After-stroke aphasia patient fMRI focal identification method based on deep learning

Abstract

The invention relates to the technical field of brain function network analysis, in particular to a post-stroke aphasia patient fMRI focal recognition method based on deep learning. The method comprises the following steps of: acquiring a region consistency value for geometric structures of matched local regions of two adjacent fMRI images in an fMRI image sequence, similarity of gray distribution and an image matching degree, and dividing a same time window by using the region consistency value; and obtaining a functional connection coefficient according to the difference of the region consistent values of the fMRI images in two adjacent time windows and the difference of the geometric structure features and the gray level distribution features of the local regions, determining the convolution kernel size of deep learning of the fMRI images in the time windows by using the functional connection coefficient, and performing focal recognition on the fMRI images in the time windows. According to the method, the function connection coefficient presenting the cooperation degree of the brain region activity modes in the adjacent time windows is analyzed, the convolution kernel size of the time windows is determined in a self-adaptive mode, and the recognition effect on the local focus of the patient is improved.

Description

Technical Field

[0001] This invention relates to the field of brain functional network analysis technology, specifically to a deep learning-based method for fMRI focal identification in post-stroke aphasia patients. Background Technology

[0002] Post-stroke aphasia is a common sequela of stroke, severely impacting patients' communication abilities and quality of life. Functional magnetic resonance imaging (fMRI) is widely used in brain function research. fMRI technology can capture the dynamic activity of different brain regions in a resting state and during specific language tasks, thereby revealing brain network regions related to language function. Deep learning, as a powerful data-driven method, can automatically extract features by constructing deep neural networks, thus effectively identifying and classifying complex patterns of brain functional activity; it can not only improve the accuracy of identifying damaged brain regions but also reveal the neural mechanisms of different language dysfunctions.

[0003] Convolutional Neural Networks (CNNs) are deep learning models specifically designed for image processing and computer vision tasks. The core characteristic of CNN models is their use of convolutional and pooling layers to extract features from images, which are then passed through fully connected layers for classification or regression tasks. Traditional CNN models extract spatial features using two-dimensional or three-dimensional convolutional kernels. However, convolutional operations only operate on single time points or short time segments, lacking the ability to capture long-term temporal dependencies. Yet, fMRI dynamic functional connectivity analysis requires analyzing the changing patterns of the functional connectivity matrix across time windows. When CNN models are used for fMRI dynamic functional connectivity analysis, their short-term nature may lead to misjudgments of the dynamic reorganization of the brain network, resulting in poor identification of local lesions in fMRI of stroke-induced aphasia patients. Summary of the Invention

[0004] To address the technical problem of poor identification of focal lesions in fMRI of post-stroke aphasia patients due to errors in fMRI dynamic functional connectivity analysis caused by the short-term nature of CNN models, this invention aims to provide a deep learning-based method for focal lesion identification in fMRI of post-stroke aphasia patients. The specific technical solution adopted is as follows: This invention proposes a deep learning-based method for fMRI focal identification in post-stroke aphasia patients, the method comprising: Obtain fMRI image sequences of the brain of a stroke-induced aphasia patient, wherein the fMRI images in the sequences contain different local regions; Local regions of every two adjacent fMRI images in the sequence are matched. Based on the geometric structural similarity and gray-level distribution similarity of the two matched local regions of every two adjacent fMRI images, as well as the degree of image matching, the region consistency value of the corresponding two fMRI images is obtained. Based on the region consistency value, the corresponding time period of the sequence is divided into different time windows. Based on the difference in the regional consistency value of fMRI images within each two adjacent time windows, as well as the differences in the geometric structure features and gray-scale distribution features of the local region, the functional connectivity coefficients for the corresponding two time windows are obtained. The kernel size for deep learning of fMRI images within each time window is determined based on the functional connectivity coefficients, and focal identification is performed on the fMRI images within each time window.

[0005] Further, obtaining the regional consistency value corresponding to two fMRI images includes: A matching region pair is formed by two matching local regions in each pair of adjacent fMRI images; the overall regional difference value of the corresponding two fMRI images is obtained based on the difference in gray value of the pixels of the two local regions in all matching region pairs of each pair of adjacent fMRI images and the difference in slope between adjacent pixels on the edge. The regional consistency value of two fMRI images is obtained by comparing the number of region matching pairs between each pair of adjacent fMRI images with the overall difference value of the region.

[0006] Further, obtaining the overall regional difference value corresponding to the two fMRI images includes: The slope sequence is obtained by arranging the slopes between all adjacent pixels on the edge of the local region in order; the average gray value of the pixels in the local region is calculated and recorded as the overall gray value. After calculating the product of the absolute value of the difference in the overall gray values ​​of the two local regions in each matching region pair of each adjacent two fMRI images and the DTW value of the slope sequence, the mean of all products is taken as the overall regional difference value of the corresponding two fMRI images. The formula for calculating the overall regional difference between any two adjacent fMRI images is as follows: ; This represents the overall regional difference value between the c-th and c+1-th fMRI images in the fMRI image sequence; The number of matching region pairs in the c-th and c+1-th fMRI images in the fMRI image sequence; is the overall gray value of the local region located in the c-th fMRI image in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; is the overall gray value of the local region located in the (c+1)th fMRI image in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; DTW values ​​of the slope sequences of two local regions in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; It is an absolute value function.

