PET-based epilepsy postoperative prognosis information generation system and method

Abstract

The invention discloses a PET-based epilepsy postoperative prognosis information generation system and method, and the system comprises a preprocessing module, a target training data generation module, a training module and an application module which are sequentially connected. Through processing the whole brain health PET image, the preoperative whole brain PET image and the postoperative whole brain magnetic resonance imaging, the obtained target training data focuses on the metabolic connection characteristics of the focus area and the whole brain, a new view angle is provided for the postoperative prognosis prediction, so that the target training data is adopted to train the deep learning model, and the prognosis prediction accuracy is improved. A deep learning model focusing on the metabolic connection characteristics of the focus area and the whole brain can be obtained, so that the postoperative whole brain magnetic resonance imaging is accurately processed, and prognosis information is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of medical equipment, and in particular to a PET-based epilepsy postoperative prognostic information generation system, method, terminal, and computer-readable storage medium. Background Art

[0002] Epilepsy is a chronic neurological disease caused by abnormal, synchronized discharges of brain neurons. Its manifestations include sudden loss of consciousness, limb convulsions, paresthesias, and autonomic dysfunction. For patients with drug-refractory epilepsy, surgical resection of the epileptic lesion is one of the most effective treatments. However, postoperative prognosis varies widely, with some patients achieving complete seizure freedom after surgery, while others may experience recurrences. Therefore, predicting postoperative prognosis is crucial for optimizing treatment strategies.

[0003] However, when predicting postoperative prognosis, it is necessary to obtain corresponding information based on magnetic resonance imaging and corresponding PET images. However, the current method used to extract information from magnetic resonance imaging and corresponding PET images lacks feature extraction capabilities, making it difficult to capture high-dimensional information and whole-brain metabolic connection characteristics. In addition, the model has limited generalization capabilities and is difficult to adapt to the complex brain network changes of different groups, resulting in the inability to accurately predict postoperative prognosis with the help of magnetic resonance imaging and corresponding PET images.

[0004] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0005] The main purpose of the present invention is to provide a PET-based epilepsy postoperative prognostic information generation system, method, terminal and computer-readable storage medium, aiming to solve the problem in the prior art that when predicting postoperative prognosis, it is necessary to obtain corresponding information based on magnetic resonance imaging and corresponding PET images. However, when extracting information from magnetic resonance imaging and corresponding PET images, the method currently used has insufficient feature extraction capability and is difficult to capture high-dimensional information and whole-brain metabolic connection characteristics. In addition, the model has limited generalization capability and is difficult to adapt to the complex brain network changes of different groups, resulting in the inability to accurately predict postoperative prognosis with the help of magnetic resonance imaging and corresponding PET images.

[0006] To achieve the above-mentioned object, the present invention provides a PET-based epilepsy postoperative prognostic information generation system, which comprises a preprocessing module, a target training data generation module, a training module and an application module, which are connected in sequence; The preprocessing module is configured to obtain a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images, and perform preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging images; The target training data generation module is used to locate the lesion area on the corresponding pre-operative whole-brain PET image after pre-processing according to each pre-processed post-operative whole-brain magnetic resonance imaging, and extract voxels; based on all pre-processed healthy whole-brain PET images, perform a single sample test on each voxel of each pre-processed whole-brain PET image after pre-processing, and perform false discovery rate correction to obtain a third preset number of voxel clusters of each pre-processed whole-brain PET image; based on each voxel cluster and the voxel of the corresponding pre-processed whole-brain PET image, obtain a third preset number of correlation maps of each pre-processed whole-brain PET image after pre-processing; and generate target training data based on the category corresponding to each pre-processed whole-brain PET image and the corresponding third preset number of correlation maps; The training module is used to obtain multiple pre-selected deep learning models, train each deep learning model according to the target training data, and select a target deep learning model after the training is completed based on the training effect of each deep learning model; The application module is used to receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps based on the first preset number of whole-brain healthy PET images, and input the third preset number of target correlation maps into the target deep learning model to generate prognostic information.

[0007] Optionally, the preprocessing module includes: A data acquisition unit, configured to acquire a first preset number of whole-brain healthy PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images; a PET image processing unit configured to obtain, using a brain extraction tool, a brain parenchymal extent from each of the whole-brain healthy PET image and the preoperative whole-brain PET image, register the corresponding brain parenchymal extent to a standard Montreal Neurological Institute space using a linear image registration tool, and then perform normalization based on the intensity of the registered brain parenchymal extent to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images; The postoperative whole-brain MRI processing unit is configured to register each postoperative whole-brain MRI to the standard Montreal Neurological Institute space using a linear image registration tool, and to align the images with the corresponding preoperative whole-brain PET images to obtain a second preset number of preprocessed postoperative whole-brain MRI images.

[0008] Optionally, the PET image processing unit includes: a registration subunit, configured to obtain a brain parenchymal extent from each of the whole-brain healthy PET images and the preoperative whole-brain PET image according to the brain extraction tool, and register the corresponding brain parenchymal extent to a standard Montreal Neurological Institute space according to the linear image registration tool, thereby obtaining a registered brain parenchymal extent corresponding to each of the whole-brain healthy PET image and the preoperative whole-brain PET image; The normalization subunit is used to obtain the range of each registered brain substance, calculate the average intensity and the intensity standard deviation, and normalize the intensity of each registered brain substance according to the average intensity and the intensity standard deviation to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images.

[0009] Optionally, the target training data generating module includes: a voxel extraction unit, configured to locate the lesion area on the corresponding pre-processed pre-operative whole-brain PET image according to each pre-processed post-operative whole-brain MRI, and extract voxels; a voxel cluster forming unit, configured to calculate a normal metabolic intensity mean and a normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images, perform a single-sample test on each voxel of each preprocessed preoperative whole-brain PET image based on the normal metabolic intensity mean and the normal metabolic intensity standard deviation, obtain voxel points of each preprocessed preoperative whole-brain PET image through false discovery rate correction and a preset threshold, select a third preset number of seed points from the voxel points of each preprocessed preoperative whole-brain PET image, and form the third preset number of voxel clusters accordingly; The target training data forming unit is used to calculate the correlation between each voxel cluster and the corresponding whole brain, generate a third preset number of correlation maps for each preprocessed preoperative whole-brain PET image, and generate target training data according to the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps.

[0010] Optionally, the voxel cluster forming unit includes: a calculation subunit, configured to calculate a normal metabolic intensity mean and a normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images; a single sample test subunit, configured to perform a single sample test on each voxel of each pre-processed preoperative whole-brain PET image according to the normal metabolic intensity mean and the normal metabolic intensity standard deviation, to obtain a single sample test result; The voxel point formation subunit is used to generate a p-value list based on the single-sample test results of each voxel, sort all the p-value lists, and perform false discovery rate correction based on a preset threshold to obtain the voxel points of each pre-processed preoperative whole-brain PET image; The voxel cluster construction subunit is used to select the third preset number of voxel points with the largest differences from the voxel points of each preprocessed preoperative whole-brain PET image as seed points, and extract the fourth preset number of voxel points around each seed point to form a voxel cluster, thereby obtaining the third preset number of voxel clusters.

