Alzheimer's disease pathological image analysis method and system based on deep learning

Abstract

The invention relates to the technical field of image analysis, in particular to an Alzheimer's disease pathological image analysis method and system based on deep learning. Comprising the following steps: acquiring a plurality of brain medical images of a patient, aligning the brain medical images with a reference image, calculating a weight vector based on the definition, brightness, contrast ratio and area proportion of a target area, and performing weighted fusion to generate a high-quality key image; dividing the key image into a plurality of sub-images, marking the lesion probability in combination with historical definite diagnosis information, constructing a spiral stacked graph by taking the high-probability sub-image as a center, and retaining a local spatial relationship; adopting a dynamic weight adjustment algorithm to train a neural network model, balancing normal and diseased sample data, and enhancing generalization ability; and inputting the spiral stack graph into a model to predict a risk value, generating a comprehensive risk score in combination with a statistical weight of a historical case database, and dividing risk grades. According to the method, the problem of insufficient pathological image analysis precision in the prior art is solved, and the diagnosis accuracy and reliability are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of image analysis technology, and in particular to a method and system for analyzing Alzheimer's disease pathological images based on deep learning. Background Art

[0002] The early diagnosis of Alzheimer's disease relies on the accurate analysis of brain pathological features. Existing technologies mostly adopt multi-modal data fusion or complex image registration methods. Similar existing technologies include a Chinese patent with the publication number CN117995415A, which proposes an early diagnosis method for Alzheimer's disease based on adaptive hypergraph representation. By extracting features from multi-modal data, hypergraphs are constructed respectively. The complex high-order relationships between subjects are modeled using hypergraphs. Subsequently, each modal data is fed into an adaptive hypergraph convolutional neural network. Between two hypergraph convolutional layers, the hypergraph structure is adaptively updated. After the first layer of convolution, the high-order relationships between subjects are further explored, and the corresponding hypergraph structure is updated, thereby improving the quality of subsequent features. After convolution, the high-order features of each modality are subjected to feature fusion based on the attention mechanism. Different from ordinary feature splicing, using the attention mechanism can highlight important features in each modality and suppress noise features, and can automatically discover the correlation and complementarity between different modalities, thus realizing the early diagnosis of Alzheimer's disease. In addition, a similar existing technology is a US patent with the publication number US20230263457A1, which proposes a device and method for providing information required for dementia diagnosis. In the device and method according to the preferred embodiment of the present invention, an MRI brain image and a PET brain image of a subject for diagnosis are received, the MRI brain image is divided into multiple regions, and then the MRI brain image is registered with the PET brain image. In addition to simply providing the information required for dementia diagnosis, it is also possible to provide various information useful for dementia diagnosis to medical staff performing dementia diagnosis using PET images. The above two patent documents both solve the problem of assisting the diagnosis of Alzheimer's disease through image analysis, but the analysis accuracy is insufficient, resulting in inaccurate diagnosis results. Summary of the Invention

[0003] This application provides a method for analyzing Alzheimer's disease pathological images based on deep learning, which provides important technical support for the early diagnosis and treatment of Alzheimer's disease through the combination of deep learning and image processing. The method includes:

[0004] Obtain multiple medical images of a patient, align the multiple medical images with corresponding reference images, obtain a target region, calculate a weight vector of each target region corresponding to a target medical image, and obtain a key image corresponding to the target region based on the multiple medical images and the weight vector;

[0005] Divide each of the key images into multiple sub-images, and obtain N spiral stacked images corresponding to each key image according to the lesion probability and positional relationship of each sub-image;

[0006] Create an analysis model, and train the analysis model according to historical medical images and a dynamic weight adjustment algorithm;

[0007] Input the N spiral stacked images corresponding to each key image into the analysis model, obtain the medical data corresponding to each key image, and obtain the disease risk of the patient according to the medical data.

