Postoperative symptom prediction method for ventricular metrioma

Abstract

The invention belongs to the technical field of postoperative prediction, and particularly relates to a method for predicting postoperative symptoms of ventricular metrioma, which comprises the following steps: S1, screening out patients diagnosed as ventricular metrioma through postoperative pathological results, carrying out follow-up visit on the patients, screening out death patients irrelevant to the course of disease, and determining the death patients irrelevant to the course of disease; s2, extracting MRI images of the remaining patients, and screening out patients with image missing and unqualified patients; extracting a postoperative MRI follow-up visit patient, and checking whether postoperative spinal cord adhesion exists or not; according to the method, imaging characteristics of the ventricular membrane tumor are identified through the convolutional neural network, deep learning is carried out through data analysis of follow-up visit patients in combination with information and preoperative symptoms of the patients, more and more comprehensive imaging data are provided compared with previous research, and the credibility of the model can be improved; after deep learning is carried out on the model through a large number of MRI images, some factors which are weak in correlation with prognosis are deleted, and therefore the calculated amount is greatly reduced on the premise that the accuracy rate is not obviously reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of postoperative prediction, and specifically relates to a method for predicting postoperative symptoms of spinal cord ependymoma. Background Art

[0002] Spinal cord ependymoma is a rare spinal cord tumor that usually requires surgical resection. Currently, the specific correlation degree between syringomyelia, deposition of hemosiderin, the length of the segments of tumors and edema and the symptoms of ependymoma is unknown, and whether these different imaging manifestations are related to the postoperative function of patients to a certain extent. Due to the high disability rate of intramedullary spinal cord tumors, this technology can guide the bed doctor to have preoperative conversations and facilitate patients to understand the risks of surgery.

[0003] In the prior art, the following methods are usually used to predict the postoperative symptoms of spinal cord ependymoma: Deep learning based on preoperative magnetic resonance (MR) images improves the prediction ability of the survival model of primary spinal cord astrocytoma; Deep learning for the detection and classification of posterior fossa tumors in children; Magnetic resonance image discrimination between filum terminale ependymoma and schwannoma based on convolutional neural network; Establishment and external validation of a radiomics nomogram for differentiating circumscribed astrocytic gliomas and diffuse gliomas based on MRI; Based on the above predictions, preoperative conversations are then carried out, which can, to a certain extent, improve the understanding of the disease by patients and their families, and at the same time can improve the understanding of the characteristics of tumors by the operating surgeon and further refine the surgical technique.

[0004] In the above technologies, most of these models are based on the MRI of brain tumors, and a small number of models are for the prognosis of survival of inflammatory diseases such as multiple sclerosis of the spinal cord or malignant gliomas such as astrocytomas; However, the characteristics of tumor cells in the spinal cord are very different from those of brain tumors; Moreover, the anatomical structure and microenvironment of the spinal cord are special, the functional areas are complex, and the incidence of tumors is low. Therefore, the prediction models for the brain cannot be well matched and applied to the spinal cord; In addition, the existing MRI has disadvantages such as poor recognition accuracy and many artifacts for such fine structures of the spinal cord. Therefore, a more professional model is needed for spinal cord tumors.

[0005] Therefore, the present invention provides a method for predicting postoperative symptoms of spinal cord ependymoma. Summary of the Invention

[0006] In order to make up for the deficiencies of the prior art and solve at least one of the technical problems proposed in the background art.