[0011] In addition, to achieve the above-mentioned purpose, the present invention further provides a method for generating prognostic information after epilepsy surgery based on PET, wherein the method for generating prognostic information after epilepsy surgery based on PET comprises: Acquiring a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and performing preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging; According to each preprocessed postoperative whole-brain magnetic resonance imaging, the lesion area is located on the corresponding preprocessed preoperative whole-brain PET image, and voxels are extracted; according to all preprocessed healthy whole-brain PET images, a single sample test is performed on each voxel of each preprocessed preoperative whole-brain PET image, and false discovery rate correction is performed to obtain a third preset number of voxel clusters of each preprocessed preoperative whole-brain PET image; according to each voxel cluster and the voxel of the corresponding preprocessed preoperative whole-brain PET image, a third preset number of correlation maps of each preprocessed preoperative whole-brain PET image are obtained; according to the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps, target training data is generated; Obtain multiple pre-selected deep learning models, train each deep learning model according to the target training data, and after the training is completed, select the target deep learning model based on the training effect of each deep learning model; Receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps based on the first preset number of whole-brain healthy PET images, input the third preset number of target correlation maps into the target deep learning model to generate prognostic information.

[0012] Optionally, acquiring a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and performing preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging, includes: Acquire a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images; obtaining, using a brain extraction tool, a brain parenchymal extent from each of the whole-brain healthy PET image and the preoperative whole-brain PET image, registering the corresponding brain parenchymal extent to a standard Montreal Neurological Institute space using a linear image registration tool, and then normalizing the intensity of the registered brain parenchymal extent to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images; Each postoperative whole-brain MRI was registered to the standard Montreal Neurological Institute space using a linear image registration tool and aligned with the corresponding preoperative whole-brain PET image to obtain a second preset number of preprocessed postoperative whole-brain MRIs.

[0013] Optionally, based on each preprocessed postoperative whole-brain magnetic resonance imaging, the lesion area is located on the corresponding preprocessed preoperative whole-brain PET image, and voxels are extracted; based on all preprocessed healthy whole-brain PET images, a single sample test is performed on each voxel of each preprocessed preoperative whole-brain PET image, and false discovery rate correction is performed to obtain a third preset number of voxel clusters for each preprocessed preoperative whole-brain PET image; based on each voxel cluster and the voxel of the corresponding preprocessed preoperative whole-brain PET image, a third preset number of correlation maps for each preprocessed preoperative whole-brain PET image is obtained; based on the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps, target training data is generated, including: Based on each pre-processed post-operative whole-brain MRI, the lesion area was located on the corresponding pre-processed pre-operative whole-brain PET image, and voxels were extracted; Calculating a normal metabolic intensity mean and a normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images, performing a single-sample test on each voxel of each preprocessed preoperative whole-brain PET image based on the normal metabolic intensity mean and the normal metabolic intensity standard deviation, and obtaining voxel points of each preprocessed preoperative whole-brain PET image through false discovery rate correction and a preset threshold, selecting a third preset number of seed points from the voxel points of each preprocessed preoperative whole-brain PET image, and forming a third preset number of voxel clusters accordingly; Calculate the correlation between each voxel cluster and the corresponding whole brain, generate a third preset number of correlation maps for each preprocessed preoperative whole-brain PET image, and generate target training data based on the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps.

[0014] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a PET-based epilepsy postoperative prognostic information generation program stored in the memory and runnable on the processor, and when the PET-based epilepsy postoperative prognostic information generation program is executed by the processor, the steps of the PET-based epilepsy postoperative prognostic information generation method described above are implemented.

[0015] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a PET-based epilepsy postoperative prognostic information generation program, and when the PET-based epilepsy postoperative prognostic information generation program is executed by a processor, it implements the steps of the PET-based epilepsy postoperative prognostic information generation method as described above.

[0016] In the present invention, the PET-based epilepsy postoperative prognosis information generation system includes a preprocessing module, a target training data generation module, a training module and an application module, which are connected in sequence; wherein the preprocessing module is used to obtain a first preset number of whole-brain healthy PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and perform preprocessing to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging; the target training data generation module is used to locate the lesion area on the corresponding preprocessed preoperative whole-brain PET image according to each preprocessed postoperative whole-brain magnetic resonance imaging, and extract voxels, and perform a single sample test on each voxel of each preprocessed preoperative whole-brain PET image according to all the preprocessed whole-brain healthy PET images, and obtain each preprocessed whole-brain magnetic resonance imaging through false discovery rate correction. A third preset number of voxel clusters of the preprocessed preoperative whole-brain PET image are obtained, and a third preset number of correlation maps of each preprocessed whole-brain PET image are obtained according to each voxel cluster and the corresponding voxel of the preprocessed whole-brain PET image. Target training data are generated according to the category corresponding to each preprocessed whole-brain PET image and the corresponding third preset number of correlation maps; the training module is used to obtain multiple pre-selected deep learning models, train each deep learning model according to the target training data, and select a target deep learning model according to the training effect of each deep learning model after the training; the application module is used to receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps according to the first preset number of whole-brain healthy PET images, and input the third preset number of target correlation maps into the target deep learning model to generate prognosis information. The present invention processes whole-brain healthy PET images, preoperative whole-brain PET images, and postoperative whole-brain MRI images to obtain target training data that focuses on the metabolic connection characteristics between the lesion area and the whole brain, providing a new perspective for postoperative prognosis prediction. By using this target training data to train a deep learning model, a deep learning model that focuses on the metabolic connection characteristics between the lesion area and the whole brain can be obtained, thereby accurately classifying postoperative whole-brain MRI images and helping corresponding staff to achieve accurate postoperative prognosis prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a structural diagram of a preferred embodiment of a PET-based epilepsy postoperative prognostic information generation system of the present invention; Figure 2 This is a schematic diagram of an implementation flow of the PET-based epilepsy postoperative prognostic information generation system of the present invention; Figure 3This is a schematic diagram showing the brain map of the lesion location in the method for generating prognostic information after epilepsy surgery based on PET of the present invention; Figure 4 is a schematic diagram of a correlation diagram in the method for generating prognostic information after epilepsy surgery based on PET of the present invention; Figure 5 This is a bar chart showing the accuracy of different models in the PET-based method for generating prognostic information after epilepsy surgery under five-fold and three-fold cross validation; Figure 6 This is a bar chart of the accuracy of different models in the PET-based epilepsy postoperative prognostic information generation method of the present invention under five-fold and three-fold cross validation; Figure 7 This is a histogram of the recall rates of different models in the PET-based epilepsy postoperative prognostic information generation method under five-fold and three-fold cross validation; Figure 8 It is a histogram of the specificity rates of different models in the PET-based method for generating prognostic information after epilepsy surgery under five-fold and three-fold cross validation; Figure 9 It is a histogram of the F1 values of different models in the PET-based epilepsy postoperative prognostic information generation method of the present invention under five-fold and three-fold cross validation; Figure 10 The figure is a histogram of the AUC of different models in the PET-based method for generating prognostic information after epilepsy surgery under five-fold and three-fold cross validation; Figure 11 It is a flow chart of a preferred embodiment of the method for generating prognostic information after epilepsy surgery based on PET of the present invention; Figure 12 FIG. 4 is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] Epilepsy is a chronic neurological disease caused by abnormal synchronous discharges of brain neurons. Its manifestations include sudden loss of consciousness, limb convulsions, paresthesias, and autonomic dysfunction. For patients with drug-refractory epilepsy, surgical resection of the epileptic lesion is one of the most effective treatments. However, postoperative prognosis varies greatly. Some patients can be completely seizure-free after surgery, while others may still relapse. Therefore, predicting postoperative prognosis is important for optimizing treatment strategies. Currently, positron emission tomography (PET) has become an important tool for localizing epileptic lesions and evaluating postoperative prognosis because it can reflect brain metabolic levels.