[0008] As a preferred technical solution of the present invention, the acquisition of the key images includes:

[0009] Align each medical image with the corresponding reference image, and obtain the target region. Based on the pixel clarity, brightness, contrast of the target region in each target image and the proportion of the target region in the target image, score it, and obtain an evaluation vector composed of the scores of the target region corresponding to each target image. The target region is the image region corresponding to the pathological feature;

[0010] Calculate the weight vector corresponding to the target region based on the evaluation vector, and take the ratio of the score corresponding to each target medical image to the total score as the weight in the corresponding weight vector. The total score is the sum of each score in the evaluation vector;

[0011] Weight the target image based on the weight vector to obtain the key image corresponding to the target region.

[0012] As a preferred technical solution of the present invention, the acquisition of N spiral stacked images corresponding to each key image includes:

[0013] Divide each key image into multiple sub-images, obtain the lesion probability of the corresponding region of each sub-image based on historical diagnosis information, and take the region with the highest lesion probability corresponding to the sub-image as the starting sub-image. Clockwise obtain other adjacent sub-images within the preset range of the starting sub-image. Arrange the starting sub-image and the corresponding adjacent sub-images in clockwise order to form a spiral stacked image, and add the lesion probability corresponding to the starting sub-image and the position information of each sub-image to the spiral stacked image. Repeat this step to obtain N spiral stacked images corresponding to the key image.

[0014] As a preferred technical solution of the present invention, the training of the analysis model includes:

[0015] Obtain historical medical images, perform affine transformation and non-linear transformation on each of the historical medical images to globally and locally align the historical medical images with the reference image, obtain aligned images, enhance the aligned images, and based on the enhanced aligned images and the reference image, locate and cut out each core image. According to the annotation information of the core images, count the first number of normal images and the second number of diseased images among the core images of the same type. By weighting the normal images and / or the diseased images, generate a weighted image such that the difference between the first number and the second number is less than or equal to a set threshold. Repeat the method of obtaining the spiral stack diagram corresponding to the key image, obtain and train the analysis model through N historical spiral stack diagrams corresponding to each core image and the weighted image, and also dynamically adjust the weights of the historical spiral stack diagrams corresponding to each core image or weighted image through a loss function.

[0016] As a preferred technical solution of the present invention, obtaining the disease risk of the patient according to the medical data includes:

[0017] Input the N spiral stack diagrams corresponding to each key image into the analysis model, obtain the medical data corresponding to each key image, the medical data being the risk value corresponding to each key image. Based on the historical case database, determine the weight coefficient corresponding to each key image, where the weight coefficient is calculated by a statistical learning algorithm;

[0018] Perform weighted summation on the medical data and the corresponding weight coefficients to generate a comprehensive risk score;

[0019] According to the comparison result between the comprehensive risk score and a preset threshold, output the disease risk level of the patient, the risk level including low risk, medium risk, and high risk.

[0020] As a preferred technical solution of the present invention, the pathological features at least include hippocampal atrophy, ventricular enlargement, medial temporal lobe, frontal lobe atrophy, and entorhinal cortex atrophy.

[0021] As a preferred technical solution of the present invention, the analysis model is a neural network model.

[0022] The present invention also provides a deep learning-based Alzheimer's disease pathological image analysis system for implementing the above method. The system includes:

[0023] An acquisition unit for acquiring multiple medical images of a patient, aligning the multiple medical images with corresponding reference images, and obtaining a target region, calculating a weight vector of the target medical image corresponding to each target region, and obtaining a key image corresponding to the target region based on the multiple medical images and the weight vector;

[0024] A dividing unit, configured to divide each of the key images into a plurality of sub-images, and obtain N spiral stacked images corresponding to each key image according to the lesion probability and positional relationship of each sub-image;

[0025] A training unit, configured to create an analysis model and train the analysis model according to historical medical images and a dynamic weight adjustment algorithm;

[0026] A prediction unit, configured to input the N spiral stacked images corresponding to each key image into the analysis model, obtain medical data corresponding to each key image, and obtain the disease risk of the patient according to the medical data.