[0007] The technical solution adopted by the present invention to solve its technical problems is: A method for predicting postoperative symptoms of spinal cord ependymoma according to the present invention includes the following steps:

[0008] S1: Screen out the patients diagnosed with spinal cord ependymoma through postoperative pathological results, follow up these patients, and exclude the patients with lost to follow-up and those whose deaths are clearly unrelated to the disease course;

[0009] S2: Extract the MRI images of the remaining patients, and exclude the patients with missing or unqualified images; Extract the patients with postoperative MRI follow-up, and check for postoperative spinal cord adhesions;

[0010] S3: Standardize the images, and manually draw the ROI on the T2 image by a neurosurgeon;

[0011] S4: Generate neurons a, b, c, and d respectively by analyzing the MRI features and according to the patient information;

[0012] S5: Arrange neurons a, b, c, and d side by side to form a new direct neural network B, denoted as the second-level neurons; Relay the direct neural network B to generate a new direct neural network, denoted as the third-level neurons;

[0013] S6: Use a random function to take 60% of the patient cases as the training set, 20% of the patient cases as the validation set, and 20% of the patient information as the test set, and generate different cohorts multiple times for multiple trainings;

[0014] S7: Automatically read the partial image groups α, β, θ of the training set and the validation set and the preoperative symptom features, MRI features, and pathological diagnoses of the corresponding cases into the model for training and deep learning, and set a random dropout of 10% - 20% to obtain the initial models α, β, θ;

[0015] S8: Input the test set and the corresponding features into the corresponding models α, β, θ respectively, obtain the average of the corresponding model accuracies, which is the prediction accuracy of the model, and select the model with the highest accuracy;

[0016] S9: Delete a single neuron in the direct neural network B in turn to obtain a new model with different neuron nodes removed; Delete the neuron nodes with an influence less than 0.1% to obtain a new model;

[0017] S10: Calculate the accuracy of the new model and save the model, denoted as the postoperative prediction model for mature spinal cord ependymoma.

[0018] Preferably, the method for standardizing the images in S3 includes the following steps:

[0019] Standardization:

[0020] where x is the image matrix, μ is the image mean, σ is the image variance, and N is the number of pixels;

[0021] Normalization:

[0022] where x is the image matrix;

[0023] Balancing: y(x) = x + d(x),

[0024] where d(x) is the deformation field;

[0025] Use ITK-SNAP to perform balancing by combining T2 images and T1+C images for comparison.

[0026] Preferably, the ROI in S3 is the region of interest, and the features of the ROI include tumor entity, cystic change, syringomyelia, spinal cord edema, hemorrhage, and fluid level.

[0027] Preferably, in the T2 image and T1+C image, the largest tumor-exposed image in the 2D images of T2 image and T1+C, the 3D images in T2 image and T1+C, the 2D images in T2 image and T1+C, and the ROI are respectively denoted as three imaging picture recognition feature methods α, β, θ.

[0028] Preferably, the MRI features in S4 include the position and length of the tumor, cavity, and edema, the enhancement pattern of the tumor, the boundary clarity, the hemorrhage condition, and the fluid level condition.

[0029] Preferably, in S4,

[0030] The neuron a is the imaging picture recognition feature generated by constructing a ResNet model using PyTorch, selecting appropriate structures and the number of neuron nodes.

[0031] The neuron b constructs the preoperative symptoms and function scores of the patient, and the time from the appearance of symptoms to seeking medical treatment for surgery into separate several neuron nodes.

[0032] The neuron c constructs the evaluated MRI feature information into separate several neuron nodes.

[0033] The neuron d is the postoperative pathological neuron formed by extracting the WHO and pathological classification of the tumor.

[0034] Preferably, the expression of the second-level neuron in S5:

[0035]

[0036] where x a , x b , x c , x d are the input features of each neuron respectively, and W is the learnable weight;

[0037] Expression of the third - level neuron:

[0038] C(X) = C(x a ,x b ,x c ,x d ) = W·a(x a )·b(x b )·c(x c )·d(x c )。

[0039] Preferably, the method for forming the direct neural network B in S5 includes the following steps:

[0040] A1: Set the data represented by the above neuron nodes: Each sample contains the pre - processed image feature data x a , preoperative symptom factor feature x b , imaging feature data x c , postoperative pathological feature data x d ;

[0041] A2: Set the final result model and the corresponding folder output for the corresponding prediction result accuracy;

[0042] A3: Set different MRI imaging: T1W, T2W, T1 + C;

