A hippocampus abnormity atrophy recognition processing method of a brain function magnetic resonance image

Abstract

This invention discloses a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain, comprising: acquiring functional MRI image data of the brain to be processed; preprocessing the functional MRI image data of the brain to obtain preprocessed functional MRI image data of the brain; extracting the hippocampal region from the preprocessed functional MRI image data of the brain; constructing a three-dimensional image of the hippocampus based on the hippocampal region; performing feature extraction on the three-dimensional image of the hippocampal region based on a deep learning network to obtain feature extraction results; and classifying and identifying abnormal atrophy of the hippocampal region according to the feature extraction results to obtain classification and identification results.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more particularly to a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain. Background Technology

[0002] Alzheimer's disease (AD) is a neurodegenerative disease characterized by progressive cognitive impairment and behavioral abnormalities. With the accelerating aging of the global population, the incidence of AD is showing a significant upward trend. Studies indicate that AD accounts for 50%-70% of all dementia cases, making it a major public health problem threatening the health of the elderly population.

[0003] Currently, there are approximately 50 million people living with dementia worldwide, and this number is projected to increase to 152 million by 2050. Age is the most significant risk factor for Alzheimer's disease (AD), with about 8% of people over 65 years of age affected, and the prevalence rate reaching as high as 30% in those over 85. my country has approximately 10 million AD patients, and this number is increasing by 300,000 annually. Early diagnosis and intervention are crucial for improving patient prognosis.

[0004] Currently, traditional methods for detecting Alzheimer's disease (AD) primarily rely on clinical symptom assessment, cerebrospinal fluid biomarker detection, and imaging examinations. However, these methods often only provide a definitive diagnosis in the middle to late stages of the disease, severely impacting the effectiveness of early intervention. Hippocampal atrophy is one of the earliest and most significant imaging changes in AD. Studies have shown that the rate of hippocampal atrophy in AD patients is 2-3 times faster than in normal individuals. Therefore, accurately identifying and quantifying the degree of hippocampal atrophy is crucial for the early diagnosis of AD.

[0005] Traditional diagnostic methods primarily rely on image detection and recognition methods from traditional imaging studies, i.e., computer-aided diagnostic systems. Their application in the early diagnosis of Alzheimer's disease (AD) has become a research hotspot. These methods mainly fall into three categories: traditional image processing, machine learning, and deep learning. In image processing, threshold segmentation, region growing, and edge detection are commonly used for hippocampal region extraction. However, these methods require high image quality and are easily affected by noise. Feature extraction mainly includes morphological features (volume, surface area, etc.), statistical features (gray-level histogram, co-occurrence matrix, etc.), and texture features (wavelet coefficients, fractal dimension, etc.). While these features can reflect some characteristics of the hippocampus, they often require manual design and selection, making it difficult to comprehensively describe complex pathological changes. In classification and recognition, classic algorithms such as Support Vector Machines (SVM), Random Forests (RF), and K-Nearest Neighbors (KNN) are commonly used. However, these methods are highly dependent on the expressive power of features and struggle to handle high-dimensional data. In recent years, some studies have begun to introduce deep learning methods, but the difficulty in acquiring medical image data, high annotation costs, and generally insufficient sample sizes limit the performance of deep learning models. In addition, most existing studies use uniform assessment criteria and do not fully consider the impact of age on hippocampal atrophy, which can easily lead to diagnostic bias.

[0006] The main challenge faced by traditional image detection and recognition methods in identifying hippocampal atrophy lies in the limitations of feature extraction. First, hand-designed features are often based on expert experience, making it difficult to guarantee their completeness and representativeness. For example, simple volume measurements cannot reflect the fine structural changes of the hippocampus, while complex texture features are easily affected by image quality. Second, the feature selection process lacks theoretical guidance, often employing trial-and-error methods or statistical significance tests, making it difficult to ensure that the selected features have optimal discriminative power. Third, different features may be correlated, leading to information redundancy and increasing computational complexity. Furthermore, traditional feature extraction methods struggle to adapt to image differences caused by different scanning devices and parameters, resulting in limited generalization ability. Regarding feature fusion, simple feature stitching or weighted combination fails to fully utilize the complementarity between different features, affecting the final classification result. These limitations make it difficult for traditional machine learning methods to achieve ideal diagnostic accuracy in practical applications.

[0007] Furthermore, current image detection and recognition methods for hippocampal atrophy suffer from significant inconsistencies, particularly the lack of age-based stratification standards. Studies show that hippocampal volume reduction occurs during normal aging, but the rate and pattern of atrophy differ from pathological atrophy. Existing assessment methods often employ uniform threshold standards, neglecting the impact of age on hippocampal volume, potentially leading to higher false-positive rates in older populations and higher false-negative rates in younger populations. Moreover, different studies use varying measurement methods and scoring criteria, lacking comparability. This inconsistency in assessment standards not only affects diagnostic accuracy but also hinders clinical practice and the widespread application of research findings. Therefore, establishing a standardized assessment system based on age stratification is an urgent issue that needs to be addressed. Summary of the Invention

[0008] The purpose of this invention is to provide a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain, which solves the aforementioned technical problems pointed out in the prior art.

[0009] This invention provides a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain, comprising the following steps:

[0010] Acquire functional magnetic resonance imaging (MRI) data of the brain to be processed;

[0011] The functional MRI images of the brain are preprocessed to obtain preprocessed functional MRI images of the brain.

[0012] Extract the hippocampus region from the preprocessed functional magnetic resonance imaging (fMRI) data of the brain; construct a three-dimensional image of the hippocampus based on the hippocampus region;

[0013] Feature extraction is performed on the three-dimensional image of the hippocampus region using a deep learning network to obtain the feature extraction results.

[0014] Based on the feature extraction results, abnormal atrophy of the hippocampal region is classified and identified to obtain the classification results.

[0015] Preferably, after acquiring the functional magnetic resonance imaging data of the brain to be processed, the following steps are also included:

[0016] Obtain the age information of the subjects corresponding to the functional magnetic resonance imaging data of the brain to be processed;

[0017] Extract hippocampal region data from the coronal T1-weighted image of the hippocampal 3D image;

[0018] A quantitative assessment model for the degree of hippocampal atrophy was established based on the hippocampal region data.

[0019] The age information A, the quantitative assessment model for hippocampal atrophy H, and the feature extraction result F are combined to calculate the hippocampal atrophy grading, resulting in the grading result S.

[0020] Preferably, the hippocampal region data includes hippocampal volume data and morphological characteristic parameter data;

[0021] The quantitative assessment model for the degree of hippocampal atrophy includes a hippocampal atrophy rate calculation model, a hippocampal volume loss rate calculation model, and a normal reference value database based on age stratification.