[0007] Further, the step of dividing the corresponding time period of the sequence into different time windows based on the regional consistency value includes: Based on the region consistency value, the fMRI images in the fMRI image sequence are clustered to obtain different clusters; Each fMRI image in the fMRI image sequence corresponds to a time point; determine whether the fMRI images within each cluster are continuous in the fMRI image sequence. If so, take the time points corresponding to the fMRI images with the smallest and largest indices within each cluster as the start and end times of a time window, respectively. If not, based on the continuity of fMRI images within each cluster in the fMRI image sequence, the fMRI images within each cluster are divided into different continuous image segments, and the times corresponding to the fMRI images with the smallest and largest subscripts in each continuous image segment are respectively used as the start and end times of a time window.

[0008] Furthermore, obtaining the functional connectivity coefficients corresponding to the two time windows includes: Based on the geometric structure and grayscale distribution characteristics of the local region of the fMRI image within each time window, the image feature values ​​for the corresponding time window are obtained. Calculate the mean of the region consistency value of all fMRI images within each time window, and negatively correlate and normalize the product of the absolute value of the difference between the mean of each two adjacent time windows and the absolute value of the difference between the image feature values ​​to obtain the functional connectivity coefficients of the corresponding two time windows. The formula for calculating the functional connectivity coefficient between any two adjacent time windows is as follows: In the formula, For the functional connectivity coefficients of the d-th and d+1-th time windows; The mean of the regional consistency values ​​of all two adjacent fMRI images within the d-th time window; The mean of the regional consistency values ​​of all two adjacent fMRI images within the (d+1)th time window; Let be the image feature value for the d-th time window; is the image feature value of the (d+1)th time window; exp is an exponential function with the natural constant as the base; It is an absolute value function.

[0009] Further, obtaining the image feature values ​​corresponding to the time window includes: For each fMRI image, the absolute value of the difference between the overall gray value of each local region of the fMRI image and the mean gray value of the pixels in the fMRI image is calculated, and the ratio of the gray value variance of the pixels in each local region is calculated. The sum of all ratios is then used as the gray feature value. The absolute value of the difference between the mean of the elements in the slope sequence of each local region of the fMRI image and the mean of the elements in the slope sequence of all local regions is calculated, and the ratio of the variance of the elements in the slope sequence of each local region is used as the sum of all ratios as the structural feature value. The mean of the products of the grayscale feature values ​​and the structural feature values ​​of all fMRI images within each time window is taken as the image feature value for each time window; the formula for calculating the image feature value is as follows: , , In the formula, K is the image feature value of each time window; V is the total number of fMRI images within each time window; The grayscale feature value of the v-th fMRI image within each time window; The structural feature values ​​of the v-th fMRI image within each time window; The total number of local regions in the v-th fMRI image within each time window; The overall grayscale value of the u-th local region in the v-th fMRI image within each time window; The mean gray value of a pixel in the v-th fMRI image within each time window; The variance of grayscale values ​​of pixels in the u-th local region of the v-th fMRI image within each time window; The mean value of the elements in the slope sequence of the u-th local region of the v-th fMRI image within each time window; The mean value of the elements in the slope sequence of the local region of the v-th fMRI image within each time window; The variance of the elements in the slope sequence of the u-th local region of the v-th fMRI image within each time window.

[0010] Furthermore, the method for obtaining the convolution kernel size includes: The preset convolution kernel size is weighted by the sum of the constant 1 and the functional connection coefficient, and the weighted result is used as the convolution kernel size for deep learning of all fMRI images in the latter time window of each two adjacent time windows.

[0011] Furthermore, the method for performing deep learning on the fMRI images is a CNN model.

[0012] Furthermore, the method for matching local regions of each pair of adjacent fMRI images is a template matching algorithm.

[0013] Furthermore, the preset convolution kernel size is 9.