[0043] A4: Set different slice directions: axial, coronal, transverse;

[0044] A5: Set the number of training channels,

[0045] The expression for training the number of channels using the GPU video memory size is:

[0046]

[0047] where M is the GPU video memory capacity, s is the batch size, and h and w are the height and width of the feature map respectively;

[0048] A6: Set epoch = 100, and the expression for using early stopping during training is:

[0049] L val (t)-L val (t - 1)>Δ threshold

[0050] where, L val (t) is the training loss at time t, and Δ threshold is the set early - stopping threshold. When the above expression is not satisfied continuously for N patience , stop training;

[0051] A7: The downloaded dataset in the folder is used to detect the spinal cord. Before detection, preprocessing is required to ensure a clear spinal cord area.

[0052] A8: Conduct appropriate tests to configure the safety factor. Regarding the clipping of voxels, it is necessary to ensure that the region of interest for detection is clear and complete, and the irrelevant background and noise voxels are clipped and scaled.

[0053] A9: Start training and configure the optimal threshold at the end of training. The expression is:

[0054]

[0055] where Precision is the precision rate and Recall is the recall rate.

[0056] A10: Select an appropriate learning rate to optimize the stable convergence of the model. The following formula is used to select the Adam adaptive learning rate. The expression is:

[0057]

[0058] where is the moving average square of the gradient and is a small constant.

[0059] Preferably, the random dropout in S7 is achieved by setting a dropout layer in the neural network. The expression is:

[0060]

[0061] where the dropout ratio p is between 0.1 and 0.2, and n total is the total number of neurons.

[0062] Preferably, the calculation method for the prediction accuracy of the model in S8:

[0063]

[0064] Then compare a, β, and θ to obtain the model with the highest accuracy.

[0065] The beneficial effects of the present invention are as follows:

[0066] 1. For the postoperative symptom prediction method of spinal cord ependymoma described in the present invention, the imaging features of ependymoma are recognized through a convolutional neural network, combined with the patient's information and preoperative symptoms, and deep learning is carried out through the analysis of the data of the follow-up patients. Compared with previous studies, it has more and more comprehensive imaging data, which can improve the credibility of the model; after deep learning the model with a large number of MRI images, some factors with weak correlation with prognosis are deleted, thereby greatly reducing the calculation amount without significantly reducing the accuracy.

[0067] 2. The postoperative symptom prediction method for spinal cord ependymoma according to the present invention can select more neuron nodes by comprehensively understanding the preoperative information of patients before training, which is convenient for studying the remission weights of different preoperative factors on postoperative symptoms. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The present invention will be further described below with reference to the accompanying drawings.

[0069] Figure 1 is the method flowchart of the postoperative symptom prediction method in the present invention;

[0070] Figure 2 is the method framework diagram for constructing the direct neural network B in the present invention;

[0071] Figure 3 is the method framework diagram for generating a new model in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0072] In order to make the technical means, creative features, achieved purposes and functions of the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.

[0073] As Figures 1 to 3 shown, a postoperative symptom prediction method for spinal cord ependymoma according to an embodiment of the present invention includes the following steps:

[0074] S1: Screen out the patients diagnosed as spinal cord ependyma through postoperative pathological results, follow up these patients, and screen out the lost-to-follow-up and clearly death patients unrelated to the disease course;

[0075] S2: Extract the MRI images of the remaining patients, and screen out the patients with missing or unqualified images; extract the patients with postoperative MRI follow-up, and check for postoperative spinal cord adhesions;

[0076] S3: Standardize the images, and manually draw the ROI on the T2 image by a neurosurgeon;

[0077] S4: Generate neurons a, b, c, and d respectively by analyzing the MRI features and according to the information of the patients;

[0078] S5: Arrange neurons a, b, c, and d side by side to form a new direct neural network B, denoted as the second-level neurons; relay the direct neural network B to generate a new direct neural network, denoted as the third-level neurons;