[0027] The present invention also provides a computer-readable storage medium, on which instructions are stored, and when the instructions are executed by a processor, the above method is implemented.

[0028] Effect

[0029] By performing affine transformation and non-linear transformation registration on multiple brain medical images of a patient, the present invention eliminates equipment differences and morphological interferences, accurately extracts target regions related to pathological features, such as the hippocampus, ventricles, etc., and generates high-quality key images by weighted fusion of multiple images through a weight vector, effectively solving the problem of uneven image quality in traditional methods; divides the key images into multiple sub-images, labels the lesion probability based on historical confirmed data, and constructs spiral stacked images centered on high-probability regions, retaining local spatial relationships and enhancing the model's perception ability of lesion distribution features; in the model training stage, balances normal and diseased sample data through a dynamic weight adjustment algorithm, combines data augmentation techniques to generate lesion process images, improves the model's generalization ability, and at the same time dynamically optimizes the training weights using a loss function to strengthen the learning of complex pathological features; finally, inputs the spiral stacked images into a neural network model to predict local risk values, generates a comprehensive risk score by combining the statistical weight coefficients of the historical case database, and outputs low, medium, and high risk levels through threshold division to achieve multi-dimensional evaluation. This method overcomes the defect of insufficient image analysis accuracy in the prior art through image fusion optimization, spiral stacked structure modeling, dynamic training strategies, and comprehensive scoring mechanisms, significantly improves the accuracy and reliability of early diagnosis of Alzheimer's disease, provides efficient and interpretable auxiliary decision-making support for clinical practice, and at the same time has strong adaptability and scalability, and can be extended to the field of pathological image analysis of other neurodegenerative diseases. Description of the Drawings

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0031] Figure 1 It is a flowchart of a method for analyzing Alzheimer's disease pathological images based on deep learning in an embodiment of this application;

[0032] Figure 2 It is a flowchart of a method for obtaining N helically stacked images corresponding to each key image in an embodiment of this application;

[0033] Figure 3 It is a flowchart of a training method for an analysis model in an embodiment of this application;

[0034] Figure 4 It is a structural diagram of a system for analyzing Alzheimer's disease pathological images based on deep learning in an embodiment of this application. Specific embodiments

[0035] The embodiments of this application provide a method and system for analyzing Alzheimer's disease pathological images based on deep learning. The terms "first", "second", "third", "fourth", etc. (if any) in the description, claims, and accompanying drawings of this application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0036] For ease of understanding, the following describes the specific process of the embodiments of this application. As Figure 1 shown, an embodiment of a method for analyzing Alzheimer's disease pathological images based on deep learning in this application includes:

[0037] Obtain multiple medical images of a patient, align the multiple medical images with corresponding reference images, obtain target regions, calculate the weight vector of each target region corresponding to the target medical image, and obtain the key images corresponding to the target regions based on the multiple medical images and the weight vector;

[0038] Specifically, the above-mentioned medical images are magnetic resonance images of the patient's brain. Multiple brain medical images of the patient are registered with the reference image to eliminate morphological differences caused by shooting angles, resolutions, or equipment differences, and to locate pathological feature regions related to Alzheimer's disease, such as the hippocampus, ventricles, etc. The target regions are extracted, and based on indicators such as the clarity, brightness, contrast, and area ratio of the target regions, the target regions of each image are scored to form an evaluation vector. The weight vector is calculated through scoring, which is the proportion of the score of a single image in the total score. The target regions of multiple images are weighted and fused to generate high-quality key images. Through weighted fusion, the clearest and highest-contrast target region images are selected to reduce noise interference. Alignment and registration ensure that images from different sources have a consistent anatomical structure reference, providing a reliable input for subsequent analysis.

[0039] Each of the key images is divided into multiple sub-images, and N helical stacked images corresponding to each key image are obtained according to the lesion probability and positional relationship of each sub-image.