[0014] The present invention has the following beneficial effects: In this embodiment of the invention, because a brain region in a state of continuous activation will have a high similarity in grayscale distribution and geometric structure of consecutive fMRI images, the degree of matching of local regions in two adjacent fMRI images is analyzed based on the similarity of geometric structure and grayscale distribution of matching local regions, combined with the degree of image matching from a global image perspective. The resulting region consistency value can measure the probability that adjacent fMRI images are in a state of continuous activation. Using it to divide time windows, the dynamic evolution characteristics of brain functional connectivity patterns can be captured more accurately, making the analysis of neural activity more reliable. The difference in region consistency value of fMRI images within two adjacent time windows, as well as the difference in the geometric structure and grayscale distribution characteristics of local regions, can reflect the brain region's state of continuous activation in adjacent time windows. The possibility of a time window being in a continuously activated state is analyzed, and the degree of coordinated activity of brain regions in adjacent time windows is obtained. The degree of coordinated activity directly reflects the spatiotemporal stability of brain regions. In order to maintain the spatiotemporal synchronization of dynamic changes in brain functional connectivity, the size of the convolution kernel can be adaptively adjusted based on the functional connectivity coefficient. This enables the deep learning model to effectively capture the dynamic interaction relationship of fMRI images in spatiotemporal space. This effectively solves the problem that when CNN models are used for fMRI dynamic functional connectivity analysis, the short-term nature may lead to misjudgment of the dynamic reorganization of the brain network. This enhances the explanatory power of deep learning models for complex neurological diseases such as post-stroke aphasia, improves the analytical ability of deep learning models on fMRI images, and thus improves the recognition effect of local lesions in fMRI images of post-stroke aphasia patients. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the steps of a deep learning-based fMRI focal identification method for post-stroke aphasia patients, as provided in one embodiment of the present invention. Figure 2A flowchart illustrating a method for obtaining a region consistency value according to an embodiment of the present invention; Figure 3 A flowchart illustrating a method for obtaining functional connectivity coefficients according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a computer device for identifying focal lesions in post-stroke aphasia patients based on deep learning, as provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the deep learning-based fMRI focal identification method for post-stroke aphasia patients proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of the fMRI focal identification method for post-stroke aphasia patients based on deep learning provided by this invention.

[0020] Example 1: This invention proposes a deep learning-based method for fMRI focal identification in post-stroke aphasia patients. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates a flowchart of a deep learning-based fMRI focal identification method for post-stroke aphasia patients, according to an embodiment of the present invention. The method includes: Step S1: Obtain fMRI image sequences of the brain of stroke patients with aphasia, where the fMRI images contain different local regions.

[0021] The brains of stroke-induced aphasia patients were scanned using magnetic resonance imaging (fMRI). During the scan, patients performed specific tasks such as word naming, sentence repetition, and finger tapping, alternating between 30-second tasks and rest periods 10 times, with 5-minute rest periods before and after each task. Patients remained relaxed during rest periods, allowing for analysis of the dynamic connectivity strength between the default mode network and the language network. The fMRI machine acquired images every 2 seconds, and all images were arranged chronologically to form an fMRI image sequence, reflecting changes in brain activity during specific tasks or in different states.

[0022] It should be noted that, in order to reduce head movement artifacts during the scanning process caused by patient movement, a special head coil and soft pad were used to fix the head of the stroke-induced aphasia patient. An MRI machine with a magnetic field strength greater than 3.0 Tesla, such as the Philips Ingenia 3.0T CX, was selected, and the temporal resolution was set to one layer every 2 seconds to capture the fine activation patterns of language-related brain regions.

[0023] Step S2: Match the local regions of every two adjacent fMRI images in the sequence. Based on the geometric structural similarity and gray-scale distribution similarity of the two matched local regions of every two adjacent fMRI images, as well as the degree of image matching, obtain the region consistency value of the corresponding two fMRI images. Divide the corresponding time period of the sequence into different time windows based on the region consistency value.

[0024] Brain functional connectivity is not static; its dynamic evolutionary characteristics fluctuate with cognitive tasks, emotional states, or spontaneous activities. To capture these dynamic characteristics, it is necessary to analyze dynamic functional connectivity through time window segmentation. When a brain region is in a state of continuous activation, its fMRI signal exhibits high temporal correlation across consecutive time frames. This is manifested in regular changes in local BOLD signal intensity, such as stable increases or decreases, and relatively stable spatial extent of the activated area. Specifically, BOLD signal intensity determines the grayscale distribution of a local region, while spatial extent determines its geometric structure. Therefore, if a brain region is in a state of continuous activation, the similarity in grayscale distribution and geometric structure of local regions in consecutive fMRI images will be higher.

[0025] To capture subtle changes between localized areas of the brain more precisely, local regions of two adjacent fMRI images are matched to avoid the influence of large-scale image differences. The degree of image matching between two adjacent fMRI images is analyzed from a global image perspective to assess their similarity. This is combined with the geometric and gray-scale distribution similarities of the matched local regions between the two fMRI images to analyze the degree of matching. Higher image matching, geometric similarity, and gray-scale distribution similarity indicate a higher degree of matching between the local regions of the two fMRI images, thus yielding a region consistency value.

[0026] Region consistency values ​​measure the likelihood of sustained activation in adjacent fMRI images by the degree of matching of local regions between two adjacent fMRI images. Dividing the corresponding time periods of an fMRI image sequence into different time windows based on region consistency values ​​allows for more precise capture of the dynamic evolution of brain functional connectivity patterns, making the analysis of neural activity more reliable. The corresponding time period of the fMRI image sequence refers to the entire time span during which the patient is in both a task-oriented and resting state.

[0027] In one implementation of this invention, a template matching algorithm is used to match local regions of two adjacent fMRI images. Other embodiments may use scale-invariant feature transformation algorithms, histogram matching algorithms, and Siamese neural networks.