[0020] At present, the methods for predicting the prognosis of epilepsy surgery mainly include the following: traditional statistical methods, using PET, MRI (Magnetic Resonance Imaging, magnetic resonance imaging) and DTI (Diffusion Tensor PET images are used to predict the lateralization of temporal lobe epilepsy through regression analysis. The low metabolic area of PET is a strong indicator of epileptic focus lateralization. However, traditional statistical methods have obvious defects. First, it is difficult to capture the high-dimensional features of PET images and the feature extraction ability is limited. Second, it relies on linear assumptions, while the postoperative brain network changes of epilepsy patients usually show nonlinear characteristics, resulting in limited prediction accuracy. Traditional machine learning methods, such as support vector machines and random forests, are widely used in epilepsy image analysis. These methods rely on manual feature engineering, have difficulty in automatically learning deep features, and lack the ability to capture the metabolic connectivity patterns of the whole brain, which limits the accuracy and generalization of predictions. Deep learning is effective in medical image analysis. The quantitative asymmetric features of PET images are combined with deep learning models to predict postoperative seizure recurrence in patients with refractory epilepsy. Although deep learning improves prediction performance, existing methods are mostly limited to metabolic analysis of specific brain regions (such as the temporal lobe) and fail to fully utilize the metabolic connectivity patterns of the whole brain, resulting in limited prediction performance and generalization ability.

[0021] Therefore, when predicting postoperative prognosis, it is necessary to obtain corresponding information based on magnetic resonance imaging and corresponding PET images. However, the current method used to extract information from magnetic resonance imaging and corresponding PET images lacks feature extraction capabilities, making it difficult to capture high-dimensional information and whole-brain metabolic connection characteristics. In addition, the model has limited generalization capabilities and is difficult to adapt to the complex brain network changes of different groups, resulting in the inability to accurately predict postoperative prognosis with the help of magnetic resonance imaging and corresponding PET images.

[0022] In response to one or more of the above problems, the present invention provides a PET-based epilepsy postoperative prognostic information generation system, which includes a preprocessing module, a target training data generation module, a training module and an application module, which are connected in sequence; wherein the preprocessing module is used to obtain a first preset number of whole-brain healthy PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and perform preprocessing to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging; the target training data generation module is used to locate the lesion area on the corresponding preprocessed preoperative whole-brain PET image according to each preprocessed postoperative whole-brain magnetic resonance imaging, and extract voxels, and perform single sample detection on each voxel of each preprocessed preoperative whole-brain PET image according to all preprocessed whole-brain healthy PET images. The method comprises the following steps: testing and correcting the false discovery rate to obtain a third preset number of voxel clusters of each preprocessed preoperative whole-brain PET image, obtaining a third preset number of correlation maps of each preprocessed preoperative whole-brain PET image according to each voxel cluster and the corresponding voxel of the preprocessed preoperative whole-brain PET image, and generating target training data according to the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps; the training module is used to obtain a plurality of pre-selected deep learning models, train each deep learning model according to the target training data, and select a target deep learning model according to the training effect of each deep learning model after the training; the application module is used to receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps according to the first preset number of whole-brain healthy PET images, input the third preset number of target correlation maps into the target deep learning model, and generate prognostic information.

[0023] The PET-based epilepsy postoperative prognostic information generation system described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the PET-based epilepsy postoperative prognostic information generation system includes a preprocessing module 11, a target training data generation module 12, a training module 13 and an application module 14, which are connected in sequence.

[0024] Specifically, in the present invention, for the PET-based epilepsy postoperative prognosis information generation system, the preprocessing module, the target training data generation module and the training module are mainly used to train the model. After the model is trained, the data can be continuously processed through the application module. Figure 2As shown, data preprocessing and voxel screening are completed in the preprocessing module, the target training data generation module and the training module, so as to train the deep learning model and conduct experiments, and the target deep learning model is screened out by performance. After that, the application module continues to accept the data to be processed and processes it through the target deep learning model.

[0025] Among them, the preprocessing module is used to obtain a first preset number of whole-brain healthy PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and perform preprocessing to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging.

[0026] Specifically, in the present invention, the preprocessing module includes: a data acquisition unit, configured to acquire a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging; a PET image processing unit, configured to acquire the brain parenchyma range from each of the healthy whole-brain PET images and the preoperative whole-brain PET image according to a brain extraction tool, and to register the corresponding brain parenchyma range to the standard Montreal Neurological Institute space according to a linear image registration tool, and then normalize the range according to the intensity of the registered brain parenchyma range to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images; and a postoperative whole-brain magnetic resonance imaging processing unit, configured to register each postoperative whole-brain magnetic resonance imaging to the standard Montreal Neurological Institute space according to the linear image registration tool, and to align the images according to the corresponding preoperative whole-brain PET images to obtain a second preset number of preprocessed postoperative whole-brain magnetic resonance imaging.

[0027] The data acquisition unit of the present invention acquires a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images, wherein the first preset number and the second preset number are pre-set by the user. In a preferred embodiment of the present invention, the first preset number is set to 18 and the second preset number is set to 20, and the healthy whole-brain PET images are derived from brain PET images of 18 healthy individuals in a public dataset, and the preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images are derived from preoperative brain PET images and postoperative T1-weighted Magnetic Resonance Imaging (T1-weighted Magnetic Resonance Imaging) images of patients with focal epilepsy. Preoperative whole-brain PET images are used to analyze the metabolic characteristics of epilepsy lesion regions and extract the connectivity patterns between the lesion region and the entire brain. Postoperative whole-brain MRIs mark the surgical region (i.e., the resected lesion) for each patient and provide a binary mask to guide the localization and feature extraction of the lesion region in preoperative whole-brain PET images. Whole-brain healthy PET images provide a reference for metabolic intensity and are compared with preoperative whole-brain PET images of epilepsy patients to identify voxels with significant metabolic differences within the lesion region. Furthermore, in this invention, recurrence is categorized into two categories: no seizures within two years after surgery are assigned label 1, and seizures within two years after surgery are assigned label 2.