[0079] S6: Use a random function to take 60% of the patient cases as the training set, 20% of the patient cases as the validation set, and 20% of the patient information as the test set, and generate different queues multiple times for multiple trainings;

[0080] S7: Automatically read partial image groups α, β, and θ of the training set and the validation set, as well as the preoperative symptom characteristics, MRI characteristics, and pathological diagnoses of the corresponding cases, into the model for training and deep learning, and set a random dropout of 10% - 20% to obtain the initial models α, β, and θ;

[0081] S8: Input the test set and the corresponding features into the corresponding models α, β, and θ respectively, obtain the average of the accuracy rates of the corresponding models, which is the prediction accuracy rate of the model, and select the model with the highest accuracy rate;

[0082] S9: Sequentially delete individual neurons in the direct neural network B to obtain new models with different neuron nodes removed; delete neuron nodes with an influence less than 0.1% to obtain new models;

[0083] S10: Calculate the accuracy rate of the new model and save the model, denoted as the postoperative prediction model for mature spinal cord ependymoma.

[0084] When the prediction method provided by the present invention is used, first, patients diagnosed as spinal ependymoma through postoperative pathological results are screened out, and these patients are followed up, and patients with lost follow-up and deaths clearly unrelated to the disease course are excluded; then the MRI images of the remaining patients are extracted, patients with missing or unqualified images are excluded, and at the same time, patients who have undergone postoperative MRI follow-up are extracted to check for postoperative spinal cord adhesion; by standardizing the images, neurosurgeons manually draw the ROI on the T2 image, including features such as tumor entity, cystic change, syringomyelia, spinal cord edema, hemorrhage, fluid level, etc.; then analyze MRI features including the location and length of the tumor, cavity and edema, the enhancement pattern of the tumor, whether the boundary is clear, whether there is hemorrhage, whether there is fluid level, etc. and record them in a table; by analyzing the MRI features, neurons a, b, c, and d are generated respectively according to the patient's information; neurons a, b, c, and d are juxtaposed to form a new direct neural network B, denoted as the second-level neuron; the direct neural network B is relayed to generate a new direct neural network, denoted as the third-level neuron; finally, the model is trained. By using a random function, 60% of the patient cases are used as the training set, 20% of the patient cases are used as the validation set, and 20% of the patient information is used as the test set, and different cohorts are generated multiple times for multiple trainings; then partial image groups α, β, θ of the training set and the validation set are automatically read into the model together with the preoperative symptom features, MRI features, and pathological diagnoses of the corresponding cases for training and deep learning, and 10% - 20% is randomly set to be lost to prevent overfitting of the model, and the initial models α, β, θ are obtained; the remaining test groups and corresponding features are respectively input into the corresponding models α, β, θ, and the average of the corresponding model accuracies is obtained, which is the prediction accuracy of the model, and the model with the highest accuracy is selected; single neurons in the direct neural network B are sequentially deleted to obtain models with different neuron nodes removed; factors that reduce the accuracy by 0.1% are deleted because removing these factors can significantly reduce the preliminary preparation and the computational amount of the model without significantly reducing the prediction accuracy; finally, the model is saved and denoted as the postoperative prediction model for mature spinal ependymoma.

[0085] In summary, by using a convolutional neural network to identify the imaging features of ependymoma, combining the patient's information and preoperative symptoms, and performing deep learning through the data analysis of the followed-up patients, there are more and more comprehensive imaging data compared with previous studies, which can improve the credibility of the model; and the preoperative information of the patients is more comprehensively understood before training, and more neuron nodes can be selected to facilitate the study of the remission weights of different preoperative factors on postoperative symptoms; after deep learning of the model through a large number of MRI images, some factors weakly related to the prognosis are deleted, thereby significantly reducing the computational amount without significantly reducing the accuracy.

