Visual auxiliary diagnosis method and system for pathological PET (positron emission tomography) of Alzheimer disease

Abstract

The invention provides a visual auxiliary diagnosis method and system for Alzheimer's disease pathology PET, and relates to the technical field of medical image processing.The method comprises the steps that a target PET image of a subject is preprocessed, and a preprocessed image is obtained; the preprocessed image is put into a trained antagonism decomposition learning model so as to decompose the preprocessed image, and a pathology probability graph and a diagnosis probability are generated based on a decomposition result; calculating an Alzheimer's disease resistance decomposition score based on a decomposition result of the resistance decomposition learning model, and performing visual evaluation in combination with the diagnosis probability to judge whether the subject is an Alzheimer's disease patient; and displaying the pathology probability graph in a thermodynamic diagram form so as to visually display the potential lesion area. Pathological and physiological uptake can be accurately decomposed, the accuracy and interpretability of AD diagnosis are improved, and meanwhile efficient human-machine collaborative pathological assessment is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to a visual assisted diagnosis method and system for Alzheimer's disease pathological PET. Background Art

[0002] Alzheimer's disease (AD) is a progressive neurodegenerative disease, and its pathological features include amyloid-β (Aβ) plaques and Tau neurofibrillary tangles. Positron Emission Computed Tomography (PET) technology is an important tool for the early diagnosis of AD. Researchers have developed a large number of methods to objectively and semi-quantitatively analyze AD pathological PET, but the existing analysis methods still have the following limitations:

[0003] 1. Influence of non-specific uptake in the brain: PET images not only capture pathological Aβ and Tau depositions, but may also be affected by normal physiological uptake, such as blood perfusion, metabolic activity, and other non-pathological factors. This makes image interpretation challenging, especially for potentially weakly positive patients, whose physiological uptake may interfere with the radiologist's interpretation of pathological uptake.

[0004] 2. Dependence on spatial normalization: Traditional semi-quantitative methods (such as the Centiloid for AβPET and the CenTauRz index for tau PET) usually rely on complex spatial normalization techniques, which involve processing transformation patterns such as rotation, translation, and flipping of images. Usually, manual coarse registration is required, and the calculation process is complex and time-consuming.

[0005] 3. Ignoring the heterogeneity of AD pathological marker deposition patterns: Traditional semi-quantitative methods often assume that the voxel weights in the region of interest are equal, ignoring the potential heterogeneity between individuals and within images, that is, the impact of pathological marker deposition in different individuals and different brain regions on cognitive decline is different. Therefore, it is impossible to accurately evaluate specific patients and cannot adaptively process images with highly variable structures and pathological deposition patterns.

[0006] 4. Lack of interpretability: Deep learning methods can distinguish healthy controls (Cognitive Normal, CN) from AD patients with high accuracy, but the decision-making process of these models is difficult to intuitively understand, resulting in doctors' difficulty in trusting their diagnostic results in clinical practice. Existing methods mainly rely on global or local feature extraction for classification, but cannot clearly indicate which specific image regions or features play a decisive role in the final diagnostic result.

[0007] Therefore, there is an urgent need for a new type of imaging analysis system that can effectively decompose pathological and physiological uptake, improve the diagnostic accuracy of AD, and achieve human-machine collaborative pathological evaluation. Summary of the Invention

[0008] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a visual-aided diagnosis method and system for Alzheimer's disease pathological PET, which can accurately decompose pathological and physiological uptake, improve the accuracy and interpretability of AD diagnosis, and at the same time achieve efficient human-machine collaborative pathological evaluation.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A visual-aided diagnosis method for Alzheimer's disease pathological PET, comprising:

[0011] Preprocess the target PET image of the subject to obtain a preprocessed image;

[0012] Input the preprocessed image into a trained adversarial decomposition learning model to decompose the preprocessed image, and generate a pathological probability map and a diagnostic probability based on the decomposition result;

[0013] Calculate the Alzheimer's disease adversarial decomposition score based on the decomposition result of the adversarial decomposition learning model, and perform visual evaluation in combination with the diagnostic probability to determine whether the subject is a patient with Alzheimer's disease;

[0014] Display the pathological probability map in the form of a heat map to visually display potential lesion areas.