[0022] Preferably, the calculation of the hippocampal atrophy grading determination by combining the age information A, the quantitative assessment model H for hippocampal atrophy degree, and the feature extraction result F to obtain the grading determination result S includes the following steps:

[0023] Based on the age information A, the normal reference value data corresponding to age information A is retrieved from the age-stratified normal reference value database;

[0024] Based on the normal reference value data corresponding to the age information A, the hippocampal atrophy rate calculation model, and the hippocampal volume loss rate calculation model, a quantitative assessment model H' of the degree of hippocampal atrophy after retrieval is established.

[0025] The age information A, the quantitative assessment model for the degree of hippocampal atrophy after retrieval H', and the feature extraction result F are standardized to obtain standardized age information A', standardized quantitative assessment model for the degree of hippocampal atrophy after retrieval H', and standardized feature extraction result F'.

[0026] Based on the standardized age information A', the standardized quantitative assessment model for hippocampal atrophy after retrieval H”, and the standardized feature extraction result F', the grading judgment result S is calculated to obtain the grading judgment result S;

[0027] The calculation method for the hierarchical judgment result S is as follows:

[0028] S=w1×A′+w2×H”+w3×F′+b;

[0029] In the formula, w1, w2, and w3 are weighting coefficients, b is the bias term, A' is the standardized age information, H” is the standardized quantitative assessment model for hippocampal atrophy, and F' is the standardized feature extraction result.

[0030] Preferably, the preprocessing of the functional MRI images of the brain to obtain preprocessed functional MRI images of the brain includes the following steps:

[0031] A nonlocal mean filtering algorithm was used to denoise the functional magnetic resonance imaging data of the brain, resulting in denoised functional magnetic resonance imaging data of the brain.

[0032] The denoised functional magnetic resonance imaging data of the brain were standardized by grayscale value processing to obtain grayscale standardized functional magnetic resonance imaging data of the brain.

[0033] Affine transformation method was used to spatially register the grayscale-normalized functional magnetic resonance imaging data of the brain with a preset standard template to obtain the registered image.

[0034] The registered images were subjected to bias field correction to obtain preprocessed functional magnetic resonance imaging data of the brain.

[0035] Preferably, the step of performing bias field correction on the registered images to obtain preprocessed functional magnetic resonance imaging (fMRI) data of the brain includes the following steps:

[0036] The registered image is subjected to Fourier transform processing to obtain a low-frequency image;

[0037] Based on the low-frequency image as the initial estimate, iterative optimization processing is performed using the N4ITK algorithm to obtain preprocessed functional MRI images of the brain.

[0038] Preferably, the step of using the low-frequency image as an initial estimate and performing iterative optimization processing using the N4ITK algorithm to obtain preprocessed functional magnetic resonance imaging (MRI) data of the brain includes the following steps:

[0039] Initialize iteration parameters;

[0040] The iteration parameters include the low-frequency image L(i,j), the bias field matrix B(i,j), the B-spline grid spacing d, the regularization weight coefficient λ, the spline basis function matrix W(i,j), the learning rate α, and the energy numerical function E(i,j).

[0041] The first image to be corrected, C(i,j), is calculated based on the low-frequency image L(i,j) and the deviation field matrix B(i,j).

[0042] A local region image matrix is ​​obtained by searching the first image to be corrected using a search window of a preset size; the local energy value E is calculated based on the local region image matrix and the energy value function. local(i,j) ;

[0043] Determine the local energy value E local(i,j) Is the local energy value greater than a preset minimum threshold P? If so, then adjust the local energy value E based on the learning rate α. local(i,j)The corresponding local region image matrix is ​​subjected to local iterative optimization to obtain the optimized local region image matrix;

[0044] A second image to be corrected is constructed based on the optimized local image matrix;

[0045] Based on the second image to be corrected and the low-frequency image L(i,j), the bias field is estimated to obtain the updated bias field matrix B'(i,j);

[0046] Based on the second image to be determined and the updated deviation field matrix B'(i,j), the corrected energy value E'(i,j) is calculated using the energy value function E(i,j);

[0047] Determine whether the corrected energy value E'(i,j) is greater than or equal to the preset minimum threshold of the total energy function; if yes, output the second image to be determined as preprocessed functional MRI image data of the brain; if no, return to step S2422 above and iterate again until the preprocessed functional MRI image data of the brain is output.

[0048] Preferably, the energy numerical function E(i,j) is calculated as follows:

[0049]

[0050] In the formula, N and M are the number of spline basis functions and the number of bias fields used for calculation, respectively; L(i,j) is the low-frequency image; B(i,j) is the bias field matrix; d is the B-spline grid spacing; λ is the regularization weight coefficient; and W(i,j) is the spline basis function matrix.

[0051] Preferably, the local energy value E is based on the learning rate α. local(i,j) The corresponding local region image matrix is ​​subjected to local iterative optimization to obtain the optimized local region image matrix, including the following steps:

[0052] Based on the learning rate, gradient descent is used to process the local region image matrix Q. local(i,j) Local optimization is performed to obtain the first optimized image matrix;

[0053] The energy of the first optimized image matrix is ​​calculated based on the energy numerical function E(i,j) to obtain the first local energy value E'. local(i,j) ;

[0054] Determine the first local energy value E' local(i,j)If the value is greater than a preset minimum threshold P for local energy, and if not, output the first optimized image matrix as the optimized local region image matrix; if so, further determine the first local energy value E'. local(i,j) Whether the energy ratio with the minimum local energy value threshold P is greater than or equal to the preset step size adjustment threshold; if so, the learning rate α is adjusted based on the energy ratio to obtain the adjusted first learning rate α';

[0055] The first learning rate α' and the first optimized image matrix are returned to the above operation for re-iteration until the optimized local region image matrix is ​​output.

[0056] Preferably, the first optimized image matrix is ​​calculated as follows:

[0057]

[0058] In the formula, Represents the local energy value E local(i,j) Partial derivatives with respect to the pixel values ​​of each pixel in the current local region image matrix;

[0059] The first learning rate α' is calculated as follows:

[0060]

[0061] Where β is the step size adjustment coefficient.

[0062] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages:

[0063] Analysis of the above-mentioned method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain provided by this invention reveals that, in practical applications, functional MRI images of the brain containing information on brain function, structure, and potential lesions are acquired using medical imaging equipment (such as MRI equipment) to provide raw input for subsequent analysis. Furthermore, preprocessing operations such as bias field correction are performed on the acquired image data to provide high-quality input for subsequent feature extraction and modeling. Then, the hippocampal region is extracted from the preprocessed image data, and a three-dimensional image of the hippocampus is created based on this region. By accurately locating and extracting the hippocampal region, this information is then used for three-dimensional reconstruction. This method presents the morphological features of the hippocampus in greater detail, especially when assessing changes in hippocampal volume (such as atrophy). Furthermore, it utilizes deep learning networks (such as multi-scale convolutional neural networks) to extract features from 3D hippocampal images, acquiring the shape, volume, and texture features of the hippocampus. The combination of multi-scale convolutional neural networks and attention mechanisms enhances the model's adaptability to changes under different image conditions, improving the accuracy and comprehensiveness of feature extraction and more accurately reflecting the pathological characteristics of the hippocampus. Finally, based on the extracted features, abnormal atrophy of the hippocampus is classified and identified, generating classification results (such as normal or abnormal), providing an effective decision support tool for clinicians and assisting doctors in developing treatment plans. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the main process of a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain.

