomographic image of the target object, performing landmark segmentation through an image segmentation model to obtain corresponding coordinates of the boundary points of the abnormal image area, and constructing a loss function based on the coordinates of the boundary points of the abnormal image area and the coordinates of the boundary points of the abnormal neural signal area. In some embodiments, step 4 may further include:

[0090] Step 4.1: Slice the three-dimensional image acquired by MRI (Magnetic Resonance Imaging) tomography according to the layer where the vertical channel of the array electrode is located, and obtain Num MRI two-dimensional slices.

[0091] Step 4.2: Use image segmentation models, such as Mask-Rcnn, Unet and other deep learning network models, to segment the two-dimensional slices and obtain the set of boundary coordinates of the markers in each slice layer (fx i ,fy i ).

[0092] Step 4.3: Extend the horizontal and vertical lines of each vertical section of the array electrode and draw them into a grid shape, and count (x i ,y i The grid where each point is located in ) (i.e., the boundary coordinates in each transverse section obtained in step 3.4).

[0093] Step 4.4: Calculate the coordinate points (fx i ,fy i ), and the Euclidean distance from each grid center point in step 4.3, statistics (fx i ,fy i ) is the grid where each point is located.

[0094] Step 4.5, construct the loss function:

[0095]

[0096] Where, Num represents the number of layers of the two-dimensional slice; Count represents the number of grids in the current slice layer; K1 represents the number of coordinates obtained by image segmentation in the current grid; K2 represents the number of coordinates obtained by neural signals in the current grid; x fj_mean and fj_mean They respectively represent the average values ​​of the horizontal coordinates and the average values ​​of the vertical coordinates of the boundary points of the abnormal image region corresponding to the boundary of the landmark obtained by performing MRI tomographic image segmentation by the image segmentation model in the j-th layer cross section; j_mean and j_mean They respectively represent the average abscissa and average ordinate of the corresponding corrected boundary points of the abnormal neural signal region obtained through electrode signal analysis in the j-th layer cross section.

[0097] Step 5: Iteratively optimize the loss function using a gradient descent method to obtain optimized neural signal abnormal region boundary determination parameters related to the smoothing, straight line fitting and interpolation processing. In some embodiments, step 5 may further include:

[0098] Step 5.1, define the objective function, loss = 0.

[0099] Step 5.2: Use the back-propagation control method to take the parameters v, w and α in step 1, the parameters k1, b1 and s in step 2, and the parameter β in step 3 as weight parameters.

[0100] Step 5.3: Use gradient descent as the optimization function.

[0101] Step 5.4: When loss < 0.001, the optimization is terminated, and the optimized parameters for determining the abnormal region boundary of the neural signal related to smoothing processing (parameters v, w and α), straight line fitting processing (parameters k1, b1 and s) and interpolation processing (parameter β) are obtained.

[0102] Step 6: extracting spike segments from the target object's neural signals acquired by the array electrodes using a threshold method, clustering the spike segments to obtain the coordinates of each neuron, adding the coordinates of the neuron to the corresponding coordinates of the boundary points of the abnormal image region to form a new training set image, and retraining and optimizing the image segmentation model based on the new training set image. In some embodiments, step 6 may further include:

[0103] Step 6.1: Through steps 1 to 5, the neural signal is used as the gold standard to identify the boundary of the abnormal area, which is used as the training set of the image segmentation model.

[0104] Step 6.2: Extract spike segments from the neural signals collected by the array electrodes using the threshold method, calculate the equivalent coordinates of each spike segment, cluster the spike segments into neurons, and then average the coordinates of all spike segments of the same neuron to obtain the coordinates of each neuron.

[0105] Step 6.3, vertically add the coordinate points of the neurons to the coordinate positions corresponding to the MRI slice layers in step 4.2. This processing can enhance the display of the marker area features in the magnetic resonance image.

[0106] Step 6.4: Based on the new segmentation data set from step 6.1 to step 6.3, the MRI marker image segmentation model is retrained. That is, in the marker area in the MRI slice image, color is artificially added to the location of the neuron coordinate points to form a new training set image, and the MRI marker image segmentation model is retrained with the new training set image.

[0107] Step 6.5: Specially display the pixel with the highest confidence in the segmented area and compare it with the neuron coordinates in step 6.3. In this way, the abnormal neuron area can be found through MRI image segmentation.