[0040] Specifically, the key image is divided into multiple sub-images, and the lesion probability of each sub-image is labeled based on historical diagnosis data. Starting from the sub-image with the highest lesion probability, adjacent sub-images are selected clockwise within its preset range and stacked in a three-dimensional helical structure. This process is repeated to generate N helical stacked images, ensuring that sub-images with high lesion probabilities in the key region and their surrounding context information are covered. The above technical solution, through lesion probability labeling, gives priority to sub-images with obvious pathological features, improving the analysis efficiency. The helical stacked images retain the spatial relationship of the local anatomical structure through the sequential arrangement of adjacent sub-images, enhancing the model's perception ability of the distribution of the lesion region.

[0041] An analysis model is created, and the analysis model is trained according to historical medical images and a dynamic weight adjustment algorithm.

[0042] Specifically, the historical medical images are subjected to affine / nonlinear alignment and enhancement, such as brightness adjustment and noise addition, to generate diverse training data. Through weighting, an intermediate state between normal images and diseased images, that is, the lesion process images, is generated to balance the number of positive and negative samples and solve the problem of data imbalance. During the training process, the weights of different helical stacked images are dynamically adjusted according to the loss function. For example, the weight of samples with a large prediction deviation is increased to optimize the model's attention to key features. A neural network model, such as CNN or Transformer, is used. The helical stacked images are input, and the risk value corresponding to the key image is output. The above technical solution, through data augmentation and sample balancing, enables the model to better adapt to differences between different devices and patients. The dynamic weight adjustment enables the model to focus on samples with larger prediction errors, improving the ability to identify complex pathological features.

[0043] Input the N helical stack images corresponding to each of the key images into the analysis model, obtain the medical data corresponding to each of the key images, and obtain the disease risk of the patient according to the medical data.

[0044] Specifically, input the N helical stack images of the key images into the trained model to output the local risk value of each key image. Based on the historical case database, calculate the weight coefficients of different key images through a statistical learning algorithm (such as logistic regression), perform a weighted sum on the local risk values to generate a comprehensive risk score, compare the comprehensive score with a preset threshold, such as low risk: comprehensive risk score < 0.3, medium risk: comprehensive risk score between 0.3 and 0.7, high risk: comprehensive risk score > 0.7, and output the final disease risk level. The above technical solution avoids misjudgment in a single region by weighted fusion of the risk values in different key regions, improves the reliability of the overall assessment, and the risk level division intuitively guides clinical decisions to help doctors quickly judge the disease progression of the patient.

[0045] Further, the acquisition of the key images includes:

[0046] Align each of the medical images with the corresponding reference image, and obtain the target region. Based on the pixel clarity, brightness, contrast of the target region in each target image, and the proportion of the target region in the target image, perform a scoring to obtain an evaluation vector composed of the scores of the target region corresponding to each target image. The target region is the image region corresponding to the pathological feature;

[0047] Calculate the weight vector corresponding to the target region based on the evaluation vector, and use the ratio of the score corresponding to each target medical image to the total score as the weight in the corresponding weight vector. The total score is the sum of each score in the evaluation vector;

[0048] Weight the target image based on the weight vector to obtain the key image corresponding to the target region.

[0049] Specifically, when a patient seeks medical treatment, they often carry examination information obtained from other hospitals or institutions, including the above-mentioned medical images, i.e., brain MRI images. To accurately assess the patient's disease risk, the medical images obtained from this hospital and other medical institutions closest to the current time are integrated. Since the shooting angles or image resolutions of different medical images are different, each medical image is aligned with the corresponding reference image, and the target region is obtained. The method for obtaining the target region is the same as that for obtaining the core image in the following text. The target image is scored based on the pixel clarity, brightness, contrast, and the proportion of the target region in the target image at the position corresponding to the target region in the target image that all include the same target region. The higher the clarity, brightness, contrast, and the larger the proportion, the higher the score of the corresponding target image. The score of the target image is also used as the score vector of the target region. The ratio of each score in the score vector to the total score is used as the weight corresponding to the target image, and the weights of each target image form the weight vector. The weighted image can be obtained from the target image based on the weighted vector through a machine learning model or an image weighting algorithm, and the weighted image is used as the key image. The target region, i.e., the pathological feature reflection area of Alzheimer's disease, at least includes: the hippocampal region, the ventricular region, the medial temporal lobe region, the frontal lobe region, and the entorhinal cortex region. The above technical solution realizes the optimal fusion of different medical images, improves the image quality, and lays a foundation for accurately predicting the patient's disease risk in the future.