[0086] Such as Figure 1 and Figure 2As shown, the method for normalizing the image in S3 includes the following steps:

[0087] Normalization:

[0088] where x is the image matrix, μ is the image mean, σ is the image variance, and N is the number of pixels;

[0089] Normalization:

[0090] where x is the image matrix;

[0091] Trimming: y(x) = x + d(x),

[0092] where d(x) is the deformation field;

[0093] Use ITK-SNAP to perform trimming by combining the T2 image and the T1+C image for comparison.

[0094] As Figure 1 and Figure 2 shown, the ROI in S3 is the region of interest, and the features of the ROI include tumor entity, cystic degeneration, syringomyelia, spinal cord edema, hemorrhage, and fluid level;

[0095] In the T2 image and the T1+C image, the largest tumor-exposed image in the 2D images of the T2 image and the T1+C, the 3D images in the T2 image and the T1+C, the 2D images in the T2 image and the T1+C, and the ROI are respectively denoted as three radiographic image recognition feature methods α, β, and θ.

[0096] As Figure 1 and Figure 2 shown, the MRI features in S4 include the location and length of the tumor, cavity, and edema, the enhancement pattern of the tumor, the boundary clarity, the hemorrhage condition, and the fluid level condition;

[0097] In S4,

[0098] the neuron a is a radiographic image recognition feature generated by constructing a ResNet model using PyTorch and selecting appropriate structures and the number of neuron nodes;

[0099] the neuron b constructs the preoperative symptoms and function scores of the patient and the time from the appearance of symptoms to surgical treatment as separate several neuron nodes;

[0100] the neuron c constructs the evaluated MRI feature information as separate several neuron nodes;

[0101] the neuron d is a postoperative pathological neuron formed by extracting the WHO and pathological types of the tumor.

[0102] When the neurons a, b, c, and d provided by the present invention are used, a ResNet model is constructed by adopting PyTorch, appropriate structures and the number of neuron nodes are selected, and finally a certain number of imaging picture recognition feature neurons a are generated. The neurons of the neural network constructed by ResNet are used to recognize image features;

[0103] Among them, ResNet is composed of an input layer, multiple residual layers, and an output layer.

[0104] The implementation structure of the residual layer is: H(x) = ReLU(BN(Conv(x))) + x,

[0105] where x is the input image, Conv is the convolution operation, BN is the batch normalization, and ReLU is the non-linear activation function;

[0106] The output layer implementation is: Output(x) = Softmax(W·GlobalAvgpool(H(x))),

[0107] where GlobalAvgPool is the global average pooling function, W is the output layer weight, and Softmax is the normalized output function.

[0108] The scores of the patient's preoperative symptoms and functions, and the time from the appearance of symptoms to seeking surgical treatment in the hospital are constructed into several separate neuron nodes, denoted as preoperative symptom factor neuron b;

[0109] The MRI feature information is evaluated, and the location and length of the tumor, cavity, and edema (C / T / L / Sx + ymm), the enhancement pattern of the tumor (homogeneous / heterogeneous / nodular / ring enhancement), clear boundary (yes / no), bleeding (yes / no), and fluid level (yes / no) are constructed into several separate neuron nodes, denoted as imaging feature neuron c;

[0110] Factors such as the WHO and pathological classification of the tumor are denoted as postoperative pathology neuron d.

[0111] Such as Figure 1 and Figure 3 As shown, the expression of the second-level neuron in S5:

[0112]

[0113] where x a , x b , x c , x d are the input features of each neuron respectively, and W is the learnable weight;

[0114] The expression of the third-level neuron:

[0115] C(X) = C(x a , x b , x c , x d ) = W·a(x a )·b(x b )·c(x c )·d(x c )。

[0116] When the second-level neurons and third-level neurons provided by the present invention are used, since the model can guide preoperative conversations, and the pathological examination of tumors occurs after surgery, two unique models can be generated, namely, a preoperative model without d and a postoperative model including d; adding pathological factors during the learning process may improve the accuracy of model prediction, and the postoperative model can predict the long-term prognosis of patients when they are discharged, further guiding the lives of patients; when forming the new direct neural network B and the direct neural network, it includes two models of neurons a, b, and c, and neurons a, b, c, and d.