[0028] Step S3: Based on the difference in regional consistency values ​​of fMRI images within each two adjacent time windows, as well as the differences in geometric structural features and gray-scale distribution features of local regions, obtain the functional connectivity coefficients for the corresponding two time windows.

[0029] fMRI images reveal dynamic changes in brain region activity, with activity in adjacent time windows typically exhibiting continuity and correlation. By analyzing the degree of coordinated activity among brain regions in adjacent time windows, the spatiotemporal coupling strength of brain activity can be quantified, thereby tracing its evolution. Since the degree of coordinated activity directly reflects the spatiotemporal stability of brain regions, to maintain the spatiotemporal synchronization of dynamic changes in brain functional connectivity, the convolutional kernel size needs to be adaptively adjusted according to the degree of coordinated activity among brain regions in adjacent time windows to optimize the sensitivity of feature extraction in deep learning models.

[0030] Regional consistency values ​​measure the likelihood of adjacent fMRI images being in a state of sustained activation. Differences in regional consistency values ​​between adjacent time windows reflect the stability of neural activity in brain regions within those windows; smaller differences indicate greater stability and a higher degree of coordination in brain region activity patterns within adjacent time windows. Differences in geometric structural features and grayscale distribution characteristics between adjacent time windows can measure the likelihood of brain regions being in a state of sustained activation; smaller differences indicate a higher likelihood of sustained activation and a higher degree of coordination in brain region activity patterns within adjacent time windows. Therefore, by combining the differences in regional consistency values ​​of fMRI images within adjacent time windows with differences in the spatial and grayscale distribution characteristics of local regions, the degree of coordination in brain region activity patterns within adjacent time windows can be analyzed to obtain functional connectivity coefficients.

[0031] Step S4: Determine the convolution kernel size for deep learning of fMRI images within each time window based on the functional connectivity coefficients, and perform focal identification on fMRI images within each time window.

[0032] Functional connectivity coefficients measure the degree of coordination of brain region activity patterns within adjacent time windows. Higher coordination indicates a greater likelihood of stable collaborative networks forming within the brain regions. This necessitates the use of small convolutional kernels to identify local lesions in aphasia patients within fMRI images, more sensitively capturing lesion features specific to aphasia patients, avoiding widespread blurring, and thus more completely segmenting local lesions. Conversely, lower coordination allows for the analysis of abnormal brain functional connectivity using large convolutional kernels, capturing cross-regional coordination anomalies. Using functional connectivity coefficients to guide the selection of convolutional kernel size for time windows achieves a balance between precise localization and wide coverage. Deep learning models can effectively capture the dynamic spatiotemporal interactions of fMRI images, adapting to the diagnostic needs of different brain network states, enhancing the interpretability of deep learning models for complex neurological diseases such as post-stroke aphasia, and ultimately improving the analytical capabilities of deep learning models for fMRI images.

[0033] In one implementation of this invention, the deep learning method is a CNN model.

[0034] The process of identifying local lesions in fMRI images using a CNN model based on the convolutional kernel size of fMRI images within a time window is as follows: First, the fMRI images in the fMRI image sequence are preprocessed using spatial standardization, such as MNI spatial registration and temporal layer correction, to eliminate individual differences and noise interference. Second, based on clinical diagnostic results such as the anatomical location of tumors or epileptic foci, doctors perform mask annotation on the 3D voxel data to generate binary label maps, where the label value for lesion areas is 1 and the label value for normal brain areas is 0, thus achieving the annotation of local lesion areas. Then, according to the convolutional kernel size corresponding to each time window, a multi-level 3D convolutional module is constructed, prioritizing the use of small-sized convolutional kernels to reduce the number of parameters and capture local subtle features such as gray matter abnormal signals. Finally, Dice Loss + Focal Loss are used for joint optimization to alleviate the class imbalance problem caused by extremely low proportions of local lesion areas, such as tiny epileptic foci. The fully connected layers are fine-tuned on the basis of pre-trained 3D medical image models such as 3D ResNet-50 on BraTS dataset to accelerate convergence and improve generalization ability in small data scenarios. It should be noted that deep learning can accurately identify minute changes, detailed features, morphology and location of local lesions, which helps doctors to more accurately determine the nature of local lesions.

[0035] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the regional consistency value is described in [reference needed]. Figure 2 The diagram illustrates a flowchart of a method for obtaining a region consistency value according to an embodiment of the present invention. The method includes: Step S210: A matching region pair is formed by two matching local regions of each two adjacent fMRI images; based on the difference in gray values ​​of the pixels of the two local regions in all matching region pairs of each two adjacent fMRI images and the difference in slope between adjacent pixels on the edge, the overall difference value of the region of the corresponding two fMRI images is obtained.