[0015] Preferably, preprocessing the target PET image of the subject to obtain a preprocessed image includes:

[0016] Perform rigid alignment on the target PET image to obtain an aligned image;

[0017] Identify the cerebellar gray matter region in the aligned image;

[0018] Extract the standard uptake value of the cerebellar gray matter region and perform intensity normalization to obtain the normalized preprocessed image.

[0019] Preferably, the adversarial decomposition learning model includes a decoupler and a discriminator; the decoupler is used to decompose the preprocessed image into a linear superposition of a physiological uptake part and a pathological uptake part, obtain a decomposition result, and determine the pathological probability map according to the decomposition result; the discriminator is used to distinguish whether the input image has been modified by the decoupler and stripped of the components of Alzheimer's disease, evaluate the proportion of the decomposed part of the image in the original image, and diagnose the input image as in a normal state or with Alzheimer's disease; the input image includes an image of the physiological uptake part and the original preprocessed image.

[0020] Preferably, the decoupler uses 6 convolutional layers with residual connections, and the number of channels of each convolutional layer is 16, 32, 32, 32, 32, and 32 respectively, and the stride is 2 for all; the activation function of the decoupler uses a parametric rectified linear unit.

[0021] Preferably, the discriminator adopts a multi-layer convolutional network; the discriminator includes a real / fake discrimination classification head and a diagnosis classification head.

[0022] Preferably, the total loss function of the adversarial decomposition learning model is:

[0023]

[0024] Among them, is the total loss function, is the loss function of the decoupler, E V～data denotes taking the expectation, D TrueFake (V') represents the result of the real / fake discrimination classification head evaluating the image processed by the decoupler, D diagnosis (V') represents the result of the diagnosis classification head evaluating the image processed by the decoupler; is the L1 regularization loss, P(x, y, z) is the three-dimensional representation of the probability field P, and x, y, and z respectively represent the x-axis coordinate value, y-axis coordinate value, and z-axis coordinate value of the coordinates of the voxels in the image; is the loss function of the real / fake discrimination classification head, Among them, D TrueFake (V) represents the result of the real / fake discrimination classification head evaluating the preprocessed image; H(·, ·) is the cross-entropy loss function, and V represents the input preprocessed image; is the loss function of the diagnosis classification head, y is the preset image label, D diagnosis (V) represents the result of the diagnosis classification head evaluating the preprocessed image; λ is the preset constraint coefficient.

[0025] Preferably, the calculation formula for the Alzheimer's disease against decomposition score is as follows:

[0026]

[0027] where ADAD score is the Alzheimer's disease against decomposition score, V(x, y, z) is the three-dimensional representation of the preprocessed image V, and v is the physical volume of the voxel.

[0028] Preferably, the value of the constraint coefficient is 5.

[0029] Preferably, the physical volume of the voxel is 0.15 cm 3 .

[0030] A visual assistant diagnosis system for Alzheimer's disease pathological PET includes:

[0031] An image preprocessing unit, configured to preprocess the target PET image of the subject to obtain a preprocessed image;

[0032] An adversarial decomposition unit, configured to input the preprocessed image into a trained adversarial decomposition learning model to decompose the preprocessed image, and generate a pathological probability map and a diagnosis probability based on the decomposition result;

[0033] A visual evaluation unit, configured to calculate the Alzheimer's disease against decomposition score based on the decomposition result of the adversarial decomposition learning model, and perform visual evaluation in combination with the diagnosis probability to determine whether the subject is a patient with Alzheimer's disease;

[0034] A visualization display unit, configured to display the pathological probability map in the form of a heat map to visually display potential lesion areas.

[0035] According to the specific embodiments provided by the present invention, the following technical effects are disclosed:

[0036] The present invention provides a visual assistant diagnosis method and system for Alzheimer's disease pathological PET. The method includes: preprocessing the target PET image of the subject to obtain a preprocessed image; inputting the preprocessed image into a trained adversarial decomposition learning model to decompose the preprocessed image, and generating a pathological probability map and a diagnosis probability based on the decomposition result; calculating the Alzheimer's disease against decomposition score based on the decomposition result of the adversarial decomposition learning model, and performing visual evaluation in combination with the diagnosis probability to determine whether the subject is a patient with Alzheimer's disease; displaying the pathological probability map in the form of a heat map to visually display potential lesion areas. It can accurately decompose pathological and physiological uptake, improve the accuracy and interpretability of AD diagnosis, and at the same time achieve efficient human-computer collaborative pathological evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0038] Figure 1 It is a flowchart of the method provided by the embodiment of the present invention;