[0108] Step 7: Based on the segmentation boundary of the optimized image segmentation model, the loss function is re-iteratively optimized, and the image segmentation model is continuously re-optimized, and the pixel point area with the highest confidence in the re-optimized image segmentation model is compared with the neuron clustering result. When the comparison result error is less than the preset expected value threshold, the iteration is stopped, and the boundary of the abnormal area of ​​the target object is determined by the image segmentation of the image segmentation model. In some embodiments, step 7 may also specifically include:

[0109] Step 7.1, using the new image segmentation model to segment the boundary, as a guide, repeat step 5 to optimize the boundary determination parameters of those neural signal abnormal areas (i.e., the aforementioned v, w, α, k1, b1, s, and β).

[0110] Step 7.2: Based on the optimized neural signal abnormal area recognition results and neuron clustering results, continue to optimize the MRI image segmentation model.

[0111] Step 7.3, compare the pixel area with the highest confidence in the optimized MRI image segmentation model with the neuron clustering result. When the comparison result error is less than the expected value, the abnormal discharge neuron activity area can be intuitively seen from the image segmentation of the MRI image segmentation model. That is, the optimized neural signal marker boundary and neuron discharge area are displayed as markers on the MRI image, thereby optimizing the accuracy of the image segmentation model.

[0112] The present application proposes a method for determining the boundary of an abnormal region, which analyzes and processes the neural signals of the target object obtained by the array electrode to obtain the coordinates of the boundary points of the modified neural signal abnormal region; then obtains the coordinates of the boundary points of the image abnormal region by segmenting the tomographic image of the target object + deep learning image segmentation model; according to the coordinates of the boundary points of the image abnormal region, optimizes the parameters for determining the boundary of the neural signal abnormal region, and continues to optimize the image segmentation model; compares the pixel point area with the highest confidence of the re-optimized image segmentation model with the neuron clustering result, stops iteration when the comparison result error is less than the preset expected value threshold, and determines the boundary of the abnormal region of the target object by image segmentation of the image segmentation model. The method can realize accurate identification of the boundary of the abnormal region through mutual optimization and iteration of the two abnormal region boundary determination ideas of neural signal and tomographic image + image segmentation model. The spatial accuracy of neural signals is at the micron level, while the accuracy of MRI tomographic images is at the millimeter level. The present application optimizes the accuracy of the image algorithm segmentation model again with the spatial accuracy of neural signals, which is of great significance for the accurate identification of the boundary of the body lesion area.

[0113] It should be noted that the functional modules in the various embodiments of the present application can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product.

[0114] The flowchart and / or block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method or computer program product according to various embodiments of the present application. In this regard, each box in the flowchart and / or block diagram can represent a part of a module, program segment or code, and a part of the above-mentioned module, program segment or code includes one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from that marked in the accompanying drawings. It should also be noted that each box in the block diagram or flowchart, and the combination of the boxes in the block diagram or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0115] Those skilled in the art will appreciate that the features described in the various embodiments and / or claims of the present application may be combined and / or combined in a variety of ways, even if such combinations and / or combinations are not explicitly described in the present application. In particular, without departing from the spirit and teachings of the present application, the technical features described in the various embodiments and / or claims of the present application may be combined and / or combined in a variety of ways, and all of these combinations and / or combinations fall within the scope of the present application.

[0116] Although the present application has been shown and described with reference to specific exemplary embodiments of the present application, it should be understood by those skilled in the art that various changes in form and details may be made to the present application without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents. Therefore, the scope of the present application should not be limited to the above-mentioned embodiments, but should be determined not only by the appended claims, but also by the equivalents of the appended claims.