[0065] Figure 2 A schematic diagram of a three-dimensional hippocampal region image in a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain.

[0066] Figure 3 A schematic diagram of the pons in a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain;

[0067] Figure 4 A schematic diagram of a coronal image in a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain.

[0068] Figure 5 This is a schematic diagram illustrating the workflow of obtaining preprocessed functional MRI image data of the brain through iterative optimization processing in a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain. Detailed Implementation

[0069] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0071] like Figure 1 As shown, this embodiment of the invention provides a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain, including the following steps:

[0072] Step S10: Acquire the functional magnetic resonance imaging data of the brain to be processed;

[0073] It should be noted that the above-mentioned functional MRI images of the brain to be processed are acquired by medical imaging equipment (such as MRI equipment) and are the brain functional MRI images of the patient to be processed. They can capture the functional activity, structure and possible lesions of the brain. The above-mentioned functional MRI images of the brain to be processed are multiple planar slices in two-dimensional form (such as transverse or sagittal planes, etc.), and each slice represents a different brain section.

[0074] Step S20: Preprocess the functional magnetic resonance imaging data of the brain to obtain preprocessed functional magnetic resonance imaging data of the brain;

[0075] The preprocessing includes deviation field correction processing;

[0076] It should be noted that in the embodiments of this application described above, the functional magnetic resonance imaging (fMRI) data of the brain often exhibits a low-frequency, gradually changing intensity non-uniformity, called a bias field, due to factors such as magnetic field inhomogeneity and detector response differences. The bias field typically manifests as uneven brightness in the image, which affects quantitative analysis and subsequent processing tasks. Therefore, bias field correction is an important step in improving image quality and the accuracy of subsequent analysis. The main goal of bias field correction is to remove or reduce this low-frequency, non-uniform brightness variation from the image, enabling the image to more accurately reflect the true physical characteristics of the tissue. In particular, removing the bias field can significantly improve the accuracy of the analysis results in subsequent image analysis tasks (such as tissue segmentation and quantitative analysis).

[0077] Step S30: Extract the hippocampus region from the preprocessed functional magnetic resonance imaging (fMRI) data of the brain; construct a three-dimensional image of the hippocampus based on the hippocampus region;

[0078] It should be noted that the aforementioned three-dimensional hippocampal images refer to three-dimensional hippocampal region images (i.e., three-dimensional images, or three-dimensional reconstructed images) constructed based on preprocessed functional magnetic resonance imaging (fMRI) data of the brain. Figure 2 As shown in the figure); the above embodiments of this application, by accurately extracting the hippocampal region and performing three-dimensional modeling, can perform more detailed analysis of the hippocampus, and more accurately assess its morphological changes (such as atrophy or hyperplasia), which is helpful for disease diagnosis; normally, cranial functional magnetic resonance imaging data can be in two-dimensional or three-dimensional form. However, the two-dimensional form (i.e., the cranial functional magnetic resonance imaging data to be processed mentioned above) is presented in the form of two-dimensional images, which is not strong in representing complex structures and is difficult to predict the volume and atrophy of the hippocampus; while the three-dimensional form is a volumetric dataset acquired by the magnetic resonance equipment, which can reconstruct continuous two-dimensional slices in any direction or directly display them in the form of a three-dimensional model. This type of imaging provides more detailed information, especially for complex structures such as the brain, heart or joints, which can provide better anatomical details. However, 3D MRI may require more scanning time and has higher requirements for equipment;

[0079] Based on this, the embodiments of this application adopt the acquisition of two-dimensional functional magnetic resonance imaging (fMRI) images of the brain to be processed, then extract the hippocampal region from the two-dimensional functional fMRI images of the brain to be processed, and then perform three-dimensional reconstruction based on the hippocampal region corresponding to multiple planar slices to obtain a three-dimensional image of the hippocampal region. This reduces the scanning time of the MRI equipment while displaying the loaded brain structure.

[0080] Step S40: Extract features from the three-dimensional image of the hippocampus region using a deep learning network to obtain the feature extraction results;

[0081] It should be noted that the features extracted by deep learning networks in the above embodiments of this application can usually capture potential patterns or subtle changes that are difficult to detect by traditional methods.

[0082] Specifically, the feature extraction described above includes the following steps: First, a multi-scale convolutional neural network is constructed for feature learning to capture feature information at different scales, so that the network can effectively handle objects of different sizes and shapes, enhance the network's robustness to image deformation and scale changes, enable it to perform stably under diverse image conditions in the real world, improve feature expression capabilities, fuse features at different levels, enhance the ability to recognize complex patterns, reduce information loss, retain important global and local information, adapt to diverse inputs, and improve the model's generalization ability; then, the hippocampal shape features, volume features, and texture features are extracted based on the feature learning results; further, an attention mechanism is used to weight and fuse the hippocampal shape features, volume features, and texture features to obtain fused features; finally, a dimensionality reduction algorithm is used to optimize the fused features to obtain the feature extraction results.

[0083] Step S50: Based on the feature extraction results, perform abnormal atrophy classification and identification on the hippocampal region to obtain the classification and identification results.

[0084] It should be noted that the above-described embodiments of this application can quantitatively assess the health status of the hippocampus through classification and identification, and provide reliable decision support for clinical practice. At the same time, classification methods based on deep learning generally have high accuracy and reliability, and can effectively distinguish between normal and abnormal states.

[0085] It should be noted that the above-described embodiments of this application acquire functional MRI images of the brain, containing information on brain function, structure, and potential lesions, from patients using medical imaging equipment (such as MRI equipment), providing raw input for subsequent analysis. Furthermore, preprocessing operations such as bias field correction are performed on the acquired image data to provide high-quality input for subsequent feature extraction and modeling. Then, the hippocampus region is extracted from the preprocessed image data, and a three-dimensional image of the hippocampus is created based on this region. By accurately locating and extracting the hippocampus region, and then using this information for three-dimensional reconstruction, the morphological features of the hippocampus are presented in more detail. It is particularly advantageous in assessing hippocampal volume changes (such as atrophy). Furthermore, deep learning networks (such as multi-scale convolutional neural networks) are used to extract features from 3D hippocampal images to obtain the shape, volume, and texture features of the hippocampus. The combination of multi-scale convolutional neural networks and attention mechanisms improves the model's adaptability to changes under different image conditions, enhances the accuracy and comprehensiveness of feature extraction, and more accurately reflects the pathological characteristics of the hippocampus. Finally, based on the extracted features, abnormal atrophy of the hippocampus is classified and identified, and classification results (such as normal or abnormal) are generated, providing an effective decision support tool for clinicians to assist doctors in formulating treatment plans.