[0039] Figure 2 It is a schematic diagram of the algorithm framework provided by the embodiment of the present invention;

[0040] Figure 3 It is a schematic diagram of the output result of the typical PiB AβPET image provided by the embodiment of the present invention;

[0041] Figure 4 It is a schematic diagram of the output result of the tau PET image provided by the embodiment of the present invention;

[0042] Figure 5 It is a schematic diagram of the comparison between the output result of the PET image of cognitively normal patients provided by the embodiment of the present invention and the interpretable method of the Guided Grad-CAM classification network;

[0043] Figure 6 It is a schematic diagram of the optimal cut-off point for the diagnosis of the Alzheimer's Disease Adversarial Decomposition (ADAD) score in two modalities of Aβ and tau provided by the embodiment of the present invention;

[0044] Figure 7 It is a schematic diagram of the ADAD score consistency assessment provided by the embodiment of the present invention;

[0045] Figure 8 It is a schematic diagram of the neural network structure used in the ADL model provided by the embodiment of the present invention;

[0046] Figure 9 It is a schematic diagram of the system structure provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0048] The object of the present invention is to provide a visual assisted diagnosis method and system for Alzheimer's disease pathological PET, which can accurately decompose pathological and physiological uptake, improve the accuracy and interpretability of AD diagnosis, and at the same time achieve efficient human-machine collaborative pathological evaluation.

[0049] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] Figure 1 As shown in the method flow chart provided by the embodiment of the present invention, Figure 1 As shown, the present invention provides a visual assisted diagnosis method for Alzheimer's disease pathological PET, including:

[0051] Step 100: Preprocess the target PET image of the subject to obtain a preprocessed image;

[0052] Step 200: Input the preprocessed image into a trained adversarial decomposition learning model to decompose the preprocessed image, and generate a pathological probability map and a diagnostic probability based on the decomposition result;

[0053] Step 300: Calculate an Alzheimer's disease adversarial decomposition score based on the decomposition result of the adversarial decomposition learning model, and perform a visual assessment in combination with the diagnostic probability to determine whether the subject is a patient with Alzheimer's disease;

[0054] Step 400: Display the pathological probability map in the form of a heat map to visually display potential lesion areas.

[0055] Specifically, as Figure 2 shown, the architecture of the adversarial decomposition learning model in this embodiment is as follows:

[0056] (1) Decoupler (G): Adopt a variant based on the U-Net structure, input the PET image, and output a pathological probability map. The decoupler is responsible for predicting the probability map P and modifying the original image V to generate a stripped image accordingly. This modified image is designed to resemble a "cognitively normal" brain that does not have any AD-related pathology.

[0057] Exemplarily, in this embodiment, a 6-layer convolutional layer with residual connections is used, and the number of channels in each layer is 16, 32, 32, 32, 32, 32 respectively, and the stride is 2 for all layers. The parametric rectified linear unit (PReLU) is used as the activation function. In addition, the decoder of the U-Net structure uses 2x upsampling combined with a 1x1 convolutional layer as the main structure, rather than the traditional transposed convolutional layer, to reduce "checkerboard artifacts", so as to achieve the balance between the fastest calculation speed and the optimal decomposition effect. The activation function of the last layer of the decoupler selects the Sigmoid function to convert the neural network output into a probability field of 0 to 1.

[0058] (2) Discriminator (D): A multi-layer convolutional network (CNN) is adopted, which contains two classification heads D TrueFake , D diagnosis (the real / fake discrimination head and the diagnosis head) are respectively used to distinguish whether the input image has been modified by the decoupler and the AD component has been stripped (if the image is decomposed, evaluate the proportion of the decomposed part in the original image), and diagnose whether the input image is CN or AD.

[0059] As Figure 8 shown, Figure 8 In the upper row of [Figure], the basic structure of the decoupler is shown. The size of the input PET image after preprocessing is 160x160x96. Subsequently, it is processed by a 6-layer convolutional neural network with residual connections (encoder), and then upsampled to the original resolution in a mirror image manner (the upsampling layer does not use the transposed convolution but uses the upsampling layer + 1x1 convolution to avoid checkerboard artifacts). Finally, a probability map of 0 to 1 is generated through the Sigmoid function; Figure 7 In the lower row of [Figure], the basic structure of the discriminator is shown. The size of the input PET image after preprocessing is 160x160x96. Subsequently, it is flattened into a one-dimensional vector after passing through 4 layers of convolutional layers with residual connections, and finally converted into 2 probability values through a linear layer and the Sigmoid function, respectively predicting whether the input image is AD and to what extent the input image has been tampered with by the decoupler.