[0117] As Figure 1 and Figure 3 shown, the method for forming the direct neural network B in S5 includes the following steps:

[0118] A1: Set the data represented by the above neuron nodes: Each sample contains the image feature data x of the corresponding subject after preprocessing a , preoperative symptom factor features x b , imaging feature data x c , postoperative pathological feature data x d ;

[0119] A2: Set the final result model and the corresponding folder output for the corresponding prediction result accuracy;

[0120] A3: Set different MRI imaging: T1W, T2W, T1+C;

[0121] A4: Set different slice directions: axial, coronal, transverse;

[0122] A5: Set the number of training channels,

[0123] Use the expression of the number of training channels for the GPU video memory size:

[0124]

[0125] where M is the GPU video memory capacity, s is the batch size, and h and w are the height and width of the feature map respectively;

[0126] A6: Set epoch = 100. The expression for using early stopping during training is:

[0127] L val (t) - L val (t - 1)>Δ threshold

[0128] where L val (t) is the training loss at time t, and Δ threshold is the set early stopping threshold. When the above expression is not satisfied for consecutive N patience times, training is stopped;

[0129] A7: The downloaded dataset in the folder is used to detect the spinal cord. Preprocessing is required before detection to ensure a clear spinal cord region;

[0130] A8: Conduct appropriate tests to configure the safety factor. Regarding the clipping of voxels, ensure that the region of interest for detection is clear and complete, and clip and scale irrelevant background and noise voxels;

[0131] A9: Start training and configure the optimal threshold at the end of training. The expression is:

[0132]

[0133] where Precision is the precision rate and Recall is the recall rate;

[0134] A10: Select an appropriate learning rate to optimize the stable convergence of the model. The following formula is used to select the Adam adaptive learning rate. The expression is:

[0135]

[0136] where is the moving average square of the gradient and is a small constant.

[0137] As Figure 1 and Figure 3 shown, the random dropout in S7 is achieved by setting a dropout layer in the neural network. The expression is:

[0138]

[0139] where the dropout ratio p is between 0.1 and 0.2, and n total is the total number of neurons.

[0140] When the deep learning provided by the present invention is in use, by automatically reading partial image groups α, β, and θ of the training set and the validation set and the preoperative symptom characteristics, MRI characteristics, and pathological diagnoses of the corresponding cases into the model for training and deep learning, and setting a function to randomly delete 10% - 20% of the missing neurons to prevent overfitting of the model.

[0141] As Figure 1 and Figure 3 shown, the calculation method of the prediction accuracy of the model in S8:

[0142]

[0143] Then compare a, β, and θ to obtain the model with the highest accuracy.

[0144] Working principle: First, patients diagnosed with spinal cord ependymoma through postoperative pathological results are screened out, and these patients are followed up. Patients with lost to follow-up and deaths clearly unrelated to the disease course are excluded; then, the MRI images of the remaining patients are extracted, and patients with missing or unqualified images are excluded. At the same time, patients who undergo postoperative MRI follow-up are extracted to check for postoperative spinal cord adhesions; by standardizing the images, neurosurgeons manually outline the ROI on the T2 image, including features such as tumor entity, cystic change, syringomyelia, spinal cord edema, hemorrhage, and fluid level; then, MRI features are analyzed, including the location and length of the tumor, cavity, and edema, the enhancement pattern of the tumor, whether the boundary is clear, whether there is hemorrhage, whether there is fluid level, etc., and are recorded in a table; by analyzing the MRI features, neurons a, b, c, and d are generated respectively according to the patient's information; neurons a, b, c, and d are arranged side by side to form a new direct neural network B, denoted as the second-level neuron; the direct neural network B is relayed to generate a new direct neural network, denoted as the third-level neuron; finally, the model is trained. By using a random function, 60% of the patient cases are used as the training set, 20% of the patient cases are used as the validation set, and 20% of the patient information is used as the test set, and different cohorts are generated multiple times for multiple trainings; then, partial image groups α, β, θ of the training set and the validation set are automatically read into the model together with the preoperative symptom features, MRI features, and pathological diagnoses of the corresponding cases for training and deep learning, and a random dropout of 10% - 20% is set to prevent overfitting of the model, and the initial models α, β, θ are obtained; the remaining test groups and corresponding features are input into the corresponding models α, β, θ respectively, and the average of the corresponding model accuracies is obtained, which is the prediction accuracy of the model, and the model with the highest accuracy is selected; single neurons in the direct neural network B are deleted in turn to obtain models with different neuron nodes removed; factors that reduce the accuracy by 0.1% are deleted because removing these factors significantly reduces the amount of preparatory work and model calculations without significantly reducing the prediction accuracy; finally, the model is saved and denoted as the postoperative prediction model for mature spinal cord ependymoma.