[0086] Specifically, after acquiring the functional magnetic resonance imaging data of the brain to be processed in step S10, the following steps are also included:

[0087] Step S60: Obtain the age information of the subject corresponding to the functional magnetic resonance imaging data of the brain to be processed;

[0088] Step S70: Extract hippocampal region data from the coronal T1-weighted image of the hippocampal three-dimensional image;

[0089] The hippocampal region data includes hippocampal volume data and morphological characteristic parameter data;

[0090] It should be noted that the extraction of the hippocampus region mentioned above first involves locating the anterior pons at the horizontal level, such as... Figure 3 As shown, the pons is a structure located in the brainstem, between the medulla oblongata and the midbrain. Its position is relatively fixed in brain imaging and can usually be used as a reference for localization. The anterior horizontal level of the pons can be located by manually marking or setting a reference level based on anatomical knowledge.

[0091] Furthermore, after locating the anterior pons at a horizontal level, selecting a slice passing through the hippocampus in coronal T1-weighted images allows for accurate extraction of hippocampal structural information. The hippocampus, typically located medially in the temporal lobe, primarily functions in memory storage and processing. T1-weighted images offer good tissue contrast, particularly in the imaging of brain structures like the hippocampus. The coronal plane refers to a slicing method from anterior to posterior, such as... Figure 4 As shown, the morphology of the hippocampus can usually be displayed relatively accurately on this plane; by analyzing the coronal image sequence, slices related to the hippocampus can be identified, and these slices are usually located in the medial temporal lobe, passing through the cross section of the hippocampus.

[0092] Furthermore, the hippocampal contour is obtained by pre-training a deep learning segmentation network (such as a convolutional neural network CNN). Then, the hippocampal volume and morphological feature parameters are calculated. Specifically, the hippocampal volume is estimated by calculating the number of voxels contained in the segmented contour of the hippocampal region. The surface area of ​​the hippocampus is calculated by performing surface reconstruction on the segmented contour of the hippocampus (such as using the MarchingCubes algorithm). Geometric features of the hippocampus, such as symmetry, shape factor, and curvature, are extracted using morphological feature extraction algorithms (such as region-based features, boundary features, etc.). By statistically analyzing the gray values ​​in the hippocampal region, some features related to shrinkage (such as gray matter density, texture features, etc.) can be extracted.

[0093] Step S80: Establish a quantitative assessment model for the degree of hippocampal atrophy based on the hippocampal region data;

[0094] The quantitative assessment model for the degree of hippocampal atrophy includes a hippocampal atrophy rate calculation model, a hippocampal volume loss rate calculation model, and a normal reference value database based on age stratification.

[0095] Step S90: Combine the age information A, the quantitative assessment model H of hippocampal atrophy degree (definitely incorrect), and the feature extraction result F to calculate the hippocampal atrophy grading determination, and obtain the grading determination result S (the grading determination result is the classification recognition result in step S50 above).

[0096] It should be noted that the calculated result S of the above-mentioned grading judgment is a numerical value. After the grading judgment result is calculated, it is necessary to compare the numerical value of the grading judgment result with the preset grading value range. When the grading judgment result is less than or equal to the mild atrophy threshold S1, the grading judgment result is mild atrophy level; when the grading judgment result is greater than the mild atrophy threshold S1 and less than or equal to the moderate atrophy threshold S2, the grading judgment result is moderate atrophy; when the grading judgment result is greater than the moderate atrophy threshold S2, the grading judgment result is severe atrophy.

[0097] The above-described embodiments of this application first obtain the age information of the subject corresponding to the functional magnetic resonance imaging (fMRI) images of the brain to be processed. Age is one of the important factors in hippocampal atrophy; the hippocampus may atrophy with age, so age information is crucial for subsequent atrophy assessment. Introducing age information provides a physiological background for hippocampal atrophy assessment, as normal hippocampal atrophy is closely related to age, especially in the elderly. Combining age information with age-stratified normal reference values ​​allows for comparison to determine whether the subject's hippocampus has experienced abnormal atrophy, avoiding misdiagnosis or misassessment caused by neglecting age differences in traditional hippocampal atrophy analysis methods. Furthermore, hippocampal region data are extracted from coronal T1-weighted images of the hippocampus. Coronal images have good structural contrast and can accurately display the morphology of the hippocampus, providing a clear basic image for subsequent analysis. Weighted images can effectively distinguish brain structures, especially suitable for displaying the structure of the hippocampus. Their high contrast makes the details of the hippocampus in the image more obvious, accurately capturing atrophy or other morphological changes. Furthermore, the extracted hippocampal region data (volume and morphological features) are applied to a quantitative assessment model of hippocampal atrophy, enabling quantitative analysis of hippocampal atrophy. It is particularly important that an age-stratified reference database is crucial, as hippocampal atrophy is not only related to diseases (such as Alzheimer's disease) but also to the normal aging process. By comparing with normal reference values, physiological and pathological atrophy can be effectively distinguished. Finally, combining the age information obtained in step S60, the quantitative assessment model of hippocampal atrophy (H) established in step S80, and the feature extraction results (F) extracted in step S40, a grading determination of hippocampal atrophy is made, outputting the grading result (S), indicating whether abnormal hippocampal atrophy exists.

[0098] Specifically, in step S90, the age information A, the quantitative assessment model H for the degree of hippocampal atrophy, and the feature extraction result F are combined to calculate the hippocampal atrophy grading, resulting in the grading result S. This includes the following steps:

[0099] Step S91: Based on the age information A, retrieve the normal reference value data corresponding to the age information A from the age-stratified normal reference value database;

[0100] Step S92: Based on the normal reference value data corresponding to the age information A, the hippocampal atrophy rate calculation model, and the hippocampal volume loss rate calculation model, a quantitative assessment model H' of the degree of hippocampal atrophy after retrieval is established;

[0101] Step S93: Standardize the age information A, the quantitative assessment model H' of the degree of hippocampal atrophy after retrieval, and the feature extraction result F to obtain standardized age information A', standardized quantitative assessment model H' of the degree of hippocampal atrophy after retrieval, and standardized feature extraction result F'.