[0036] In this embodiment of the invention, the method for obtaining the overall regional difference value includes: arranging the slopes between all adjacent pixels on the edge of a local region in order to obtain a slope sequence; calculating the average gray value of the pixels within the local region as the overall gray value; calculating the product of the absolute value of the difference between the overall gray values ​​of the two local regions in each matching region pair of each pair of adjacent fMRI images and the DTW value of the slope sequence, and then taking the average of all products as the overall regional difference value of the corresponding two fMRI images. The formula for calculating the overall regional difference value of each pair of adjacent fMRI images is as follows: ; This represents the overall regional difference value between the c-th and c+1-th fMRI images in the fMRI image sequence; The number of matching region pairs in the c-th and c+1-th fMRI images in the fMRI image sequence; is the overall gray value of the local region located in the c-th fMRI image in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; is the overall gray value of the local region located in the (c+1)th fMRI image in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; DTW values ​​of the slope sequences of two local regions in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; It is an absolute value function.

[0037] It should be noted that the DTW value measures the spatial distribution similarity by analyzing the degree of matching of features such as the curvature direction and inflection point distribution of the edges of two local regions in a region matching pair. The smaller the DTW value, the more similar the spatial distribution of the two local regions. The overall gray value reflects the overall gray level of the local region. The smaller the overall gray value of the two local regions in a region matching pair, the more similar the gray distribution of the two local regions. If the overall difference value of the region is smaller, the spatial distribution similarity and gray level similarity of the local regions of two adjacent fMRI images are greater.

[0038] It is important to note that, to avoid deviations in DTW values ​​due to the random starting point of pixels on the edges of local regions, when obtaining the slope sequence, the pixels on the edges of the local region should first be arranged clockwise from a preset point to obtain the pixel sequence. The slope order between each pair of adjacent pixels in the pixel sequence should then be arranged. A vertical line passing through the center point of the local region should be drawn, and the intersection point of this line and the edge of the local region with the maximum pixel coordinate is recorded as the preset point.

[0039] Step S220: Obtain the regional consistency value of two fMRI images based on the number of regional matching pairs between each pair of adjacent fMRI images and the overall regional difference value.

[0040] The number of region matching pairs between two adjacent fMRI images reflects the degree of image matching between them; a higher number of region matching pairs indicates a higher degree of matching. It is known that when brain regions are in a state of sustained activation, the similarity in grayscale distribution and geometric structure of local regions in consecutive fMRI images is higher. When the overall regional difference value is smaller and the number of region matching pairs is larger, the similarity in geometric structure and grayscale distribution between two adjacent fMRI images is greater, and the degree of image matching is higher; therefore, the two adjacent fMRI images are more similar. Thus, the overall regional difference value and the regional consistency value are negatively correlated, while the number of region matching pairs and the regional consistency value are positively correlated.

[0041] In one specific implementation of this invention, taking the first and second fMRI images in an fMRI image sequence as an example for analysis, the regional consistency value of the first and second fMRI images is analyzed. Expressed as a formula: ; In the formula, N is the minimum number of local regions in the first and second fMRI images of the fMRI image sequence; n is the number of matching region pairs in the first and second fMRI images of the fMRI image sequence. represents the overall regional difference between the first and second fMRI images in the fMRI sequence; exp is an exponential function with the natural constant as the base. It should be noted that if... The larger the value, the more similar the first and second fMRI images are, and the greater the likelihood that the brain region was in a state of continuous activation at the corresponding time points of the two fMRI images.

[0042] Preferably, in some possible implementations of the embodiments of the present invention, the method for dividing the time window includes: clustering fMRI images in the fMRI image sequence based on region consistency values ​​to obtain different clusters; each fMRI image in the fMRI image sequence corresponds to a time point; determining whether the fMRI images within each cluster are continuous in the fMRI image sequence; if so, then taking the time points corresponding to the fMRI images with the smallest and largest indices within each cluster as the start and end times of a time window, respectively; if not, then dividing the fMRI images within each cluster into different continuous image segments based on the continuity of the fMRI images within the fMRI image sequence, and taking the time points corresponding to the fMRI images with the smallest and largest indices in each continuous image segment as the start and end times of a time window, respectively. It should be noted that brain regions exhibit similar neural activity patterns within the same time window.

[0043] As an example, suppose fMRI image sequence A= In sequence A, elements represent fMRI images, and fMRI images within cluster J include... Within cluster J , If the fMRI images are consecutive in sequence A, then the fMRI images within cluster J are divided into... , Two consecutive image segments.

[0044] In one implementation of this invention, the DBSCAN algorithm is used to cluster fMRI images in an fMRI image sequence.

[0045] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the functional connectivity coefficients is described in [reference needed]. Figure 3 The diagram illustrates a flowchart of a method for obtaining functional connectivity coefficients according to an embodiment of the present invention. The method includes: Step S310: Based on the geometric structure features and grayscale distribution features of the local region of the fMRI image within each time window, obtain the image feature values ​​for the corresponding time window.