[0145] In summary, by identifying the imaging features of ependymoma through a convolutional neural network, combining the patient's information and preoperative symptoms, and performing deep learning through the data analysis of the followed-up patients, there are more and more comprehensive imaging data compared with previous studies, which can improve the credibility of the model; and a more comprehensive understanding of the patient's preoperative information before training can select more neuron nodes, facilitating the study of the remission weights of different preoperative factors on postoperative symptoms; after deep learning the model with a large number of MRI images, some factors weakly related to the prognosis are deleted, thus significantly reducing the amount of calculation without significantly reducing the accuracy.

[0146] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting postoperative symptoms of spinal ependymoma, characterized in that: The following steps are involved: S1: Screen patients diagnosed with ventricular membrane in the spinal cord by postoperative pathological results, follow up these patients, and screen out patients who were lost to follow-up and patients whose deaths were clearly unrelated to the course of the disease; S2: Extract the MRI images of the remaining patients and screen out patients with missing images and unqualified patients; extract patients with postoperative MRI follow-up to check for postoperative spinal cord adhesion; S3: The images were standardized, and the ROI was manually outlined on the T2 images by a neurosurgeon; S4: By analyzing the MRI features and according to the patient's information, neuron a, neuron b, neuron c and neuron d are generated respectively; S5: neurons a, b, c and d are arranged in parallel to form a new direct neural network B, which is recorded as the second-level neurons; the direct neural network B is relayed to generate a new direct neural network, which is recorded as the third-level neurons; S6: Use a random function to select 60% of the patient cases as the training set, 20% of the patient cases as the validation set, and 20% of the patient information as the test set, and generate different cohorts multiple times for multiple training; S7: Part of the image groups α, β, and θ of the training set and validation set and the preoperative symptom characteristics, MRI characteristics, and pathological diagnosis of the corresponding cases are automatically read into the model for training and deep learning, and a random loss of 10% to 20% is set to obtain the initial models α, β, and θ; S8: Input the test set and the corresponding features into the corresponding models α, β, and θ respectively, and obtain the average accuracy of the corresponding models, which is the prediction accuracy of the model, and select the model with the highest accuracy; S9: Delete individual neurons in the direct neural network B in sequence to obtain a new model by removing different neuron nodes; delete neuron nodes with an influence of less than 0.1% to obtain a new model; S10: Calculate the accuracy of the new model and save the model as a mature postoperative prediction model for spinal ependymoma.

2. The method for predicting postoperative symptoms of spinal cord ependymoma according to claim 1, characterized in that: The method for normalizing an image in S3 comprises the following steps: standardization: Where x is the image matrix, μ is the image mean, σ is the image variance, and N is the number of pixels; Normalization: Where x is the image matrix; Balancing: y(x) = x + d(x), where d(x) is the deformation field; ITK-SNAP was used to compare the T2 image with the T1+C image for balancing.