[0102] Step S92: Calculate the grading judgment result S based on the standardized age information A', the standardized retrieval post-retrieval hippocampal atrophy degree quantitative assessment model H”, and the standardized feature extraction result F', to obtain the grading judgment result S;

[0103] The calculation method for the hierarchical judgment result S is as follows:

[0104] S=w1×A′+w2×H”+w3×F′+b;

[0105] In the formula, w1, w2, and w3 are weighting coefficients, b is the bias term, A' is the standardized age information, H” is the standardized quantitative assessment model for hippocampal atrophy, and F' is the standardized feature extraction result;

[0106] It should be noted that the above embodiments of this application standardize the age information A, the quantitative assessment model H for the degree of hippocampal atrophy, and the feature extraction result F to ensure that features at different scales can be compared and calculated in the same way, thereby more accurately assessing the degree of hippocampal atrophy.

[0107] Specifically, in step S80, a quantitative assessment model for the degree of hippocampal atrophy is established based on the hippocampal region data, including the following steps:

[0108] Step S81: Construct a hippocampal atrophy rate calculation model based on hippocampal volume data from multiple time nodes; and calculate the hippocampal volume loss rate calculation model based on the initial hippocampal volume data of the initial time node and the final hippocampal volume data of the final time node, as well as the time interval between the initial and final time nodes.

[0109] The calculation method for the annual hippocampal atrophy rate calculation model is as follows:

[0110]

[0111] The calculation method of the hippocampal volume loss velocity calculation model is as follows:

[0112]

[0113] It should be noted that the above-described embodiments of this application involve collecting hippocampal volume data at multiple time points (e.g., every few years or every few months); then, calculating the annual atrophy rate of the hippocampal volume data at each time point corresponding to the hippocampal volume data at the previous time point based on the hippocampal volume data at multiple time points; and evaluating the atrophy rate of the hippocampus through the above-described hippocampal atrophy rate calculation model. Generally, a higher annual atrophy rate usually indicates more significant atrophy, which may be related to certain neurodegenerative diseases (such as Alzheimer's disease) and can serve as a reference for early disease monitoring.

[0114] Step S82: Establish an age-stratified database of normal reference values ​​by statistically analyzing hippocampal volume data from multiple normal subjects across various age groups;

[0115] It should be noted that the embodiments of this application, by collecting hippocampal volume data at multiple time points, establish a hippocampal atrophy rate calculation model and a hippocampal volume loss rate calculation model, thereby quantitatively assessing the atrophy rate of the hippocampus. By introducing multi-time point data, changes in the hippocampus can be dynamically tracked. Compared to static analysis (data from a single time point), multi-time point analysis can more accurately assess the atrophy trend of the hippocampus, providing a basis for early disease detection and monitoring. The annual atrophy rate, as an important indicator, can help clinicians understand the progression rate of hippocampal atrophy. A higher annual atrophy rate usually indicates significant atrophy, which may be a sign of neurodegenerative diseases. A clinical manifestation of diseases such as Alzheimer's can be monitored through long-term dynamic data, providing data support for the early diagnosis of the disease. The hippocampal volume loss rate calculation model provides a more refined assessment of the atrophy rate. By comparing volume changes at different time points, it can reveal the rate of atrophy and thus determine whether there is abnormal rapid atrophy, which is crucial for early intervention and treatment. Furthermore, by statistically analyzing the hippocampal volume data of multiple normal subjects at different ages, an age-stratified database of normal reference values ​​can be established as a standard reference for subsequent hippocampal atrophy assessment, helping to determine whether there are abnormalities in the subject's hippocampus.

[0116] Specifically, in step S20, the functional magnetic resonance imaging (fMRI) data of the brain is preprocessed to obtain preprocessed functional fMRI data of the brain, including the following steps:

[0117] Step S21: The nonlocal mean filtering algorithm is used to denoise the functional magnetic resonance imaging data of the brain to obtain the denoised functional magnetic resonance imaging data of the brain.

[0118] Step S22: Perform grayscale standardization on the denoised functional magnetic resonance imaging data of the brain to obtain grayscale standardized functional magnetic resonance imaging data of the brain.

[0119] Step S23: Use affine transformation to spatially register the grayscale-normalized functional MRI images of the brain with a preset standard template to obtain the registered images;

[0120] Step S24: Perform bias field correction on the registered images to obtain preprocessed functional MRI images of the brain.

[0121] It should be noted that the above-described embodiments of this application first employ a nonlocal mean filtering algorithm to denoise the functional MRI images of the brain. Denoising is an important step in MRI image processing because functional MRI images of the brain are often affected by various noises (such as thermal noise, motion artifacts, etc.), affecting image quality. Denoising the images can more accurately reflect the structure and function of the brain, ensuring the reliability of subsequent analysis. It can better balance noise removal and detail preservation, and is suitable for situations in medical imaging where the preservation of fine structures is required, such as the evaluation of small structures like the hippocampus. Furthermore, grayscale standardization is used to unify the grayscale range of the images, making the data from different images consistent and preventing analytical biases caused by grayscale differences, ensuring the accuracy of model training and testing. Then, an affine transformation method is used to denoise the grayscale-standardized functional MRI images of the brain. Spatial registration is performed between the stromal image data and a preset standard template. This is because the brains of different patients may differ spatially (such as head position, brain size, etc.). Registration aligns the brains of different individuals into a unified spatial coordinate system, ensuring accurate matching of anatomical structures and providing a consistent spatial framework for the assessment of atrophy of small structures such as the hippocampus, avoiding assessment errors caused by inconsistent spatial positions. Finally, bias field correction is used to remove low-frequency signal deviations caused by magnetic field inhomogeneity or scanning equipment. These deviations usually lead to uneven image brightness, affecting image contrast and quality, resulting in a more uniform image, especially with clearer boundaries between gray matter, white matter, and cerebrospinal fluid. This improves image contrast and accuracy, further enhancing the reliability of subsequent processing steps (such as feature extraction and modeling), and thus improving the accuracy of atrophy assessment.

[0122] Specifically, in step S24, the registered image is subjected to bias field correction to obtain preprocessed functional magnetic resonance imaging data of the brain, including the following steps:

[0123] Step S241: Perform Fourier transform processing on the registered image to obtain a low-frequency image;

[0124] It should be noted that in the above embodiments of this application, the registered image is processed by Fourier transform to separate the low-frequency components and high-frequency components of the registered image, and then the noise interference of the high-frequency components is filtered out, while the low-frequency components (i.e. the low-frequency image mentioned above) are retained, which simplifies the subsequent operation of bias field estimation.

[0125] Step S242: Based on the low-frequency image as the initial estimate, perform iterative optimization processing using the N4ITK algorithm to obtain preprocessed functional MRI image data of the brain.