[0050] Furthermore, as Figure 2 shown, the acquisition of N helical stacked images corresponding to each key image includes:

[0051] Each key image is divided into multiple sub-images. The lesion probability of the corresponding region of each sub-image is obtained based on historical diagnosis information. The sub-image corresponding to the region with the highest lesion probability is used as the starting sub-image. Other adjacent sub-images are obtained clockwise within the preset range of the starting sub-image. The starting sub-image and the corresponding adjacent sub-images in clockwise order are sorted to form a helical stacked image, and the lesion probability corresponding to the starting sub-image and the position information of each sub-image are added to the helical stacked image. This step is repeated to obtain N helical stacked images corresponding to the key image.

[0052] Specifically, the above-mentioned key image is the image area corresponding to the pathological features of Alzheimer's disease, that is, the reflection area. In order to efficiently obtain the corresponding pathological information based on the above-mentioned key image, the image of the above-mentioned key area is divided into multiple sub-images, and position information is labeled for each of the above-mentioned sub-images, that is, the position in the above-mentioned key image. By statistically analyzing the lesion probability at the position of each sub-image in the above-mentioned key image in the historical confirmed information, and taking the sub-image corresponding to the position with the highest lesion probability as the above-mentioned starting sub-image, and sorting the other sub-images adjacent in sequence clockwise within the preset range of the above-mentioned starting sub-image, the above-mentioned starting sub-image and the other sub-images adjacent in sequence are also formed into a three-dimensional above-mentioned spiral stacked image according to the above-mentioned sorting. Using the same method, that is, selecting the sub-image corresponding to the area with the highest lesion probability from the remaining sub-images as the above-mentioned starting sub-image again, and repeating the above steps to obtain all N above-mentioned spiral stacked images in the above-mentioned key image. Through the above technical solution, it is possible to obtain the key attention area reflecting the above-mentioned key image, that is, the area with a higher disease probability, and sort the sub-images with a higher disease probability as the center and the other sub-images adjacent in sequence clockwise to obtain the corresponding spiral stacked image, which can not only improve the accuracy of subsequent analysis but also improve the analysis speed.

[0053] Further, as Figure 3 shown, the training of the analysis model includes:

[0054] Obtain historical medical images, perform affine transformation and non-linear transformation on each of the historical medical images to globally and locally align the historical medical images with the reference image to obtain aligned images, enhance the aligned images, and based on the enhanced aligned images and the reference image, locate and cut out each core image. According to the annotation information of the core image, count and based on the first quantity of normal images and the second quantity of diseased images in the same type of core images, generate a weighted image by weighting the normal images and / or the diseased images, so that the difference between the first quantity and the second quantity is less than or equal to the set threshold. Repeat the method of obtaining the spiral stacked image corresponding to the key image, obtain and train the analysis model with N historical spiral stacked images corresponding to each core image and the weighted image, and also dynamically adjust the weights of the historical spiral stacked images corresponding to each core image or weighted image through the loss function.