[0028] The PET image processing unit includes: a registration subunit, configured to obtain a brain parenchymal range from each of the whole-brain healthy PET images and the preoperative whole-brain PET image according to the brain extraction tool, and to register the corresponding brain parenchymal range to the standard Montreal Neurological Institute space according to the linear image registration tool, thereby obtaining a registered brain parenchymal range corresponding to each of the whole-brain healthy PET image and the preoperative whole-brain PET image; and a normalization subunit, configured to obtain each registered brain parenchymal range, calculate an average intensity and an intensity standard deviation, and normalize the intensity of each registered brain parenchymal range according to the average intensity and the intensity standard deviation, thereby obtaining a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images.

[0029] The PET image processing unit is configured to process a first preset number of whole-brain healthy PET images and a second preset number of preoperative whole-brain PET images. The registration subunit and normalization subunit can be performed using FSL (FMRIB Software Library) software. Specifically, in the present invention, the registration subunit utilizes the BET (Brain Extraction Tool) to remove non-brain tissue (including skull and cerebrospinal fluid) from the preoperative whole-brain PET images and the whole-brain healthy PET images, thereby limiting the analysis scope to the brain parenchyma. Subsequently, all PET images are registered to the standard Montreal Neurological Institute (MNI) space using FLIRT (FMRIB's Linear Image Registration Tool) to eliminate anatomical differences between subjects and achieve spatial consistency of the data. The registered brain parenchyma area corresponding to each whole-brain healthy PET image and the preoperative whole-brain PET image is then obtained.

[0030] The normalization subunit then processes the image using FSL's deviation field correction module to correct for signal intensity deviations caused by scanner or magnetic field inhomogeneities. Because different scanners and individual physiological states can cause significant variation in PET image pixel intensity, all data are further normalized using Z-score standardization, using the following formula: ; in, represents the intensity-normalized data, Represents the original pixel value, that is, the original pixel value of the brain parenchyma range after registration corresponding to the whole-brain healthy PET image and the preoperative whole-brain PET image, is the average intensity of the brain region, is the intensity standard deviation of the brain region, and the mean intensity and standard deviation of the brain region are the mean intensity and standard deviation of the intensity across all PET images. This normalization process conforms the data to a uniform statistical distribution, effectively reducing intensity variations between individuals and devices, and providing a reliable data foundation for subsequent analysis. After normalization, a first preset number of pre-processed healthy whole-brain PET images and a second preset number of pre-processed pre-operative whole-brain PET images are obtained.

[0031] The postoperative whole-brain MRI processing unit is configured to register each postoperative whole-brain MRI to the standard Montreal Neurological Institute space using a linear image registration tool, and to align the images with the corresponding preoperative whole-brain PET images to obtain a second preset number of preprocessed postoperative whole-brain MRI images.

[0032] In the present invention, the postoperative whole-brain MRI is a postoperative T1 MRI image, which is processed in the postoperative whole-brain MRI processing unit. Specifically, the images are registered to MNI space using the FLIRT tool and aligned with the corresponding preoperative whole-brain PET images, thereby obtaining a second predetermined number of preprocessed postoperative whole-brain MRI images. This ensures that the surgical region mask can be accurately mapped to the PET image, which is used to guide voxel screening of the lesion area.

[0033] In the present invention, by Figure 3 Five samples with lesion locations are shown, namely sample 1, sample 2, sample 4, sample 12 and sample 20, presented in the form of brain maps in MNI space; among them, sample 1, sample 2 and sample 4 are patients who did not have seizures after surgery (labeled as 1), and sample 12 and sample 20 are patients who had seizures after surgery (labeled as 2). The lesion locations cover areas such as the temporal lobe, frontal lobe and occipital lobe, reflecting the anatomical diversity of the samples.

[0034] The target training data generation module is used to locate the lesion area on the corresponding pre-processed pre-operative whole-brain PET image based on each pre-processed post-operative whole-brain magnetic resonance imaging, and extract voxels; based on all pre-processed healthy whole-brain PET images, perform a single sample test on each voxel of each pre-processed pre-operative whole-brain PET image, and correct it through false discovery rate to obtain a third preset number of voxel clusters for each pre-processed pre-operative whole-brain PET image; based on each voxel cluster and the voxel of the corresponding pre-processed pre-operative whole-brain PET image, obtain a third preset number of correlation maps for each pre-processed pre-operative whole-brain PET image; and generate target training data based on the category corresponding to each pre-processed pre-operative whole-brain PET image and the corresponding third preset number of correlation maps.

[0035] Specifically, the target training data generation module in the present invention includes: a voxel extraction unit, which is used to locate the lesion area on the corresponding pre-processed pre-operative whole-brain PET image based on each pre-processed post-operative whole-brain magnetic resonance imaging, and extract voxels; a voxel cluster formation unit, which is used to calculate the normal metabolic intensity mean and the normal metabolic intensity standard deviation based on all pre-processed whole-brain healthy PET images, and perform a single sample test on each voxel of each pre-processed pre-operative whole-brain PET image based on the normal metabolic intensity mean and the normal metabolic intensity standard deviation, and obtain the voxel points of each pre-processed pre-operative whole-brain PET image through false discovery rate correction and a preset threshold, select a third preset number of seed points from the voxel points of each pre-processed pre-operative whole-brain PET image, and form a third preset number of voxel clusters accordingly; a target training data formation unit, which is used to calculate the correlation between each voxel cluster and the corresponding whole brain, generate a third preset number of correlation maps for each pre-processed pre-operative whole-brain PET image, and generate target training data according to the category corresponding to each pre-processed pre-operative whole-brain PET image and the corresponding third preset number of correlation maps.

[0036] Specifically, the voxel extraction unit uses the surgical area mask on the preprocessed postoperative whole-brain MRI to locate the lesion area on the corresponding preprocessed preoperative whole-brain PET image and extract the voxels within the lesion area; then voxel clustering is performed in the voxel clustering unit.

[0037] The voxel cluster forming unit includes: a calculation subunit, which is used to calculate the normal metabolic intensity mean and the normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images; a single sample test subunit, which is used to perform a single sample test on each voxel of each preprocessed preoperative whole-brain PET image based on the normal metabolic intensity mean and the normal metabolic intensity standard deviation, to obtain a single sample test result; a voxel point forming subunit, which is used to generate a p-value list based on the single sample test result of each voxel, sort all the p-value lists, and perform false discovery rate correction based on a preset threshold to obtain the voxel points of each preprocessed preoperative whole-brain PET image; a voxel cluster construction subunit, which is used to select a third preset number of voxel points with the largest difference from the voxel points of each preprocessed preoperative whole-brain PET image as seed points, and extract a fourth preset number of voxel points around each seed point to form a voxel cluster, with each seed point as the center, to obtain a third preset number of voxel clusters.