[0126] It should be noted that the above-described embodiments of this application first capture global low-frequency changes through Fourier transform processing to provide a preliminary bias field estimate. Then, the N4ITK algorithm iterative optimization processing operation captures local details and inhomogeneities. The combination of Fourier transform and N4ITK algorithm iterative processing reduces the number of iterations of the traditional N4ITK algorithm and improves the convergence speed of the algorithm. The above-described N4ITK algorithm (N4ITKiasFieldCorrection) is an algorithm for medical image processing, especially for removing the influence of "bias field" in images. The bias field is a low-frequency intensity inhomogeneity that usually appears in medical images, especially in MRI (magnetic resonance imaging) images. This inhomogeneity may be caused by a variety of factors, such as scanner hardware limitations, magnetic field inhomogeneity, etc.

[0127] The aforementioned functional magnetic resonance imaging (fMRI) data of the brain often exhibits a low-frequency, gradually varying intensity non-uniformity, known as a bias field, due to factors such as magnetic field inhomogeneity and detector response differences. The bias field typically manifests as uneven image brightness, which affects quantitative analysis and subsequent processing tasks. Therefore, bias field correction is a crucial step in improving image quality and the accuracy of subsequent analysis. The main goal of bias field correction is to remove or reduce this low-frequency, non-uniform brightness variation from the image, enabling the image to more accurately reflect the true physical characteristics of the tissue. Especially in subsequent image analysis tasks (such as tissue segmentation and quantitative analysis), removing the bias field can significantly improve the accuracy of the analysis results.

[0128] Specifically, such as Figure 5 As shown, in step S242, based on the low-frequency image as the initial estimate, iterative optimization processing is performed using the N4ITK algorithm to obtain preprocessed functional MRI data of the brain, including the following steps:

[0129] Step S2421: Initialize iteration parameters;

[0130] The iteration parameters include the low-frequency image L(i,j), the bias field matrix B(i,j), the B-spline grid spacing d, the regularization weight coefficient λ, the spline basis function matrix W(i,j), the learning rate α, and the energy numerical function E(i,j).

[0131] The energy numerical function E(i,j) is calculated as follows:

[0132]

[0133] In the formula, N and M are the number of spline basis functions and the number of bias fields used for calculation, respectively; L(i,j) is the low-frequency image; B(i,j) is the bias field matrix; d is the B-spline grid spacing; λ is the regularization weight coefficient; W(i,j) is the spline basis function matrix;

[0134] It should be noted that in the above embodiments of this application, the coordinates of each pixel of the low-frequency image are obtained as the "low-frequency image" in the initial iteration parameters. It is represented in matrix form, where each element of the matrix represents the coordinates of a pixel of the low-frequency image. The deviation field matrix represents the deviation field of the initial low-frequency image, and it is initially set to represent the identity matrix. The correction image is also represented in matrix form, and it is initially equal to the low-frequency image.

[0135] Step S2422: Calculate the first image to be corrected, C(i,j), based on the low-frequency image L(i,j) and the deviation field matrix B(i,j);

[0136] It should be noted that the above embodiments of this application correct the low-frequency image by applying an initial deviation field to obtain a first version of the corrected image. The brightness and contrast of the image are adjusted by the initial deviation field correction matrix to make them more consistent. The generated first corrected image to be determined provides a preliminary reference for local optimization in subsequent steps.

[0137] Step S2423: Search the first image to be corrected based on a search window of a preset size to obtain a local region image matrix; calculate the local energy value E based on the local region image matrix and the energy value function. local(i,j) ;

[0138] It should be noted that the above embodiments of this application identify local deviations and inconsistencies in the image through local calculation of local regions and feedback of energy numerical functions, providing local energy numerical feedback for iterative optimization and guiding the optimization process to adjust the image more finely.

[0139] Step S2424: Determine the local energy value E local(i,j)Is the local energy value greater than a preset minimum threshold P? If yes (if no, then the local region image matrix is ​​determined to be the optimized local region image matrix), then the local energy value E is adjusted based on the learning rate α. local(i,j) The corresponding local region image matrix is ​​subjected to local iterative optimization to obtain the optimized local region image matrix;

[0140] It should be noted that the above embodiments of this application utilize the learning rate to adjust the step size, and refine the local area through local iteration to achieve higher image quality. Through local optimization, the image gradually approaches the real correction effect in the detail area.

[0141] Step S2425: Construct a second image to be corrected based on the optimized local image matrix;

[0142] Step S2426: Based on the second image to be corrected and the low-frequency image L(i,j), perform bias field estimation to obtain the updated bias field matrix B'(i,j);

[0143] It should be noted that the above-described embodiments of this application can gradually eliminate uneven brightness and artifacts in the image by continuously adjusting the deviation field, and the updated deviation field provides a more accurate correction basis for the next iteration.

[0144] Step S2427: Based on the second image to be determined and the updated bias field matrix B'(i,j), the corrected energy value E'(i,j) is calculated using the energy value function E(i,j);

[0145] It should be noted that in the above embodiment of this application, the second image to be corrected is taken as the low-frequency image L(i,j), and the updated bias field matrix B'(i,j) is taken as the initial bias field matrix B(i,j). The corrected energy value E'(i,j) is calculated by the energy value function E(i,j).

[0146] Step S2428: Determine whether the corrected energy value E'(i,j) is greater than or equal to the preset minimum threshold of the total energy function; if yes, output the second image to be determined as preprocessed functional MRI image data of the brain; if no, return to step S2422 above and iterate again until the preprocessed functional MRI image data of the brain is output.

[0147] It should be noted that the above embodiments of this application gradually remove the brightness unevenness and artifacts in the original image (i.e., the low-frequency image mentioned above) by updating the bias field and local optimization, thereby improving the image quality. Furthermore, through regularization terms and local iterative optimization, the intensity distribution of the image in the overall and local regions becomes more consistent. Through the feedback of the energy numerical function during the iteration process, the gradual optimization of the image is ensured, and the preset correction effect is finally achieved, avoiding overfitting.