[0060] Exemplarily, this embodiment uses four layers of convolutional layers with residual connections as the overall framework, with the number of channels being 16, 32, 64, 128, and the stride being 2 for all layers. Finally, 2 prediction heads are generated through the Sigmoid function to achieve the balance between the fastest calculation speed and the optimal diagnostic and discrimination effects.

[0061] (3) Loss function. In this embodiment, the brain pathological PET input image V of AD patients is modeled as follows:

[0062] 1) For any input image V, it can be approximately decomposed into physiological uptake Vp Partial and pathological uptake V AD Linear superposition of parts V≈V p +V AD , since the signals of all voxels in the image are greater than 0, the physiological uptake is also greater than 0, and the AD disease accumulates pathological markers on the basis of physiological uptake. Therefore, all three parts are strictly non - negative.

[0063] 2) In addition, the input image V also has the following identity transformation: V=(1 - P)V+PV, where P = P(x, y, z) is a probability field that describes the component belonging to AD in the voxel (x, y, z) at a certain position in the input image V. Combining with the linear decomposition estimation, the image decomposition task can be modeled as finding a suitable probability field such that PV exactly corresponds to V AD , V'=(1 - P)V exactly corresponds to V p , that is, the ideal healthy control image after stripping the AD component.

[0064] Exemplarily, both the image V and the probability field P are three - dimensional, x, y, z are coordinates, indicating taking the value at that point. PV is a shorthand method representing point - by - point multiplication.

[0065] 3) For the decoupler, it needs to deceive the discriminator into believing that the image V' stripped of the AD component by the decoupler is a real image and comes from a healthy control, that is

[0066]

[0067] where E V～data denotes taking the mathematical expectation of V from the distribution data of real PET images (the same below).

[0068] 4) To prevent the generator from making meaningless or excessive modifications to the input image, we introduce the L1 regularization loss.

[0069]

[0070] 5) For the D TrueFake head of the discriminator, in this embodiment, it is expected that when facing real data, it can accurately identify that it has not been modified, and when facing the image V' stripped of the AD component, it can output the proportion of the image that has been stripped. Define H(·,·) as the cross - entropy loss function

[0071]

[0072] 6) For the D diagnosis head of the discriminator, we expect it to be able to accurately diagnose AD patients on real data. Let the label of the input image be y, and there is

[0073]

[0074] In summary, the final loss function can be written in the above linear combination form. Let λ be the constraint coefficient.

[0075] Exemplarily, the constraint coefficient is set to 5 in this embodiment, and there is

[0076]

[0077] Based on the above image decomposition results, the present invention proposes a new semi - quantitative index for evaluating the deposition levels of pathological markers (such as amyloid and tau proteins, etc.) in the brains of AD patients, namely the Alzheimer’s Disease Adversarial Decomposition (ADAD) score: This result is defined as the sum of the components deleted in the image to represent the deposition level of pathological biomarkers.

[0078]

[0079] where v is the physical volume of the voxel, which is 0.15 cm in this example 3 。

[0080] Optionally, the model training and deployment of this embodiment are as follows:

[0081] (1) Data source: Collect 2053 AβPET images and 953 tau PET images of patients with a clear diagnosis of AD or healthy patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI) for the training of the discriminator's true - false discrimination head and diagnosis head; to further enhance generalization, 3256 AβPET and 224 tau PET images collected from Wuhan Union Hospital are extended, and this part of the images is only used for the training of the discriminator's true - false discrimination head of the neural network. The data is divided into a training set, a validation set, and an internal test set in a ratio of 8:1:1. The public datasets provided by the Centiloid and CenTauR projects in The Global Alzheimer’s Association Interactive Network (GAINN) are used as external tests to evaluate the generalization performance of ADL.

[0082] (2) Data pre - processing: Following the method of ADNI PET core 3, the first scan of each imaging agent is re - positioned to have a 1.5 mm 3A standard 160×160×96 voxel image grid is used, and standard AC-PC alignment is implemented to maintain a consistent spatial orientation. Cerebellar gray matter is used as a reference region for normalizing the Standard Uptake Value ratio (SUVr).