[0046] In this embodiment of the invention, the method for obtaining image feature values ​​includes: for each fMRI image, calculating the ratio of the absolute difference between the overall gray value of each local region of the fMRI image and the mean gray value of the pixels in the fMRI image, to the variance of the gray value of the pixels in each local region, and then summing all ratios as the gray-level feature value; calculating the ratio of the absolute difference between the mean of the elements in the slope sequence of each local region of the fMRI image and the mean of the elements in the slope sequence of all local regions, to the variance of the elements in the slope sequence of each local region, and then summing all ratios as the structural feature value; and taking the mean of the product of the gray-level feature value and the structural feature value of all fMRI images within each time window as the image feature value of each time window. The formula for calculating the image feature value is as follows: , , In the formula, K is the image feature value of each time window; V is the total number of fMRI images within each time window; The grayscale feature value of the v-th fMRI image within each time window; The structural feature values ​​of the v-th fMRI image within each time window; The total number of local regions in the v-th fMRI image within each time window; The overall grayscale value of the u-th local region in the v-th fMRI image within each time window; The mean gray value of a pixel in the v-th fMRI image within each time window; The variance of grayscale values ​​of pixels in the u-th local region of the v-th fMRI image within each time window; The mean value of the elements in the slope sequence of the u-th local region of the v-th fMRI image within each time window; The mean value of the elements in the slope sequence of the local region of the v-th fMRI image within each time window; The variance of the elements in the slope sequence of the u-th local region of the v-th fMRI image within each time window.

[0047] It should be noted that grayscale and structural feature values ​​simultaneously quantify the local and global contrast intensity and regional consistency of fMRI images, significantly improving the sensitivity of grayscale distribution features and geometric structural features to abnormal regions and unique structures. This makes them suitable for medical image analysis requiring high-precision localization, such as lesion segmentation or dynamic functional connectivity studies. High grayscale feature values ​​indicate high global grayscale contrast and regional uniformity in fMRI images, while high structural feature values ​​indicate significant differences between local edge patterns and global structures in fMRI images, and these differences are inherently stable. By analyzing the overall level of grayscale distribution and geometric structural features of all fMRI images within a time window, image feature values ​​are obtained. The larger the value, the higher the global grayscale contrast and regional uniformity, and the more significant and stable the global structural differences within the time window.

[0048] In other embodiments of the present invention, the variance of gray values ​​of pixels in each local region of the fMRI image can be calculated, and the mean of all variances can be used as the gray feature value; the variance of elements in the slope sequence of each local region of the fMRI image can be calculated, and the mean of all variances can be used as the structural feature value.

[0049] Step S320: Calculate the mean of the regional consistency values ​​of all fMRI images within each time window, negatively correlate the product of the absolute difference between the means of each two adjacent time windows and the absolute difference between the image feature values, and normalize it to obtain the functional connectivity coefficients for the corresponding two time windows.

[0050] The formula for calculating the functional connectivity coefficient between any two adjacent time windows is as follows: In the formula, For the functional connectivity coefficients of the d-th and d+1-th time windows; The mean of the regional consistency values ​​of all two adjacent fMRI images within the d-th time window; The mean of the regional consistency values ​​of all two adjacent fMRI images within the (d+1)th time window; Let be the image feature value for the d-th time window; is the image feature value of the (d+1)th time window; exp is an exponential function with the natural constant as the base; It is an absolute value function.

[0051] The mean regional consistency value of all fMRI images within a time window reflects the overall stability of neural activity within that time window. The smaller the difference between the means of these values ​​in two adjacent time windows, the more stable the neural activity in those two time windows, and the higher the degree of coordination in brain region activity patterns within those two time windows. Image feature values ​​reflect the overall level of grayscale distribution and geometric structure features of fMRI images within a time window. The smaller the difference in image feature values ​​between two adjacent time windows, the greater the likelihood that a brain region is in a state of sustained activation between those two time windows, as measured by the differences in geometric structure and grayscale distribution features, and the higher the degree of coordination in brain region activity patterns within those two time windows. Therefore, the functional connectivity coefficient is obtained by negatively correlated with the product of the absolute difference between the mean regional consistency values ​​of each pair of adjacent time windows and the absolute difference between the image feature values. The larger the functional connectivity coefficient, the higher the degree of coordination of brain region activity patterns in two adjacent time windows, and the more stable the functional connectivity of the brain region in the two time windows, and the better the continuity of the functional network state. The smaller the functional connectivity coefficient, the worse the functional connectivity of the brain region in two adjacent time windows, which may correspond to sudden changes in the brain network state, such as task switching, pathological activities, etc.

[0052] It should be noted that while Norm is used for normalization in this embodiment of the invention, other normalization methods can also be selected, such as function transformation, max-min normalization, etc., and are not limited here.

[0053] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the convolution kernel size includes: weighting the preset convolution kernel size using the sum of a constant 1 and functional connectivity coefficients, and using the weighted result as the convolution kernel size for deep learning of all fMRI images in the latter time window of every two adjacent time windows. It should be noted that if the functional connectivity coefficients of two adjacent time windows are larger, it indicates a higher degree of coordination in brain region activity patterns within the two time windows, and the brain regions are more likely to form a stable collaborative network. If a small-sized convolution kernel is needed to identify local details of local lesions in fMRI images of aphasia patients, then the preset convolution kernel size needs to be adjusted using negatively correlated and normalized functional connectivity coefficients to determine the convolution kernel size for each time window.