[0038] Specifically, in the voxel cluster formation unit, a single-sample T test is performed on the voxels extracted by the voxel extraction unit to compare whether there is a statistical difference in the distribution of the metabolic intensity of the voxels in the lesion area of each epileptic sample and the metabolic intensity of the corresponding voxels in the healthy subject group. The purpose is to screen out voxels with significant metabolic differences between epileptic samples and non-epileptic samples in the lesion area as seed points for subsequent analysis and to construct corresponding voxel clusters.

[0039] In the calculation subunit, the normal metabolic intensity mean and the normal metabolic intensity standard deviation of all pre-processed whole-brain healthy PET images are calculated. In the single-sample test subunit, a single-sample test is performed on each voxel of each pre-processed pre-operative whole-brain PET image according to the normal metabolic intensity mean and the normal metabolic intensity standard deviation to obtain a single-sample test result. Among them, the single-sample test is to calculate the metabolic intensity of each voxel of the lesion area extracted from each pre-processed pre-operative whole-brain PET image, and perform a single-sample T test with the metabolic intensity mean of the corresponding area of the whole-brain healthy PET image to obtain a single-sample test result, that is, a test statistic. Specifically, the test statistic t is calculated as follows: ; in, is the metabolic intensity value of each voxel in the extracted lesion area of each preprocessed preoperative whole-brain PET image, is the mean normal metabolic intensity of the corresponding voxels of all preprocessed whole-brain healthy PET images, is the standard deviation of normal voxel metabolic intensity corresponding to all preprocessed whole-brain healthy PET images, is the sample size of all preprocessed whole-brain healthy PET images.

[0040] After calculating the test statistic, the voxel point forming subunit adopts FDR (False Discovery Rate) correction to screen out voxel points with p-values less than 0.05 and significant differences. The p-value represents the probability of observing the current or more extreme results when the null hypothesis (that is, the metabolic intensity of the voxel is not significantly different from the mean of healthy individuals) is true. The smaller the p-value, the more significant the difference, and the more inclined to reject the null hypothesis, believing that the metabolic intensity of the voxel is significantly different from that of healthy individuals. Specifically, the specific process of FDR correction is, for each voxel in the lesion area of each pre-processed preoperative whole-brain PET image, the single-sample T test result, that is, the single-sample test result, first generate a list of p-values corresponding to each voxel; then, perform FDR correction, that is, first sort all p-values from small to large, recorded as ;in is the total number of voxels; then, for each p-value , i is 1 to , calculate the FDR correction threshold: ;in =0.05, which is the set false discovery rate control level, that is, the pre-set threshold. is the position of the p value in the sorting; start from the largest p value and compare forward to find the largest , making , then the p-value and all the p-values ranked before are considered significant, thus screening out voxels with p-values less than 0.05 and significant differences.

[0041] After obtaining the voxel points, the voxel cluster construction subunit selects a third preset number of voxel points with the largest differences from the significantly different voxel points screened within the lesion area as seed points. In one embodiment of the present invention, the third preset number is 9. With each voxel point as the center, a fourth preset number of voxel points surrounding it are extracted to form a voxel cluster, thereby obtaining the third preset number of voxel clusters. In one embodiment of the present invention, the third preset number is the same as the fourth preset number.

[0042] After obtaining the voxel cluster, the target training data formation unit calculates the correlation between the cluster and the corresponding voxels in the whole brain and generates a correlation map. The correlation calculation method is: first calculate the dissimilarity between the voxel cluster and the corresponding whole brain voxel, that is, , the formula is as follows: ; in, and are the normalized intensity values of voxel clusters and whole-brain voxels, respectively. Subsequently, similarity is calculated based on dissimilarity, i.e. , the formula is as follows: ; in, , and are the standardized intensity standard deviations of voxel clusters and whole-brain voxels, respectively. By this method, each pre-processed whole-brain PET image before surgery generates a third preset number of correlation maps, reflecting the connection pattern between the lesion area and the whole brain; according to the corresponding category of each pre-processed whole-brain PET image before surgery and the corresponding third preset number of correlation maps, target training data is generated, which contains corresponding labels. Furthermore, the present invention generates a third preset number of correlation maps after each pre-processed whole-brain PET image before surgery, which reflects the connection pattern between the lesion area and the whole brain, reflecting the connection pattern between the lesion area and the whole brain; according to the corresponding category of each pre-processed whole-brain PET image before surgery and the corresponding third preset number of correlation maps, target training data is generated, which contains corresponding labels. Figure 4 Five correlation maps are shown, with one representative correlation map selected for each preprocessed preoperative whole-brain PET image.

[0043] The training module is used to obtain multiple pre-selected deep learning models, train each deep learning model according to the target training data, and after the training is completed, select the target deep learning model according to the training effect of each deep learning model.

[0044] Specifically, after obtaining the target training data, a training module is used to select a target deep learning model. Multiple pre-selected deep learning models are trained, and after training, a target deep learning model is selected based on the training effect of each deep learning model. The training effect is judged by accuracy, precision, sensitivity, specificity, F1 score, and AUC (Area Under Curve).

[0045] In one embodiment of the present invention, four deep learning models were used for training and evaluation: Inception V4 (Inception Version 4), DenseNet (Densely Connected Convolutional Network), VGG (Visual Geometry Group Network), and ResNet34 (Residual Network with 34 Layers). The model input consisted of nine similarity graphs for each target training data point, and the output was the corresponding classification result, i.e., label 1 or 2. To verify the robustness of the model, the training process employed 20 rounds of five-fold cross-validation and 20 rounds of three-fold cross-validation to optimize model performance. Five-fold cross-validation involves randomly dividing the 20 target training data points into five subsets, using four subsets as training sets each time and the remaining subset as a test set for five cycles, ensuring that each subset was used as a test set at least once. Three-fold cross-validation involves randomly dividing the data into three subsets, using two subsets as training sets each time and the remaining subset as a test set for three cycles. 20-round cross-validation involves repeating the above process 20 times and taking the average performance indicator. This reduces bias caused by random partitioning and enhances the stability of model evaluation. To verify the effectiveness of this method, a control group experiment was conducted. The experimental group used the correlation map as model input, while the control group used preoperative PET images directly as model input. The prediction performance of the two groups was compared. Model evaluation metrics included accuracy, precision, sensitivity, specificity, F1 score, and AUC.