[0083] (3) Optimizer: In this embodiment, the Adam optimizer is used, and the learning rate is set to 0.0001.

[0084] (4) Data augmentation: In this embodiment, data augmentation (random flipping, affine transformation, contrast adjustment, Gaussian blur, and Gaussian noise) is used to improve the generalization ability of the model. In addition, the dataset is supplemented with different attenuation correction strategies (no attenuation correction, measured attenuation correction, and zero echo time attenuation correction).

[0085] (5) Model evaluation: The area under the receiver operator characteristic curve, F1 value, and confusion matrix are used to comprehensively evaluate the AD diagnosis ability of the discriminator on real data, and the model with the highest F1 score on the validation set is retained as the final model.

[0086] (6) Support local CPU / GPU deployment and can be integrated into the hospital PACS system to achieve automated image analysis.

[0087] Furthermore, the system device and performance optimization scheme of this embodiment are as follows:

[0088] (1) Hardware device: The server uses a high-performance GPU for deep learning inference, a high-speed SSD for storage to improve the read and write speed of image data, and the network environment supports high-speed data transmission and can seamlessly connect to the hospital PACS system. In addition, the system can also be deployed using a CPU, which is suitable for institutions and individuals with limited computing resources.

[0089] (2) Computational acceleration: TensorRT is used to optimize model inference to improve computational efficiency; through mixed-precision training (FP16), the consumption of computational resources is reduced and the inference speed is increased. Support for exporting the ONNX format is provided for easy cross-platform deployment.

[0090] As an alternative implementation, this embodiment provides three processes for building an Alzheimer's disease PET image analysis system based on adversarial decomposition learning, which are specifically as follows:

[0091] (1) As shown in Figure 3 , the typical PiB Aβ PET image is input into the system for analysis, where Figure 3The top is the original image, the middle is the image after removing AD-related components, and the bottom is the probability map output by the decoupler.

[0092] Image preprocessing: Rigidly align the PET images and extract the SUV of the cerebellar gray matter for intensity normalization.

[0093] Pathological decomposition: Input the images into the decoupler to generate pathological probability maps.

[0094] Diagnostic evaluation: The program calculates the ADAD Score (the optimal cut-off point can be obtained according to Figure 6 ), and the discriminator outputs the AD probability. Clinicians combine visual evaluation to determine whether the subject is an AD patient.

[0095] Result visualization: Display the pathological marker deposition probability map calculated by the decoupler based on the input images in the form of a heat map to show potential lesion areas.

[0096] ADL accurately identifies abnormal Aβ deposits in areas such as the frontal, parietal, occipital, temporal lobes, and precuneus, and effectively decomposes the signal into an imaging pattern approximating that of normal CN

[0097] (2) Input the tau PET images into the system for analysis

[0098] Image preprocessing: Rigidly align the PET images and extract the SUV of the cerebellar gray matter for intensity normalization.

[0099] Pathological decomposition: Input the images into the decoupler to generate pathological probability maps.

[0100] Diagnostic evaluation: Calculate the ADAD Score and classify based on the discriminator results.

[0101] Result visualization: Generate an interpretable heat map to show the lesion areas.

[0102] The output results are as Figure 4 , Figure 4 The top is the original image, the middle is the image after removing AD-related components, and the bottom is the probability map output by the decoupler.

[0103] ADL successfully identifies abnormal deposits distributed in areas such as the parahippocampal gyrus, parietal lobe, frontal lobe, and precuneus, and selectively deletes these components to simulate the characteristics of normal Tau PET images

[0104] (3) Input the PET images of cognitively normal subjects into the system for analysis

[0105] Image preprocessing: Rigidly align the PET images and extract the SUV of the cerebellar gray matter for intensity normalization.

[0106] Pathological decomposition: Input the image into the decoupler to generate a pathological probability map.

[0107] Diagnostic evaluation: Calculate the ADAD Score and classify it in combination with the discriminator result.

[0108] Result visualization: Generate an interpretable heat map to display the lesion area.

[0109] The output results are as Figure 5 , Figure 5 Far left: The probability map output by the decoupler for a negative AβPET, with no significant deposition of pathological markers found; Figure 5 Middle left: The attribution map obtained by the Guided Grad-CAM interpretable method for the same AβPET, showing that its interpretability is not high; Figure 5 Middle right: The probability map output by the decoupler for a negative tau PET; Figure 5 Far right: The attribution map obtained by the Guided Grad-CAM interpretable method for the same tau PET.