[0148] The embodiments described above first provide the necessary initial conditions for subsequent bias field correction by initializing parameters, ensuring that the iterative process has a clear starting point and can continuously approach the target (removing the bias field). Furthermore, the initialization of the low-frequency image and the bias field matrix provides the foundation for subsequent optimization and iteration, enabling subsequent steps to eliminate inhomogeneities in the image through gradual correction. Then, by initially applying the bias field matrix to correct the low-frequency image, a first-version corrected image is obtained, removing low-frequency uneven brightness variations and providing a preliminary unified brightness correction effect. This provides the foundation for subsequent local optimization, especially when there are more complex biases in details and local regions of cranial functional MRI images, allowing for more refined adjustments through subsequent iterations. Further, by selecting local regions in the image and calculating local energy values ​​based on energy value functions, biases and inconsistencies in local regions are identified. Through local region optimization, bias field correction is not limited to global adjustment but can also finely handle local brightness unevenness, helping to identify subtle and irregular biases in the image, thereby improving the image correction accuracy, especially for small-area biases in cranial functional MRI images. The problem of local brightness unevenness is addressed. Further, through local iterative optimization, local regions in the image are finely adjusted to gradually eliminate bias and brightness unevenness in the functional MRI images of the brain. The learning rate controls the step size of each iteration to avoid over-adjustment or slow convergence. Furthermore, the second version of the corrected image undergoes local iterative optimization, further improving image quality in detail. This image is then used as input to update the bias field matrix, providing a basis for the next round of correction. Further, by comparing the second image to be corrected with the low-frequency image, the bias field matrix is ​​estimated and updated, further improving the brightness uniformity of the image. Further, the corrected energy value is calculated using an energy numerical function to evaluate the bias field correction effect and provide a basis for determining whether correction is complete. By judging whether the bias field correction has reached the predetermined accuracy threshold, if it has, the final preprocessed functional MRI image of the brain is output; otherwise, the iteration step is returned to continue optimization until the bias field problem is completely resolved. The final output corrected image accurately reflects the true physical characteristics of the brain, avoiding the impact of brightness unevenness caused by the bias field on subsequent analysis.

[0149] Specifically, in step S2424, the local energy value E is calculated based on the learning rate α. local(i,j) The corresponding local region image matrix is ​​subjected to local iterative optimization to obtain the optimized local region image matrix, including the following steps:

[0150] Step S24241: Based on the learning rate, use gradient descent to process the local region image matrix Q. local(i,j) Local optimization is performed to obtain the first optimized image matrix;

[0151] The first optimized image matrix is ​​calculated as follows:

[0152]

[0153] In the formula, Represents the local energy value E local(i,j) The partial derivatives of the pixel values ​​of each pixel in the current local region image matrix (representing the influence of that pixel on the first optimized image matrix);

[0154] Step S24242: Calculate the energy of the first optimized image matrix based on the energy numerical function E(i,j) to obtain the first local energy value E'. local(i,j) ;

[0155] Step S24243: Determine the first local energy value E' local(i,j) If the value is greater than a preset minimum threshold P for local energy, and if not, output the first optimized image matrix as the optimized local region image matrix; if so, further determine the first local energy value E'. local(i,j) If the energy ratio to the minimum local energy value threshold P is greater than or equal to the preset step size adjustment threshold, then if yes (if no, return directly to step S24241 above to iterate again), then adjust the learning rate α based on the energy ratio to obtain the adjusted first learning rate α'.

[0156] The first learning rate α' is calculated as follows:

[0157]

[0158] Where β is the step size adjustment coefficient (controlling the sensitivity of learning rate adjustment);

[0159] Step S24244: Return the first learning rate α' and the first optimized image matrix to step S24241 above for re-iteration until the optimized local region image matrix is ​​output.

[0160] It should be noted that the above embodiments of this application first optimize the local region of the image using gradient descent to reduce the image's energy function value, optimize deviation and brightness unevenness; further, the "quality" of the optimized image in the current local region is quantified by calculating the energy value of the updated image matrix; further, by comparing the first local energy value with a preset minimum threshold, it is determined whether the current optimization has met the requirements. If the requirements have not been met, it indicates that the optimization degree of the image is insufficient, and there is still a large brightness unevenness or deviation field problem, requiring continued iterative optimization; further, if the first local energy value is greater than... The minimum threshold is then used to compare the ratio of the first local energy value to the minimum threshold with a preset step size adjustment threshold. If the ratio is large, it means that the optimization step size is too large, which may lead to unstable optimization or over-adjustment. Therefore, the learning rate needs to be adjusted to obtain the adjusted first learning rate (α'). When it is determined that optimization needs to continue (i.e., the energy value has not reached the minimum threshold), the learning rate is adjusted according to the energy ratio. The adjusted learning rate will affect the step size of the next iteration to avoid the step size being too large or too small. The adjusted learning rate allows for better control of the optimization degree of image details during each iteration when performing local optimization.

[0161] The method for identifying and processing hippocampal atrophy in functional MRI images of the brain provided by this invention innovatively introduces age stratification standards and establishes a database of normal value ranges based on age. The MTA score calculation employs an improved automatic scoring algorithm, transforming qualitative assessment into quantitative indicators. Studies have found that the dynamic assessment method described in this invention can better monitor disease progression, and the standardized scoring system improves the comparability of results between different centers. Clinical trials have validated the assessment model provided by this invention, showing an overall diagnostic accuracy rate of 89.5%–90% (which is also the model's accuracy rate), indicating high reliability of the model in clinical applications.

[0162] In summary, the present invention proposes a method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain. This method uses medical imaging equipment (such as MRI) to acquire functional MRI images of the brain containing information on brain function, structure, and potential lesions, providing raw input for subsequent analysis. Furthermore, preprocessing operations such as bias field correction are performed on the acquired image data to provide high-quality input for subsequent feature extraction and modeling. Then, the hippocampal region is extracted from the preprocessed image data, and a three-dimensional image of the hippocampus is created based on this region. By accurately locating and extracting the hippocampal region, and then using this information for three-dimensional reconstruction, the morphological characteristics of the hippocampus are presented in more detail, especially when assessing changes in hippocampal volume (such as atrophy). Finally, deep learning networks (such as multi-scale convolutional neural networks) are used to extract features from the three-dimensional hippocampal image. This method extracts the shape, volume, and texture features of the hippocampus. By combining multi-scale convolutional neural networks and attention mechanisms, it improves the model's adaptability to changes under different image conditions, enhancing the accuracy and comprehensiveness of feature extraction and more accurately reflecting the pathological characteristics of the hippocampus. Finally, based on the extracted features, it classifies and identifies abnormal hippocampal atrophy and generates classification results (e.g., normal or abnormal), providing an effective decision support tool for clinicians to assist in formulating treatment plans. Furthermore, by introducing age information, it provides a physiological background for hippocampal atrophy assessment, as normal hippocampal atrophy is closely related to age, especially in the elderly. Combining age information allows for comparison with age-stratified normal reference values ​​to determine whether the subject's hippocampus has experienced abnormal atrophy, avoiding misdiagnosis or misassessment caused by neglecting age differences in traditional hippocampal atrophy analysis methods.

[0163] Specifically, in the preprocessing operation of bias field correction, Fourier transform is used to capture global low-frequency changes to provide an initial bias field estimate. Then, the N4ITK algorithm iterative optimization operation is used to capture local details and inhomogeneities. The combination of Fourier transform and N4ITK algorithm iterative processing reduces the number of iterations of the traditional N4ITK algorithm and improves the convergence speed of the algorithm, thereby enabling faster and more accurate output of preprocessed cranial functional magnetic resonance images and improving subsequent hippocampal atrophy identification and analysis.