3. The method for predicting postoperative symptoms of spinal cord ependymoma according to claim 1, characterized in that: The ROI in S3 is a region of interest, and the features of the ROI include tumor entity, cystic change, syringomyelia, spinal cord edema, hemorrhage and fluid level.

4. The method for predicting postoperative symptoms of spinal ependymoma according to claim 2, characterized in that: In the T2 image and T1+C image, the largest tumor exposure image in the 2D image of T2 image and T1+C, the 3D image in T2 image and T1+C, the 2D image of T2 image and T1+C, and ROI are respectively recorded as three imaging picture recognition feature modes α, β, and θ.

5. The method for predicting postoperative symptoms of spinal cord ependymoma according to claim 1, characterized in that: The MRI features in S4 include the location and length of the tumor, cavity and edema, the enhancement pattern of the tumor, the boundary clarity, the bleeding situation and the fluid level situation.

6. The method for predicting postoperative symptoms of spinal ependymoma according to claim 1, characterized in that: In said S4, The neuron a is constructed by using PyTorch to build a ResNet model, selecting a suitable structure and number of neuron nodes, and generating imaging image recognition features; The neuron b constructs the patient's preoperative symptom and function scores, and the time from symptom onset to surgical treatment into several separate neuron nodes; The neuron c constructs the evaluation MRI feature information into several separate neuron nodes; The neuron d is a postoperative pathological neuron formed by extracting the WHO and pathological typing of the tumor.

7. The method for predicting postoperative symptoms of spinal cord ependymoma according to claim 1, characterized in that: The expression of the second-order neurons in S5 is: where x a ,x b ,x c ,x d are the input features of each neuron, and W is the learnable weight; The expression of the third-order neuron: C(X)=C(x a ,x b ,x c ,x d )=W·a(x a )·b(x b )·c(x c )·d(x c )。 8. The method for predicting postoperative symptoms of spinal ependymoma according to claim 1, characterized in that: The method for forming the direct neural network B in S5 comprises the following steps: A1: Set the data represented by the above neuron nodes: Each sample contains the preprocessed image feature data x of the corresponding subject. a , Preoperative symptom factor characteristics x b , imaging feature data x c , postoperative pathological characteristics data x d ; A2: Set the final result model and the corresponding folder output of the prediction result accuracy; A3: Set up different MRI imaging: T1W, T2W, T1+C; A4: Set different slice directions: axial, coronal, transverse; A5: Set the number of training channels. The expression for the number of training channels using GPU memory size is: Where M is the GPU memory capacity, s is the batch size, h and w are the height and width of the feature map respectively; A6: Set epoch=100 and use the early stopping method during training as follows: L val (t)-L val (t-1)>Δ threshold Among them, L val (t) is the training loss at time t, Δ threshold is the early stop threshold set. patience If the above expression is not satisfied, the training is stopped; A7: The downloaded dataset in the folder is used to detect the spinal cord. It needs to be preprocessed before detection to ensure a clear spinal cord area. A8: Perform appropriate tests to configure the safety factor. Regarding voxel clipping, ensure that the region of interest to be detected is clear and complete, and that irrelevant background and noise voxels are clipped and scaled; A9: Start training and configure the optimal threshold at the end of training. The expression is: Where Precision is the precision rate and Recall is the recall rate; A10: Select an appropriate learning rate to optimize the model for stable convergence. The following formula is selected using the Adam adaptive learning rate: in is the moving average square of the gradient, which is a small constant.

9. The method for predicting postoperative symptoms of spinal ependymoma according to claim 1, characterized in that: The random loss in S7 is achieved by setting a dropout layer in the neural network, and the expression is: Among them, the loss ratio p is between 0.1 and 0.2, n total is the total number of neurons.

10. The method for predicting postoperative symptoms of spinal ependymoma according to claim 1, characterized in that: The calculation method of the prediction accuracy of the model in S8 is: Then compare a, β and θ to get the model with the highest accuracy.

Images