[0110] No abnormal deposits were found in the probability map output by the decoupler. In contrast, the attribution map output by the Guided Grad-CAM interpretable method does not have good interpretability.

[0111] As Figure 7 shown, the ADAD score is in good agreement with the existing brain PET semi-quantitative metrics (Centiloid and CenTauRz), but has better agreement with other AD metrics such as multiple neuropsychological scales, cerebrospinal fluid, and hippocampal volume atrophy.

[0112] Corresponding to the above method, as Figure 9 shown, this embodiment also provides a visual-assisted diagnosis system for Alzheimer's disease pathology PET, including:

[0113] An image preprocessing unit for preprocessing the target PET image of the subject to obtain a preprocessed image;

[0114] An adversarial decomposition unit for inputting the preprocessed image into a trained adversarial decomposition learning model to decompose the preprocessed image and generate a pathological probability map and a diagnostic probability based on the decomposition result;

[0115] A visual evaluation unit for calculating the Alzheimer's disease adversarial decomposition score based on the decomposition result of the adversarial decomposition learning model and performing visual evaluation in combination with the diagnostic probability to determine whether the subject is a patient with Alzheimer's disease;

[0116] A visualization display unit for displaying the pathological probability map in the form of a heat map to visually display potential lesion areas.

[0117] This embodiment provides a new image decomposition mode: Existing Centiloid / CenTauRz and other semi - quantitative calculation methods rely on MR images and complex spatial normalization processes, and require manual coarse registration. The manual coarse registration process is time - consuming and vulnerable to human factors, resulting in a complex and inefficient overall process. This study addresses these challenges by introducing an interpretable deep - learning strategy, adversarial decomposition learning (ADL), to analyze AD PET images. ADL decouples disease - related biomarkers (Aβ and tau) from physiological uptake, enabling more accurate identification of AD - related changes without spatial normalization.

[0118] This embodiment is an interpretable enhanced interactive system for medical researchers to use: Traditional deep - learning diagnostic systems are usually regarded as "black - box" models, making it difficult to explain their decision - making processes and affecting doctors' trust in model results. Existing classification neural network interpretability techniques, such as Gradient - weighted Class Activation Mapping (GradCAM) and saliency map by analyzing gradient information, and input - perturbation - based interpretability techniques, such as Local interpretable model - agnostic explanations, etc., have certain hints for the decision - making process of trained neural networks, but are often unstable, contain a lot of irrelevant noise, and have extremely limited interpretability. The present invention explicitly models the distribution pattern of AD pathological biomarker pathophysiological deposition characteristics on PET images to generate a pathological probability map, enabling doctors to visually view the distribution of potential lesion areas in the image and improving interpretability. The system uses a heat map to display potential lesion areas and provides an "ideal" "healthy" image after removing AD - related components, enabling doctors to compare the images before and after removing pathological components and assisting in the diagnosis of AD through a novel human - machine interaction method.

[0119] Since this embodiment integrates high - performance computing hardware, a high - performance graphics computing graphics card, to accelerate the inference process of the ADL model and achieve efficient computing. The model can also be exported in ONNX (Open Neural Network exchange) format and deployed on ordinary computing devices, thus meeting the requirements of different medical environments. Design an optimized data storage and transmission scheme, combined with the hospital PACS system, to improve the access efficiency of image data, reduce computing latency, and ensure the usability and stability of the system in a medical environment.

[0120] The beneficial effects of the present invention are as follows:

[0121] (1) The present invention does not need to rely on spatially normalized PET pathological image decomposition.

[0122] (2) The present invention has high diagnostic accuracy. On the ADNI dataset, the AUC values of the ADL method reach 0.88 (AβPET) and 0.90 (Tau PET), which are better than the traditional Centiloid and CenTauR methods.

[0123] (3) The present invention has strong interpretability, generates pathological probability maps, enables doctors to visually view the lesion areas in the images, and improves the transparency of decision-making.

[0124] (4) The present invention can accelerate calculations by combining high-performance GPUs, is suitable for large-scale data processing, uses a distributed computing framework to optimize the inference process, and shortens the single-case image analysis time to within 5 seconds.

[0125] (5) The present invention can be integrated with the hospital PACS system to achieve automated image analysis.