[0046] Finally, the training effect is obtained and the results are shown in Table 1. Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 、 Figure 9 and Figure 10Table 1 shows the performance comparison between the experimental and control groups. The prediction performance of different models under five-fold and three-fold cross-validation is listed. Taking DenseNet as an example, in the five-fold cross-validation, the experimental group achieved an accuracy of 0.9254±0.0490, a sensitivity of 0.8958±0.1163, and an F1 score of 0.8661±0.1223, representing improvements of 17.24%, 37.65%, and 43.94% respectively compared to the control group. In the three-fold cross-validation, the experimental group achieved an accuracy of 0.8510±0.0765, a sensitivity of 0.8125±0.1524, and an F1 score of 0.8109±0.1026, representing improvements of 10.82%, 24.39%, and 27.89% respectively compared to the control group. Other models showed similar trends, with improvements in all metrics ranging from 3% to 44% for the experimental group. Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 、 Figure 9 and Figure 10 The comparison results of the different models, including accuracy, precision, sensitivity, specificity, F1 score, and AUC, are presented in bar graphs, further visualizing the performance advantages of the experimental group. Taking into account metrics such as accuracy, sensitivity, and F1 score, as well as the model's stability and robustness in five-fold and three-fold cross-validation, DenseNet was ultimately selected as the preferred model for this paper. The selection criteria were as follows: First, DenseNet achieved the highest accuracy (0.9254±0.0490) and F1 score (0.8661±0.1223) in five-fold cross-validation, demonstrating its excellent performance in identifying patients with recurrence within two years of surgery (i.e., label 2), effectively reducing the risk of missed diagnosis. Second, its sensitivity (0.8958±0.1163) was the highest among all models, reflecting the model's reliability in detecting the positive class. Furthermore, DenseNet also outperformed other models in three-fold cross-validation (accuracy 0.8510±0.0765, F1 score 0.8109±0.1026), with a small standard deviation, indicating the high stability of the model's prediction results. Therefore, DenseNet was selected as the final model, the target deep learning model.

[0047] Table 1: Prediction performance of different models

[0048] The application module is used to receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps based on the first preset number of whole-brain healthy PET images, and input the third preset number of target correlation maps into the target deep learning model to generate prognostic information.

[0049] Upon receiving the preoperative whole-brain PET images and postoperative whole-brain MRIs to be processed, the trained target deep learning model generates a third preset number of target correlation maps based on the corresponding first preset number of healthy whole-brain PET images. These correlation maps are then fed into the trained target deep learning model. The target deep learning model extracts the spatial features of each correlation map through convolutional and pooling layers, performs classification prediction through a fully connected layer, and outputs the corresponding category (label 1 or label 2) for each correlation map. Given the third preset number of correlation maps, the model generates a third preset number of classification results. The votes for label 1 and label 2 in the classification results of the third preset number of correlation maps are counted. Based on the classification results of the third preset number of correlation maps, prediction scores for label 1 and label 2 are generated, i.e., the corresponding prognostic information. The corresponding staff can then use the prediction scores for label 1 and label 2 to accurately predict the postoperative prognosis.

[0050] The PET-based epilepsy postoperative prognostic information generation system of the present invention includes a preprocessing module, a target training data generation module, a training module and an application module, which are connected in sequence; wherein the preprocessing module is used to obtain a first preset number of whole-brain healthy PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and perform preprocessing to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging; the target training data generation module is used to locate the lesion area on the corresponding preprocessed preoperative whole-brain PET image according to each preprocessed postoperative whole-brain magnetic resonance imaging, and extract voxels, and perform a single sample test on each voxel of each preprocessed preoperative whole-brain PET image according to all the preprocessed whole-brain healthy PET images, and obtain each preprocessed lesion area through false discovery rate correction. A third preset number of voxel clusters of the pre-processed whole-brain PET image after surgery are obtained, and a third preset number of correlation maps of each pre-processed whole-brain PET image are obtained according to each voxel cluster and the voxel of the corresponding pre-processed whole-brain PET image. Target training data are generated according to the category corresponding to each pre-processed whole-brain PET image and the corresponding third preset number of correlation maps; the training module is used to obtain multiple pre-selected deep learning models, train each deep learning model according to the target training data, and after the training, select the target deep learning model according to the training effect of each deep learning model; the application module is used to receive the pre-operative whole-brain PET image to be processed and the corresponding post-operative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps according to the first preset number of whole-brain healthy PET images, and input the third preset number of target correlation maps into the target deep learning model to generate prognosis information. The present invention processes whole-brain healthy PET images, preoperative whole-brain PET images, and postoperative whole-brain MRI images to obtain target training data that focuses on the metabolic connection characteristics between the lesion area and the whole brain, providing a new perspective for postoperative prognosis prediction. By using this target training data to train a deep learning model, a deep learning model that focuses on the metabolic connection characteristics between the lesion area and the whole brain can be obtained, thereby accurately classifying postoperative whole-brain MRI images and helping corresponding staff to achieve accurate postoperative prognosis prediction.

[0051] Further, if Figure 11 As shown, based on the above-mentioned PET-based epilepsy postoperative prognostic information generation system, the present invention also provides a PET-based epilepsy postoperative prognostic information generation method, wherein the PET-based epilepsy postoperative prognostic information generation method includes: Step S111: acquiring a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images, and performing preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging images; Step S112: based on each preprocessed postoperative whole-brain magnetic resonance imaging, locate the lesion area on the corresponding preprocessed preoperative whole-brain PET image and extract voxels; based on all preprocessed healthy whole-brain PET images, perform a single-sample test on each voxel of each preprocessed preoperative whole-brain PET image, and perform false discovery rate correction to obtain a third preset number of voxel clusters for each preprocessed preoperative whole-brain PET image; based on each voxel cluster and the voxels of the corresponding preprocessed preoperative whole-brain PET image, obtain a third preset number of correlation maps for each preprocessed preoperative whole-brain PET image; and generate target training data based on the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps; Step S113: Acquire multiple pre-selected deep learning models, train each deep learning model according to the target training data, and after the training is completed, select a target deep learning model based on the training effect of each deep learning model; Step S114: Receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps based on the first preset number of whole-brain healthy PET images, input the third preset number of target correlation maps into the target deep learning model to generate prognostic information.

[0052] The method of obtaining a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and performing preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging, includes: Acquire a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images; obtaining, using a brain extraction tool, a brain parenchymal extent from each of the whole-brain healthy PET image and the preoperative whole-brain PET image, registering the corresponding brain parenchymal extent to a standard Montreal Neurological Institute space using a linear image registration tool, and then normalizing the intensity of the registered brain parenchymal extent to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images; Each postoperative whole-brain MRI was registered to the standard Montreal Neurological Institute space using a linear image registration tool and aligned with the corresponding preoperative whole-brain PET image to obtain a second preset number of preprocessed postoperative whole-brain MRIs.