Claims

1. A method for determining the boundary of an abnormal area, characterized in that: include: Perform feature extraction and feature fitting on the neural signals of the target object acquired by the array electrodes, and perform smoothing on the acquired abnormal recognition results, eliminate channels without abnormal intervals, and obtain the average confidence of all the retained abnormal channels; Performing straight line fitting processing on the average confidence of the abnormal channel and the corresponding channel point coordinates, and obtaining the boundary points of the abnormal area based on a preset confidence threshold; Performing interpolation processing on the boundary points of the abnormal area to obtain corrected coordinates of the boundary points of the abnormal area of ​​the neural signal; Acquire a tomographic image of the target object, perform landmark segmentation through an image segmentation model to obtain corresponding coordinates of boundary points of abnormal image regions, and construct a loss function based on the coordinates of boundary points of abnormal image regions and the coordinates of boundary points of abnormal neural signal regions; Iteratively optimizing the loss function using a gradient descent method to obtain optimized neural signal abnormal region boundary determination parameters related to the smoothing process, straight line fitting process, and interpolation process; Performing threshold method spike segment extraction processing on the neural signal of the target object acquired by the array electrode, clustering the spike segments to obtain the coordinates of each neuron, adding the coordinates of the neuron to the corresponding coordinates of the boundary points of the abnormal area of ​​the image to form a new training set image, and retraining and optimizing the image segmentation model based on the new training set image; Based on the segmentation boundary of the optimized image segmentation model, the loss function is re-iteratively optimized, and the image segmentation model is continuously re-optimized, and the pixel point area with the highest confidence in the re-optimized image segmentation model is compared with the neuron clustering result. When the comparison result error is less than a preset expected value threshold, the iteration is stopped, and the boundary of the abnormal area of ​​the target object is determined through image segmentation of the image segmentation model.

2. The abnormal region boundary determination method according to claim 1, characterized in that: The feature extraction of the neural signal of the target object acquired by the array electrode includes: Cut each channel on each probe of the array electrode into segments based on a sliding window, and obtain a plurality of segments for each channel; Perform time domain feature extraction and frequency domain feature extraction on the frequency domain signal of each segment of each channel; Fourier transform is performed on the time domain signal of each segment of each channel to obtain a time-frequency image, and image features are extracted from the time-frequency image.

3. The abnormal region boundary determination method according to claim 2, characterized in that: Performing the feature fitting includes: The features extracted from the frequency domain signal and the time domain signal are sent as training data to a machine learning model for model training to obtain an optimal model for abnormality recognition and the selected features; Each channel on each probe of the array electrode is segmented, feature extraction is performed according to the selected features, the trained best abnormality recognition model is loaded, and the abnormality recognition result of the current segment and the confidence of the model output are obtained.

4. The abnormal region boundary determination method according to claim 3, characterized in that: The obtained abnormal recognition results are smoothed, channels without abnormal intervals are eliminated, and the average confidence of all the retained abnormal channels is obtained, including: In the abnormal identification result, if the identification result of at least the second consecutive data segments among the first consecutive data segments is abnormal, then mark the first consecutive data segments as abnormal intervals; The smoothing mechanism continues until at least the second consecutive data segments in the first consecutive data segments have normal recognition results, then this interval result is overturned as a normal interval, otherwise it continues to be an abnormal interval, and the start and end time of each channel abnormal interval is recorded; the first data is greater than the second data; For each channel on each probe, the mean interval of the start and end time of each abnormal interval is calculated respectively. If the overlap rate between the start and end time of the abnormal interval and the mean interval of the start and end time is less than a preset overlap rate threshold, the abnormal interval is reset to a normal interval. Eliminate channels without abnormal intervals, obtain all the retained abnormal channels and obtain the corresponding average confidence.

5. The abnormal region boundary determination method according to claim 4, characterized in that: The performing straight line fitting processing on the average confidence of the abnormal channel and the corresponding channel point coordinates, and obtaining the boundary points of the abnormal area based on a preset confidence threshold, comprises: Performing a straight line fitting on the average confidence of the abnormal channel and the corresponding channel point coordinates by the least square method to obtain a corresponding straight line fitting equation; The confidence threshold is substituted into the straight line fitting equation as a dependent variable to obtain boundary points on all transverse channels and longitudinal channels of the abnormal area.

6. The abnormal region boundary determination method according to claim 5, characterized in that: The interpolation processing of the boundary points of the abnormal region to obtain the corrected coordinates of the boundary points of the abnormal region of the neural signal includes: one or more correction methods of singular value correction, whole line missing correction and boundary point density increase correction; Singular value correction: For every three adjacent boundary points A, B and C on the horizontal or vertical channels, a group of judgments is made. If the corresponding horizontal or vertical coordinates x A 、x B and x C A value in When the deviation exceeds the preset boundary point correction threshold β, the coordinate is marked as a singular coordinate and needs to be corrected; Draw a straight line through points A and C, and intersect the extension of the straight line of point B at point O. Take a point B' on the line connecting BO. When , point B' is the updated coordinate position of point B; Whole line missing fix: After the singular value correction is completed, take the midpoint F of the line segment AG, and draw a straight line through point F parallel to the straight line where point A is located; the point G is a boundary point adjacent to point A in a direction different from that of point C; Boundary point density increase correction: After the entire row missing is corrected, the midpoint of line segment B'C is taken as a newly added boundary point.