[0054] To avoid the sliding offset problem caused by the inability to determine the center anchor point due to an even-numbered kernel size, it is necessary to ensure that the kernel size is odd. The specific method is as follows: when the integer part of the weighted result is odd, the weighted result is rounded down and used as the kernel size for deep learning of fMRI images in the latter time window of each two adjacent time windows; when the integer part of the weighted result is even, the weighted result is rounded up and used as the kernel size for deep learning of fMRI images in the latter time window of each two adjacent time windows.

[0055] In one implementation of this invention, the preset convolution kernel size is set to 9.

[0056] It is important to note that the kernel size refers to the combined length and width of the kernel. In this embodiment, the length and width of the kernel are equal, and the kernel size ranges from 1 to 9 (an odd number). The kernel sizes used for deep learning on fMRI images within the first two time windows are equal.

[0057] This invention is now complete.

[0058] Example 2: This invention also presents a schematic diagram of a computer device for fMRI focal identification of post-stroke aphasia patients based on deep learning. Please refer to [link / reference]. Figure 4 The computer device includes a memory 501, a processor 502, and a computer program 503 stored in the memory 501 and running on the processor 502. When the processor 502 executes the computer program 503, the computer device can perform any of the aforementioned deep learning-based fMRI focal identification methods for post-stroke aphasia patients.

[0059] Furthermore, embodiments of this application also protect an apparatus that may include a memory and a processor, wherein the memory stores executable program code, and the processor is used to call and execute the executable program code to perform a deep learning-based fMRI focal identification method for post-stroke aphasia patients provided in embodiments of this application.

[0060] This embodiment can divide the device into functional modules based on the above method example. For example, each module can correspond to a separate function, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0061] When each module is divided according to its function, the device may also include a communication module, a signal analysis module, a complexity analysis module, and a positioning module. It should be noted that all relevant content of each step involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here.

[0062] It should be understood that the device provided in this embodiment is used to perform the above-described deep learning-based fMRI focal identification method for post-stroke aphasia patients, and therefore can achieve the same effect as the above-described implementation method.

[0063] When using integrated units, the device may include a processing module and a storage module. When applied to a workpiece, the processing module can be used to control and manage the workpiece's operations. The storage module can be used to support the execution of program code by the workpiece.

[0064] The processing module may be a processor or a controller, which can implement or execute various exemplary logic blocks, modules, and circuits contained in conjunction with the disclosure of this application. The processor may also be a combination of functions that implement computing capabilities, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc., and the storage module may be a memory.

[0065] Example 3: This embodiment also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement the deep learning-based fMRI focal identification method for post-stroke aphasia patients provided in the above embodiment.

[0066] Example 4: This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to realize the deep learning-based fMRI focal identification method for post-stroke aphasia patients provided in the above embodiment.

[0067] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0068] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0069] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0070] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A deep learning-based method for identifying focal lesions on fMRI in post-stroke aphasia patients, characterized in that, The method includes: Obtain fMRI image sequences of the brain of a stroke-induced aphasia patient, wherein the fMRI images in the sequences contain different local regions; Local regions of every two adjacent fMRI images in the sequence are matched. Based on the geometric structural similarity and gray-level distribution similarity of the two matched local regions of every two adjacent fMRI images, as well as the degree of image matching, the region consistency value of the corresponding two fMRI images is obtained. Based on the region consistency value, the corresponding time period of the sequence is divided into different time windows. Based on the difference in the regional consistency value of fMRI images within each two adjacent time windows, as well as the differences in the geometric structure features and gray-scale distribution features of the local region, the functional connectivity coefficients for the corresponding two time windows are obtained. The kernel size for deep learning of fMRI images within each time window is determined based on the functional connectivity coefficients, and focal identification is performed on the fMRI images within each time window.

2. The method for identifying focal lesions on fMRI in post-stroke aphasia patients based on deep learning according to claim 1, characterized in that, The step of obtaining the regional consistency value of corresponding two fMRI images includes: A matching region pair is formed by two matching local regions in each pair of adjacent fMRI images; the overall regional difference value of the corresponding two fMRI images is obtained based on the difference in gray value of the pixels of the two local regions in all matching region pairs of each pair of adjacent fMRI images and the difference in slope between adjacent pixels on the edge. The regional consistency value of two fMRI images is obtained by comparing the number of region matching pairs between each pair of adjacent fMRI images with the overall difference value of the region.

3. The method for identifying focal lesions on fMRI in post-stroke aphasia patients based on deep learning according to claim 2, characterized in that, The step of obtaining the overall regional difference value of corresponding two fMRI images includes: The slope sequence is obtained by arranging the slopes between all adjacent pixels on the edge of the local region in order; the average gray value of the pixels in the local region is calculated and recorded as the overall gray value. After calculating the product of the absolute value of the difference in the overall gray values ​​of the two local regions in each matching region pair of each adjacent two fMRI images and the DTW value of the slope sequence, the mean of all products is taken as the overall regional difference value of the corresponding two fMRI images. The formula for calculating the overall regional difference between any two adjacent fMRI images is as follows: ; This represents the overall regional difference value between the c-th and c+1-th fMRI images in the fMRI image sequence; The number of matching region pairs in the c-th and c+1-th fMRI images in the fMRI image sequence; The gray value is the overall gray value of the local region located in the c-th fMRI image in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence. is the overall gray value of the local region located in the (c+1)th fMRI image in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; DTW values ​​of the slope sequences of two local regions in the m-th matching region pair of the c-th and c+1-th fMRI images in the fMRI image sequence; It is an absolute value function.