[0126] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0127] In this article, specific examples are used to elaborate on the principle and implementation manner of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention. At the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation on the present invention.

Claims

1. A visually assisted diagnosis method for Alzheimer's disease pathology PET, characterized in that: include: preprocessing the target PET image of the subject to obtain a preprocessed image; The preprocessed image is fed into a trained adversarial decomposition learning model to decompose the preprocessed image, and a pathology probability map and a diagnosis probability are generated based on the decomposition result; Calculating an Alzheimer's disease adversarial decomposition score based on the decomposition results of the adversarial decomposition learning model, and performing a visual assessment in combination with the diagnosis probability to determine whether the subject is an Alzheimer's disease patient; The pathology probability map is presented in the form of a heat map to visualize potential lesion areas.

2. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 1, characterized in that: The target PET image of the subject is preprocessed to obtain a preprocessed image, including: Rigidly aligning the target PET image to obtain an aligned image; identifying a cerebellar gray matter region in the aligned image; The standard uptake value of the cerebellar gray matter region is extracted and intensity normalized to obtain the normalized preprocessed image.

3. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 1, characterized in that: The adversarial decomposition learning model includes a decoupler and a discriminator; the decoupler is used to decompose the preprocessed image into a linear superposition of a physiological uptake part and a pathological uptake part to obtain a decomposition result, and determine the pathological probability map according to the decomposition result; The discriminator is used to distinguish whether the input image has been modified by the decoupler and the Alzheimer's disease component has been stripped, and to evaluate the proportion of the decomposed part of the image to the original image, and to diagnose the input image as a normal state or Alzheimer's disease; the input image includes an image of the physiologically captured part and an original preprocessed image.

4. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 3, characterized in that: The decoupler uses 6 convolutional layers with residual connections, the number of channels of each convolutional layer is 16, 32, 32, 32, 32 and 32 respectively, and the step size is 2; the activation function of the decoupler uses a parametric linear rectifier unit.

5. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 3, characterized in that: The discriminator adopts a multi-layer convolutional network; the discriminator includes a true and false identification classification head and a diagnosis classification head.

6. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 4, characterized in that: The total loss function of the adversarial decomposition learning model is: in, is the total loss function, is the loss function of the decoupler, E V～data represents the expectation, D TrueFake (V') represents the result of the evaluation of the true and false identification classification head on the image processed by the decoupler, D diagnosis (V') represents the evaluation result of the diagnostic classification head on the image processed by the decoupler; is the L1 regularization loss, P(x,y,z) is the three-dimensional representation of the probability field P, where x, y, and z represent the x-axis coordinate value, y-axis coordinate value, and z-axis coordinate value of the voxel in the image, respectively; is the loss function of the true and false identification classification head, Among them, D TrueFake (V) represents the result of the evaluation of the true and false identification classification head on the pre-processed image; H(·,·) is the cross entropy loss function, and V represents the input pre-processed image; is the loss function of the diagnostic classification head, y is the preset image label, D diagnosis (V) represents the evaluation result of the diagnostic classification head on the preprocessed image; λ is the preset constraint coefficient.

7. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 6, characterized in that: The calculation formula for the Alzheimer's disease resistance decomposition score is: Among them, ADAD score is the Alzheimer's disease resistance decomposition score, V(x, y, z) is the three-dimensional representation of the preprocessed image V, and v is the physical volume of the voxel.

8. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 6, characterized in that: The value of the constraint coefficient is 5.

9. The visually assisted diagnosis method for Alzheimer's disease pathology PET according to claim 7, characterized in that: The physical volume of the voxel is 0.15 cm 3 .

10. A visually assisted diagnosis system for Alzheimer's disease pathology PET, characterized in that: include: An image preprocessing unit, used for preprocessing a target PET image of a subject to obtain a preprocessed image; An adversarial decomposition unit, used for adding the preprocessed image to a trained adversarial decomposition learning model to decompose the preprocessed image, and generating a pathology probability map and a diagnosis probability based on the decomposition result; a visual assessment unit, configured to calculate an Alzheimer's disease adversarial decomposition score based on the decomposition result of the adversarial decomposition learning model, and to perform a visual assessment in combination with the diagnosis probability to determine whether the subject is an Alzheimer's disease patient; The visualization display unit is used to display the pathology probability map in the form of a heat map to visualize the potential lesion area.

Images