[0053] The method includes: locating the lesion area on the corresponding pre-operative whole-brain PET image after pre-processing based on each pre-processed post-operative whole-brain magnetic resonance imaging, extracting voxels, performing a single sample test on each voxel of each pre-processed whole-brain PET image after pre-processing based on all pre-processed healthy whole-brain PET images, and performing false discovery rate correction to obtain a third preset number of voxel clusters of each pre-processed whole-brain PET image after pre-processing; obtaining a third preset number of correlation maps of each pre-processed whole-brain PET image after pre-processing based on each voxel cluster and the voxel of the corresponding pre-processed whole-brain PET image; and generating target training data based on the category corresponding to each pre-processed whole-brain PET image and the corresponding third preset number of correlation maps. The method includes: Based on each pre-processed post-operative whole-brain MRI, the lesion area was located on the corresponding pre-processed pre-operative whole-brain PET image, and voxels were extracted; Calculating a normal metabolic intensity mean and a normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images, performing a single-sample test on each voxel of each preprocessed preoperative whole-brain PET image based on the normal metabolic intensity mean and the normal metabolic intensity standard deviation, and obtaining voxel points of each preprocessed preoperative whole-brain PET image through false discovery rate correction and a preset threshold, selecting a third preset number of seed points from the voxel points of each preprocessed preoperative whole-brain PET image, and forming a third preset number of voxel clusters accordingly; Calculate the correlation between each voxel cluster and the corresponding whole brain, generate a third preset number of correlation maps for each preprocessed preoperative whole-brain PET image, and generate target training data based on the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps.

[0054] The present invention analyzes the whole-brain metabolic connectivity pattern through correlation graphs, and combines them with deep learning models to significantly improve the accuracy and robustness of postoperative prognosis prediction for epilepsy. The direct effects brought about by specific technical features include: screening significantly different voxels through single-sample T test and FDR correction, constructing correlation graphs with voxel clusters as seed points, capturing the metabolic connectivity characteristics of the lesion area and the whole brain, overcoming the defects of traditional methods that are limited to local brain areas, providing more comprehensive feature expressions, and directly improving sensitivity and F1 score indicators. When selecting deep learning models, deep learning models such as Inception V4, DenseNet, VGG and ResNet34 are used, combined with grid search to optimize hyperparameters, enhance feature extraction and classification capabilities, enable the model to adapt to changes in complex brain networks, and directly improve the accuracy and generalization of predictions. The five-fold and three-fold cross-validation designs ensure the robustness of the model, are suitable for different situations, and enhance the applicability of the method.

[0055] Furthermore, in the present invention, deep learning models are pre-selected. In addition to Inception V4, DenseNet, VGG and ResNet34, other models such as Transformer (a neural network architecture based on self-attention) or Graph Neural Network (GNN) can also be used for feature extraction and classification, which are suitable for capturing global and local features of correlation graphs.

[0056] Further, if Figure 12 As shown, based on the above-mentioned PET-based method and system for generating postoperative prognostic information for epilepsy, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 12 Only some of the components of the terminal are shown, but it should be understood that implementation of all of the shown components is not required, and more or fewer components may be implemented instead.

[0057] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard drive or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash card, etc. equipped on the terminal. Furthermore, the memory 20 may include both the internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software installed on the terminal and various types of data, such as program code of the terminal. The memory 20 may also be used to temporarily store data that has been output or is about to be output. In one embodiment, the memory 20 stores a PET-based epilepsy postoperative prognostic information generation program 40, which can be executed by the processor 10, thereby implementing the PET-based epilepsy postoperative prognostic information generation method of the present invention.

[0058] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, configured to execute program codes or process data stored in the memory 20, such as executing the PET-based method for generating postoperative prognostic information for epilepsy surgery.

[0059] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch screen, etc. The display 30 is used to display information on the terminal and to display a visual user interface.

[0060] In one embodiment, when the processor 10 executes the PET-based epilepsy postoperative prognostic information generation program 40 in the memory 20 , the steps of the above-mentioned PET-based epilepsy postoperative prognostic information generation method are implemented.

[0061] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a PET-based epilepsy postoperative prognostic information generation program, and when the PET-based epilepsy postoperative prognostic information generation program is executed by a processor, the steps of the above-mentioned PET-based epilepsy postoperative prognostic information generation method are implemented.

[0062] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal comprising the element.

[0063] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium that can be read by a computer. When executed, the program can include the processes in the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0064] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A PET-based epilepsy postoperative prognostic information generation system, characterized in that: The PET-based epilepsy postoperative prognostic information generation system includes a preprocessing module, a target training data generation module, a training module and an application module, which are connected in sequence; The preprocessing module is configured to obtain a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images, and perform preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging images; The target training data generation module is used to locate the lesion area on the corresponding pre-operative whole-brain PET image after pre-processing according to each pre-processed post-operative whole-brain magnetic resonance imaging, and extract voxels; based on all pre-processed healthy whole-brain PET images, perform a single sample test on each voxel of each pre-processed whole-brain PET image after pre-processing, and perform false discovery rate correction to obtain a third preset number of voxel clusters of each pre-processed whole-brain PET image; based on each voxel cluster and the voxel of the corresponding pre-processed whole-brain PET image, obtain a third preset number of correlation maps of each pre-processed whole-brain PET image after pre-processing; and generate target training data based on the category corresponding to each pre-processed whole-brain PET image and the corresponding third preset number of correlation maps; The training module is used to obtain multiple pre-selected deep learning models, train each deep learning model according to the target training data, and select a target deep learning model after the training is completed based on the training effect of each deep learning model; The application module is used to receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps based on the first preset number of whole-brain healthy PET images, and input the third preset number of target correlation maps into the target deep learning model to generate prognostic information.

2. The PET-based epilepsy postoperative prognostic information generation system according to claim 1, characterized in that: The pre-processing module comprises: A data acquisition unit, configured to acquire a first preset number of whole-brain healthy PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images; a PET image processing unit configured to obtain, using a brain extraction tool, a brain parenchymal extent from each of the whole-brain healthy PET image and the preoperative whole-brain PET image, register the corresponding brain parenchymal extent to a standard Montreal Neurological Institute space using a linear image registration tool, and then perform normalization based on the intensity of the registered brain parenchymal extent to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images; The postoperative whole-brain MRI processing unit is configured to register each postoperative whole-brain MRI to the standard Montreal Neurological Institute space using a linear image registration tool, and to align the images with the corresponding preoperative whole-brain PET images to obtain a second preset number of preprocessed postoperative whole-brain MRI images.

3. The PET-based epilepsy postoperative prognostic information generation system according to claim 2, characterized in that: The PET image processing unit includes: a registration subunit, configured to obtain a brain parenchymal extent from each of the whole-brain healthy PET images and the preoperative whole-brain PET image according to the brain extraction tool, and register the corresponding brain parenchymal extent to a standard Montreal Neurological Institute space according to the linear image registration tool, thereby obtaining a registered brain parenchymal extent corresponding to each of the whole-brain healthy PET image and the preoperative whole-brain PET image; The normalization subunit is used to obtain the range of each registered brain substance, calculate the average intensity and the intensity standard deviation, and normalize the intensity of each registered brain substance according to the average intensity and the intensity standard deviation to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images.