4. The method for fMRI focal identification of post-stroke aphasia patients based on deep learning according to claim 1, characterized in that, The step of dividing the corresponding time period of the sequence into different time windows based on the regional consistency value includes: Based on the region consistency value, the fMRI images in the fMRI image sequence are clustered to obtain different clusters; Each fMRI image in the fMRI image sequence corresponds to a time point; determine whether the fMRI images within each cluster are continuous in the fMRI image sequence. If so, take the time points corresponding to the fMRI images with the smallest and largest indices within each cluster as the start and end times of a time window, respectively. If not, based on the continuity of fMRI images within each cluster in the fMRI image sequence, the fMRI images within each cluster are divided into different continuous image segments, and the times corresponding to the fMRI images with the smallest and largest subscripts in each continuous image segment are respectively used as the start and end times of a time window.

5. The method for identifying focal lesions on fMRI in post-stroke aphasia patients based on deep learning according to claim 3, characterized in that, The process of obtaining the functional connection coefficients corresponding to the two time windows includes: Based on the geometric structure and grayscale distribution characteristics of the local region of the fMRI image within each time window, the image feature values ​​for the corresponding time window are obtained. Calculate the mean of the region consistency value of all fMRI images within each time window, and negatively correlate and normalize the product of the absolute value of the difference between the mean of each two adjacent time windows and the absolute value of the difference between the image feature values ​​to obtain the functional connectivity coefficients of the corresponding two time windows. The formula for calculating the functional connectivity coefficient between any two adjacent time windows is as follows: In the formula, For the functional connectivity coefficients of the d-th and d+1-th time windows; The mean of the regional consistency values ​​of all two adjacent fMRI images within the d-th time window; The mean of the regional consistency values ​​of all two adjacent fMRI images within the (d+1)th time window; Let be the image feature value for the d-th time window; is the image feature value of the (d+1)th time window; exp is an exponential function with the natural constant as the base; It is an absolute value function.

6. The method for focal identification of fMRI in post-stroke aphasia patients based on deep learning according to claim 5, characterized in that, The step of obtaining the image feature values ​​corresponding to the time window includes: For each fMRI image, the absolute value of the difference between the overall gray value of each local region of the fMRI image and the mean gray value of the pixels in the fMRI image is calculated, and the ratio of the gray value variance of the pixels in each local region is calculated. The sum of all ratios is then used as the gray feature value. The absolute value of the difference between the mean of the elements in the slope sequence of each local region of the fMRI image and the mean of the elements in the slope sequence of all local regions is calculated, and the ratio of the variance of the elements in the slope sequence of each local region is used as the sum of all ratios as the structural feature value. The mean of the products of the grayscale feature values ​​and the structural feature values ​​of all fMRI images within each time window is taken as the image feature value for each time window; the formula for calculating the image feature value is as follows: , , In the formula, K is the image feature value of each time window; V is the total number of fMRI images within each time window; The grayscale feature value of the v-th fMRI image within each time window; The structural feature values ​​of the v-th fMRI image within each time window; The total number of local regions in the v-th fMRI image within each time window; The overall grayscale value of the u-th local region in the v-th fMRI image within each time window; The mean gray value of a pixel in the v-th fMRI image within each time window; The variance of grayscale values ​​of pixels in the u-th local region of the v-th fMRI image within each time window; The mean value of the elements in the slope sequence of the u-th local region of the v-th fMRI image within each time window; The mean value of the elements in the slope sequence of the local region of the v-th fMRI image within each time window; The variance of the elements in the slope sequence of the u-th local region of the v-th fMRI image within each time window.

7. The method for identifying focal lesions on fMRI in post-stroke aphasia patients based on deep learning according to claim 1, characterized in that, The method for obtaining the convolution kernel size includes: The preset convolution kernel size is weighted by the sum of the constant 1 and the functional connection coefficient, and the weighted result is used as the convolution kernel size for deep learning of all fMRI images in the latter time window of each two adjacent time windows.

8. The method for fMRI focal identification of post-stroke aphasia patients based on deep learning according to claim 1, characterized in that, The method used for deep learning of the fMRI images is a CNN model.

9. The method for identifying focal lesions on fMRI in post-stroke aphasia patients based on deep learning according to claim 1, characterized in that, The method for matching local regions of each pair of adjacent fMRI images is a template matching algorithm.

10. The method for identifying focal lesions on fMRI in post-stroke aphasia patients based on deep learning according to claim 7, characterized in that, The preset convolution kernel size is 9.

Images