[0055] Specifically, since the image quality and angles of the above-mentioned historical medical images, i.e., historical brain MRI images, captured by different devices are different, and the brain shapes and sizes of different patients are also different, in order to improve the generalization ability of the above-mentioned analysis model, the above-mentioned reference image is used as the reference image, and each of the above-mentioned medical images is globally aligned with the above-mentioned reference image through affine transformation, i.e., linear registration, and is also locally aligned with the above-mentioned reference image through the above-mentioned non-linear transformation, i.e., non-linear registration, so as to eliminate the morphological differences between the images. The quality of the aligned images is also improved by enhancing the aligned images. Since the above-mentioned aligned images are aligned with the object of the above-mentioned reference image, the core image cutting positions in the above-mentioned aligned images and the above-mentioned reference image are the same. Thus, the above-mentioned core images can be automatically aligned, located, and cut out. The first quantity and the second quantity of normal images and diseased images in the same type of core images are also statistically counted according to the annotation information in the above-mentioned core images. The above-mentioned annotation information includes whether the above-mentioned core image is a normal image or a diseased image and the corresponding risk value. Among them, the risk value corresponding to a normal image is 0, and the risk value corresponding to a diseased image is 1. The above-mentioned first quantity and the second quantity can reflect whether the training sample data is balanced. When the difference between the above-mentioned first quantity and the above-mentioned second quantity is greater than the above-mentioned set threshold, it indicates that the above-mentioned training sample data is unbalanced. By weighting between the above-mentioned normal images or between the above-mentioned diseased images, normal images or diseased images are generated, so as to eliminate the problem of unbalanced sample data. Since the above-mentioned training sample data only includes normal images and diseased images, but does not include the lesion process images between the above-mentioned normal images and the above-mentioned diseased images, and at the same time, in order to improve the generalization ability of the above-mentioned analysis model, the above-mentioned normal images and the above-mentioned diseased images are also weighted to generate multiple lesion process images, and the label values of the above-mentioned lesion process images are generated according to the weights of the normal images and the above-mentioned diseased images. For example: the weight of the normal image is 20%, the weight of the diseased image is 80%, the label value of the normal image is 0, and the label value of the diseased image is 1. The above-mentioned label value is the disease risk value, then the weight of the above-mentioned diseased image is used as the label value of the above-mentioned lesion process image, i.e., 0.8. The above-mentioned normal images, diseased images, and diseased process images generated by weighting are all used as the above-mentioned weighted images. By the method of obtaining the spiral stacked diagrams corresponding to the key images, the spiral stacked diagrams corresponding to the above-mentioned core images and the above-mentioned weighted images are respectively obtained, and the above-mentioned analysis model is trained based on the above-mentioned spiral stacked diagrams. Since the above-mentioned spiral stacked diagrams also include position information and lesion probability data, the above-mentioned core images and weighted images can also be restored through the above-mentioned spiral stacked diagrams, so that the above-mentioned analysis model can be trained from local to global, and focused training can also be carried out according to the above-mentioned diseased probability, thereby improving the accuracy of the analysis model. During the training process, the above-mentioned loss function is also used to dynamically adjust the weight of each core image or weighted image according to the deviation of the prediction result. For example, the label data of the input image is 0.3, that is, the risk value, and the output data of the above-mentioned input image input into the above-mentioned analysis model is 0.4, then the weight of the historical spiral stacked diagram corresponding to the weighted image with the label data of 0.3 is increased, thereby improving the prediction accuracy of the analysis model. Through the above technical solution, an analysis model with higher accuracy can be obtained, laying a foundation for further obtaining accurate analysis results.

[0056] Further, obtaining the disease risk of the patient according to the medical data includes:

[0057] Inputting the N spiral stacked diagrams corresponding to each of the key images into the analysis model to obtain the medical data corresponding to each key image, where the medical data is the risk value corresponding to each key image, and based on the historical case database, determining the weight coefficient corresponding to each key image, where the weight coefficient is calculated by a statistical learning algorithm;

[0058] Performing weighted summation on the medical data and the corresponding weight coefficients to generate a comprehensive risk score;

[0059] According to the comparison result between the comprehensive risk score and a preset threshold, outputting the disease risk level of the patient, where the risk level includes low risk, medium risk, and high risk.

[0060] Specifically, by inputting the N spiral stacked images corresponding to each of the above key images into the above analysis model, the risk value corresponding to each key image can be obtained, and the weight coefficient corresponding to each of the above key images is also determined according to the above historical case database, and weighted summation is performed based on the weight coefficient and risk value of each key image, that is, the medical data, to obtain the above comprehensive risk score, and based on the comparison between the above comprehensive score and the preset threshold corresponding to each risk level, the disease level of the above patient is obtained. Through the above technical solution, the disease risk of the above patient can be accurately obtained, thereby guiding medical staff and patients to carry out corresponding treatments.