[0164] Furthermore, through local region optimization, the deviation field correction is not limited to global adjustment, but can also finely process local brightness unevenness, which helps to identify subtle and irregular deviations in the image, thereby improving the image correction accuracy.

[0165] Furthermore, by optimizing local regions of the image using gradient descent and adjusting the learning rate to influence the step size of the next iteration, the step size is avoided from being too large or too small. The adjusted learning rate allows for better control over the degree of image detail optimization during each iteration, thereby obtaining more accurate preprocessed functional MRI images of the brain and improving subsequent hippocampal atrophy identification and analysis.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain, characterized in that, The following steps are included: Acquire functional magnetic resonance imaging (MRI) data of the brain to be processed; The cranial functional magnetic resonance imaging (fMRI) image data is preprocessed to obtain preprocessed cranial functional MRI image data, including: denoising the cranial functional MRI image data using a nonlocal mean filtering algorithm to obtain denoised cranial functional MRI image data; standardizing the grayscale values ​​of the denoised cranial functional MRI image data to obtain grayscale-standardized cranial functional MRI image data; and spatially registering the grayscale-standardized cranial functional MRI image data with a preset standard template using an affine transformation method to obtain a registered image. The registered image is subjected to Fourier transform processing to obtain a low-frequency image; Based on the low-frequency image as an initial estimate, iterative optimization processing using the N4ITK algorithm yields preprocessed functional magnetic resonance imaging (fMRI) data of the brain, including: S1: Initialize iteration parameters; the iteration parameters include low-frequency images. Deviation field matrix B-spline mesh spacing d, regularization weight coefficient spline basis function matrix W Learning rate and energy numerical function; The energy numerical function E(i,j) is calculated as follows: ； In the formula, and These are the number of spline basis functions and the number of bias fields used for calculation, respectively. S2: The first image to be corrected is calculated based on the low-frequency image and the deviation field matrix; S3: Search the first image to be corrected based on a search window of a preset size to obtain a local region image matrix; calculate the local energy value based on the local region image matrix and the energy value function; S4: Determine whether the local energy value is greater than a preset minimum threshold for local energy value; if so, perform local iterative optimization on the local region image matrix corresponding to the local energy value based on the learning rate to obtain the optimized local region image matrix. S5: Construct a second image to be corrected based on the optimized local region image matrix; perform bias field estimation based on the second image to be corrected and the low-frequency image to obtain an updated bias field matrix; calculate the corrected energy value based on the second image to be corrected and the updated bias field matrix using an energy value function. S6: Determine whether the corrected energy value is less than or equal to the preset minimum threshold of the total energy function; if yes, output the second image to be determined as preprocessed functional MRI image data of the brain; if no, return to the above S2 operation and iterate again until the preprocessed functional MRI image data of the brain is output. Extract the hippocampus region from the preprocessed functional magnetic resonance imaging (fMRI) data of the brain; construct a three-dimensional image of the hippocampus based on the hippocampus region; Feature extraction is performed on the three-dimensional image of the hippocampus based on a deep learning network to obtain the feature extraction results; Based on the feature extraction results, abnormal atrophy of the hippocampus region is classified and identified, and the classification and identification results are obtained.

2. The method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain according to claim 1, characterized in that, After acquiring the functional magnetic resonance imaging data of the brain to be processed, the following steps are also included: Obtain the age information of the subjects corresponding to the functional magnetic resonance imaging data of the brain to be processed; Extract hippocampal region data from the coronal T1-weighted image of the hippocampus 3D image; A quantitative assessment model for the degree of hippocampal atrophy was established based on the hippocampal region data. The age information A, the quantitative assessment model for hippocampal atrophy H, and the feature extraction result F are combined to calculate the hippocampal atrophy grading, resulting in the grading result S.

3. The method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain according to claim 2, characterized in that, The hippocampal region data includes hippocampal volume data and morphological characteristic parameter data; The quantitative assessment model for the degree of hippocampal atrophy includes a hippocampal atrophy rate calculation model, a hippocampal volume loss rate calculation model, and a normal reference value database based on age stratification.

4. The method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain according to claim 3, characterized in that, The calculation of hippocampal atrophy grading based on the age information A, the quantitative assessment model H for hippocampal atrophy, and the feature extraction result F, to obtain the grading result S, includes the following steps: Based on the age information A, the normal reference value data corresponding to age information A is retrieved from the age-stratified normal reference value database; Based on the normal reference value data corresponding to the age information A, the hippocampal atrophy rate calculation model, and the hippocampal volume loss rate calculation model, a quantitative assessment model H' of the degree of hippocampal atrophy after retrieval is established. The age information A, the quantitative assessment model for the degree of hippocampal atrophy after retrieval H', and the feature extraction result F are standardized to obtain standardized age information A', standardized quantitative assessment model for the degree of hippocampal atrophy after retrieval H'', and standardized feature extraction result F'. Based on the standardized age information A', the standardized retrieval post-retrieval hippocampal atrophy quantitative assessment model H'', and the standardized feature extraction result F', the grading determination result S is calculated to obtain the grading determination result S; The calculation method for the classification result S is as follows: ； In the formula, , and These are the weighting coefficients; For bias terms; To standardize age information; To standardize the quantitative assessment model for the degree of hippocampal atrophy after retrieval; This is the result of standardized feature extraction.

5. The method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain according to claim 4, characterized in that, The step of performing local iterative optimization on the local region image matrix corresponding to the local energy value based on the learning rate to obtain the optimized local region image matrix includes the following steps: S41: Based on the learning rate, the gradient descent method is used to process the local region image matrix. Local optimization is performed to obtain the first optimized image matrix; S42: Calculate the energy of the first optimized image matrix based on the energy numerical function E(i,j) to obtain the first local energy value. ; S43: Determine the value of the first local energy. Is it greater than the preset minimum threshold for local energy value? If not, output the first optimized image matrix as the optimized local region image matrix; if yes, further determine the value of the first local energy. With the minimum threshold of the local energy value If the energy ratio is greater than or equal to the preset step size adjustment threshold, then the learning rate α is adjusted based on the energy ratio to obtain the adjusted first learning rate α'. S44: Return the first learning rate α' and the first optimized image matrix to the operation in S41 above and iterate again until the optimized local region image matrix is ​​obtained as the output.

6. The method for identifying and processing abnormal hippocampal atrophy in functional MRI images of the brain according to claim 5, characterized in that, The first optimized image matrix is ​​calculated as follows: ; In the formula, Represents local energy values Partial derivatives with respect to the pixel values ​​of each pixel in the current local region image matrix; The first learning rate α' is calculated as follows: ； in, This is the step size adjustment factor.

Images