7. The abnormal region boundary determination method according to claim 6, characterized in that: The step of acquiring a tomographic image of the target object, performing landmark segmentation through an image segmentation model to obtain corresponding coordinates of boundary points of an abnormal image region, and constructing a loss function based on the coordinates of boundary points of the abnormal image region and the coordinates of boundary points of an abnormal neural signal region comprises: Slicing the three-dimensional magnetic resonance image according to the layer where the vertical channel of the array electrode is located to obtain a plurality of two-dimensional slices; Using the image segmentation model, the two-dimensional slices are respectively segmented for the markers, and the coordinates of the boundary points of the image abnormal area corresponding to the boundary of the markers in each slice layer are respectively obtained; Gridding the cross-sections of the array electrodes in the vertical, horizontal and longitudinal directions, respectively obtaining the Euclidean distances between the boundary points of each abnormal image region and the center points of each grid, and performing statistics to obtain the grid where the boundary points of each abnormal image region are located; The loss function is constructed, and its specific formula is as follows: Where Num represents the number of layers of the two-dimensional slice; Count represents the number of grids in the current slice layer; K1 represents the number of coordinates obtained by image segmentation in the current grid; K2 represents the number of coordinates obtained by neural signals in the current grid; x fj_mean and fj_mean They respectively represent the average values ​​of the horizontal coordinates and the vertical coordinates of the boundary points of the abnormal image region corresponding to the boundary of the landmark obtained by image segmentation by the image segmentation model in the j-th layer cross section; j_mean and j_mean They respectively represent the average abscissa and average ordinate of the corresponding corrected boundary points of the abnormal neural signal region obtained through electrode signal analysis in the j-th layer cross section.

8. The abnormal region boundary determination method according to claim 7, characterized in that: The gradient descent method is used to iteratively optimize the loss function to obtain the optimized neural signal abnormal region boundary determination parameters related to the smoothing process, the straight line fitting process and the interpolation process, including: Define the objective function, loss = 0; Adopting a back-propagation control method, taking one or more of the first data, the second data, the coincidence rate threshold, the intercept and slope of the straight line fitting equation, the confidence threshold and the boundary point correction threshold as weight parameters; The gradient descent method is used as the optimization function. When loss is less than 0.001, the iterative optimization can be terminated to obtain the corresponding boundary determination parameters of all the abnormal neural signal regions at this time.

9. The abnormal region boundary determination method according to claim 8, characterized in that: The step of performing a threshold method spike segment extraction process on the neural signal of the target object acquired by the array electrode, clustering the spike segments to obtain the coordinates of each neuron, adding the coordinates of the neuron to the corresponding coordinates of the boundary points of the abnormal image region to form a new training set image, and retraining and optimizing the image segmentation model based on the new training set image includes: Extract spike fragments from the target object's neural signals acquired by the array electrodes using the threshold method, obtain the equivalent coordinates of each spike fragment, perform neuron clustering on the spike fragments, and then calculate the average of the coordinates of all spike fragments of the same neuron to obtain the coordinates of each neuron; Add the coordinate point of each neuron vertically to the coordinate position corresponding to the slice layer of the tomographic image to obtain a new training set image; The image segmentation model is retrained and optimized based on the new training set images.

10. The abnormal region boundary determination method according to claim 9, characterized in that: The continuing to optimize the image segmentation model comprises: Based on the iteratively optimized loss function, optimizing the abnormal neural signal region recognition results and neuron clustering results; The image segmentation model is further optimized based on the optimized neural signal abnormal area recognition results and neuron clustering results.

Citation Information

Cited By

  • 3D medical image anomaly detection and classification method

    CN120635595A

  • Ground stair recognition and stimulation method and system for bionic visual system

    CN121121621A

  • Ground stair identification and stimulation method and system for a biomimetic vision system

    CN121121621B