[0061] Furthermore, the pathological features at least include hippocampal atrophy, ventricular enlargement, medial temporal lobe, frontal lobe atrophy, and entorhinal cortex atrophy.

[0062] Furthermore, the analysis model is a neural network model.

[0063] The present invention also provides an Alzheimer's disease pathological image analysis system based on deep learning for implementing the above method. As Figure 4 shown, the system includes:

[0064] An acquisition unit, configured to acquire multiple medical images of a patient, align the multiple medical images with corresponding reference images, obtain a target region, calculate a weight vector of each target region corresponding to a target medical image, and obtain a key image corresponding to the target region based on the multiple medical images and the weight vector;

[0065] A partitioning unit, configured to partition each key image into multiple sub-images, and obtain N helical stacked images corresponding to each key image according to the lesion probability and positional relationship of each sub-image;

[0066] A training unit, configured to create an analysis model and train the analysis model according to historical medical images and a dynamic weight adjustment algorithm;

[0067] A prediction unit, configured to input the N helical stacked images corresponding to each key image into the analysis model, obtain medical data corresponding to each key image, and obtain the disease risk of the patient according to the medical data.

[0068] The present invention also provides a computer-readable storage medium, on which instructions are stored. The instructions, when executed by a processor, implement the above method.

[0069] In summary, the present invention performs affine transformation and non-linear transformation registration on multiple brain medical images of patients, eliminates equipment differences and morphological interferences, accurately extracts target regions related to pathological features, such as the hippocampus, ventricles, etc., and generates high-quality key images by weighted fusion of multiple images using a weight vector, effectively solving the problem of uneven image quality in traditional methods; divides the key images into multiple sub-images, annotates the lesion probabilities based on historical confirmed data, and constructs a spiral stacked graph centered on the high-probability regions, retaining local spatial relationships and enhancing the model's perception ability of lesion distribution characteristics; in the model training stage, balances normal and diseased sample data through a dynamic weight adjustment algorithm, generates lesion process images in combination with data augmentation techniques to improve the model's generalization ability, and at the same time dynamically optimizes the training weights using a loss function to strengthen the learning of complex pathological features; finally, inputs the spiral stacked graph into a neural network model to predict local risk values, generates a comprehensive risk score in combination with the statistical weight coefficients of the historical case database, and outputs low, medium, and high risk levels through threshold division to achieve multi-dimensional evaluation. This method overcomes the defect of insufficient image analysis accuracy in the prior art through image fusion optimization, spiral stacked structure modeling, dynamic training strategies, and comprehensive scoring mechanisms, significantly improves the accuracy and reliability of early diagnosis of Alzheimer's disease, provides efficient and interpretable auxiliary decision-making support for clinical practice, and at the same time has strong adaptability and scalability, and can be extended to the field of pathological image analysis of other neurodegenerative diseases.

[0070] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0071] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0072] As described above, the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A method for analyzing Alzheimer's disease pathological images based on deep learning, characterized in that, The method includes: Obtaining multiple medical images of a patient, aligning the multiple medical images with corresponding reference images, obtaining target regions, calculating a weight vector of each target region corresponding to a target medical image, and obtaining a key image corresponding to the target region based on the multiple medical images and the weight vector; Dividing each key image into multiple sub-images, and obtaining N spiral stacked images corresponding to each key image according to the lesion probability and positional relationship of each sub-image; Creating an analysis model and training the analysis model according to historical medical images and a dynamic weight adjustment algorithm; Inputting the N spiral stacked images corresponding to each key image into the analysis model, obtaining medical data corresponding to each key image, and obtaining the disease risk of the patient according to the medical data.