4. The PET-based epilepsy postoperative prognostic information generation system according to claim 1, characterized in that: The target training data generation module includes: a voxel extraction unit, configured to locate the lesion area on the corresponding pre-processed pre-operative whole-brain PET image according to each pre-processed post-operative whole-brain MRI, and extract voxels; a voxel cluster forming unit, configured to calculate a normal metabolic intensity mean and a normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images, perform a single-sample test on each voxel of each preprocessed preoperative whole-brain PET image based on the normal metabolic intensity mean and the normal metabolic intensity standard deviation, obtain voxel points of each preprocessed preoperative whole-brain PET image through false discovery rate correction and a preset threshold, select a third preset number of seed points from the voxel points of each preprocessed preoperative whole-brain PET image, and form the third preset number of voxel clusters accordingly; The target training data forming unit is used to calculate the correlation between each voxel cluster and the corresponding whole brain, generate a third preset number of correlation maps for each preprocessed preoperative whole-brain PET image, and generate target training data according to the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps.

5. The PET-based epilepsy postoperative prognostic information generation system according to claim 4, characterized in that: The voxel cluster forming unit comprises: a calculation subunit, configured to calculate a normal metabolic intensity mean and a normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images; a single sample test subunit, configured to perform a single sample test on each voxel of each pre-processed preoperative whole-brain PET image according to the normal metabolic intensity mean and the normal metabolic intensity standard deviation, to obtain a single sample test result; The voxel point formation subunit is used to generate a p-value list based on the single-sample test results of each voxel, sort all the p-value lists, and perform false discovery rate correction based on a preset threshold to obtain the voxel points of each pre-processed preoperative whole-brain PET image; The voxel cluster construction subunit is used to select the third preset number of voxel points with the largest differences from the voxel points of each preprocessed preoperative whole-brain PET image as seed points, and extract the fourth preset number of voxel points around each seed point to form a voxel cluster, thereby obtaining the third preset number of voxel clusters.

6. A method for generating postoperative prognostic information of epilepsy based on PET based on the system according to any one of claims 1 to 5, characterized in that: include: Acquiring a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and performing preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging; According to each preprocessed postoperative whole-brain magnetic resonance imaging, the lesion area is located on the corresponding preprocessed preoperative whole-brain PET image, and voxels are extracted; according to all preprocessed healthy whole-brain PET images, a single sample test is performed on each voxel of each preprocessed preoperative whole-brain PET image, and false discovery rate correction is performed to obtain a third preset number of voxel clusters of each preprocessed preoperative whole-brain PET image; according to each voxel cluster and the voxel of the corresponding preprocessed preoperative whole-brain PET image, a third preset number of correlation maps of each preprocessed preoperative whole-brain PET image are obtained; according to the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps, target training data is generated; Obtain multiple pre-selected deep learning models, train each deep learning model according to the target training data, and after the training is completed, select the target deep learning model based on the training effect of each deep learning model; Receive the preoperative whole-brain PET image to be processed and the corresponding postoperative whole-brain magnetic resonance imaging, and generate a third preset number of target correlation maps based on the first preset number of whole-brain healthy PET images, input the third preset number of target correlation maps into the target deep learning model to generate prognostic information.

7. The method for generating postoperative prognostic information of epilepsy based on PET according to claim 6, characterized in that: The method of obtaining a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging, and performing preprocessing to obtain a first preset number of preprocessed healthy whole-brain PET images and a second preset number of preprocessed preoperative whole-brain PET images and preprocessed postoperative whole-brain magnetic resonance imaging, includes: Acquire a first preset number of healthy whole-brain PET images and a second preset number of preoperative whole-brain PET images and corresponding postoperative whole-brain magnetic resonance imaging images; obtaining, using a brain extraction tool, a brain parenchymal extent from each of the whole-brain healthy PET image and the preoperative whole-brain PET image, registering the corresponding brain parenchymal extent to a standard Montreal Neurological Institute space using a linear image registration tool, and then normalizing the intensity of the registered brain parenchymal extent to obtain a first preset number of preprocessed whole-brain healthy PET images and a second preset number of preprocessed preoperative whole-brain PET images; Each postoperative whole-brain MRI was registered to the standard Montreal Neurological Institute space using a linear image registration tool and aligned with the corresponding preoperative whole-brain PET image to obtain a second preset number of preprocessed postoperative whole-brain MRIs.

8. The method for generating postoperative prognostic information of epilepsy based on PET according to claim 6, characterized in that: The method includes: locating the lesion area on the corresponding pre-operative whole-brain PET image after pre-processing based on each pre-processed post-operative whole-brain magnetic resonance imaging, extracting voxels, performing a single sample test on each voxel of each pre-processed whole-brain PET image after pre-processing based on all pre-processed healthy whole-brain PET images, and performing false discovery rate correction to obtain a third preset number of voxel clusters of each pre-processed whole-brain PET image after pre-processing; obtaining a third preset number of correlation maps of each pre-processed whole-brain PET image after pre-processing based on each voxel cluster and the voxel of the corresponding pre-processed whole-brain PET image; and generating target training data based on the category corresponding to each pre-processed whole-brain PET image and the corresponding third preset number of correlation maps. The method includes: Based on each pre-processed post-operative whole-brain MRI, the lesion area was located on the corresponding pre-processed pre-operative whole-brain PET image, and voxels were extracted; Calculating a normal metabolic intensity mean and a normal metabolic intensity standard deviation based on all preprocessed whole-brain healthy PET images, performing a single-sample test on each voxel of each preprocessed preoperative whole-brain PET image based on the normal metabolic intensity mean and the normal metabolic intensity standard deviation, and obtaining voxel points of each preprocessed preoperative whole-brain PET image through false discovery rate correction and a preset threshold, selecting a third preset number of seed points from the voxel points of each preprocessed preoperative whole-brain PET image, and forming a third preset number of voxel clusters accordingly; Calculate the correlation between each voxel cluster and the corresponding whole brain, generate a third preset number of correlation maps for each preprocessed preoperative whole-brain PET image, and generate target training data based on the category corresponding to each preprocessed preoperative whole-brain PET image and the corresponding third preset number of correlation maps.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a PET-based epilepsy postoperative prognostic information generation program stored in the memory and runnable on the processor. When the PET-based epilepsy postoperative prognostic information generation program is executed by the processor, the steps of the PET-based epilepsy postoperative prognostic information generation method according to any one of claims 6 to 8 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a PET-based epilepsy postoperative prognostic information generation program, which, when executed by a processor, implements the steps of the PET-based epilepsy postoperative prognostic information generation method according to any one of claims 6 to 8.

Images