2. The method according to claim 1, wherein The obtaining of the key image includes: Aligning each medical image with a corresponding reference image, obtaining a target region, scoring based on the pixel clarity, brightness, contrast of the target region in each corresponding target image and the proportion of the target region in the target image, obtaining an evaluation vector composed of the scores of the target region corresponding to each target image, and the target region being an image region corresponding to pathological features; Calculating a weight vector corresponding to the target region based on the evaluation vector, taking the ratio of the score of each target medical image to the total score as the weight in the corresponding weight vector, and the total score being the sum of each score in the evaluation vector; Weighting the target image based on the weight vector to obtain the key image corresponding to the target region.

3. The method according to claim 1, wherein The obtaining of N spiral stacked images corresponding to each key image includes: Dividing each key image into multiple sub-images, obtaining the lesion probability of the region corresponding to each sub-image based on historical diagnosis information, taking the sub-image corresponding to the region with the highest lesion probability as the starting sub-image, obtaining other sequentially adjacent sub-images clockwise within a preset range of the starting sub-image, forming a spiral stacked image by sorting the starting sub-image and the sequentially adjacent sub-images clockwise, and adding the lesion probability corresponding to the starting sub-image and the position information of each sub-image to the spiral stacked image, repeating this step to obtain N spiral stacked images corresponding to the key image.

4. The method according to claim 1, wherein The training of the analysis model includes: Obtain historical medical images, perform affine transformation and non-linear transformation on each of the historical medical images to globally and locally align the historical medical images with the reference image, obtain aligned images, enhance the aligned images, and based on the enhanced aligned images and the reference image, locate and cut out each core image. According to the annotation information of the core images, count the first number of normal images and the second number of diseased images among the core images of the same type. By weighting the normal images and / or the diseased images, generate a weighted image such that the difference between the first number and the second number is less than or equal to a set threshold. Repeat the method of obtaining the spiral stack diagram corresponding to the key image, obtain and train the analysis model through N historical spiral stack diagrams corresponding to each core image and the weighted image, and also dynamically adjust the weights of the historical spiral stack diagrams corresponding to each core image or weighted image through a loss function.

5. The method according to claim 1, wherein Obtain the patient's disease risk according to the medical data, including: Input the N spiral stack diagrams corresponding to each key image into the analysis model, obtain the medical data corresponding to each key image, the medical data being the risk value corresponding to each key image, and based on the historical case database, determine the weight coefficient corresponding to each key image, where the weight coefficient is calculated by a statistical learning algorithm; Perform weighted summation on the medical data and the corresponding weight coefficients to generate a comprehensive risk score; According to the comparison result between the comprehensive risk score and a preset threshold, output the disease risk level of the patient, the risk level including low risk, medium risk, and high risk.

6. The method according to claim 2, wherein The pathological features at least include hippocampal atrophy, ventricular enlargement, medial temporal lobe, frontal lobe atrophy, and entorhinal cortex atrophy.

7. The method according to claim 1, wherein The analysis model is a neural network model.

8. A pathological image analysis system for Alzheimer's disease based on deep learning, which is used to implement the method described in any one of claims 1-7, characterized in that, The system includes: An acquisition unit for acquiring multiple medical images of a patient, aligning the multiple medical images with a corresponding reference image, obtaining a target area, calculating a weight vector of the target medical image corresponding to each target area, and obtaining a key image corresponding to the target area based on the multiple medical images and the weight vector; A division unit for dividing each key image into multiple sub-images and obtaining N spiral stack diagrams corresponding to each key image according to the lesion probability and positional relationship of each sub-image; A training unit for creating an analysis model and training the analysis model according to historical medical images and a dynamic weight adjustment algorithm; A prediction unit for inputting the N spiral stack diagrams corresponding to each key image into the analysis model, obtaining the medical data corresponding to each key image, and obtaining the disease risk of the patient according to the medical data.

9. A computer-readable storage medium having instructions stored thereon, characterized in that, When the instruction is executed by a processor, the method according to any one of claims 1-7 is implemented.

Images