A glioma image intelligent diagnosis method based on geometric optimization uncertainty correction

Abstract

This invention proposes an intelligent diagnostic method for glioma images based on geometric optimization uncertainty correction. This invention combines truth discovery, improved Veno diagrams, geometric optimization, and deep learning to achieve intelligent diagnosis of glioma images and uncertainty correction optimization of deep neural network models, effectively assisting clinical treatment. The invention includes the following steps: Step 1: Collect several glioma images, extract features from them using a deep convolutional neural network, and classify the extracted glioma image features; Step 2: Consider the joint influence of multiple glioma images, improve the Veno diagram algorithm, and perform category uncertainty correction on the glioma images based on the improved Veno diagram algorithm; Step 3: Post-correct the glioma image classification model based on the geometric optimization algorithm to improve the problem of decreased accuracy of the correction model in glioma images; Step 4: Obtain the final category information for each glioma medical image, realizing intelligent diagnosis of glioma images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a glioma image intelligent diagnosis method based on geometric optimization uncertainty correction. BACKGROUND

[0002] Glioma is a cancer of the brain and spinal cord glial cells, which is derived from the neural system glial cells and is the most common intracranial primary tumor, accounting for 40% to 50% of intracranial malignant tumors. The World Health Organization (WHO) classifies glioma into four levels, and the malignancy increases with the level. From the classification method in the fifth edition of the World Health Organization Classification of Tumors of the Central Nervous System, it can be seen that molecular diagnosis is gradually adopted by doctors in the current treatment of glioma. IDH (isocitrate dehydrogenase) is a common molecular pathological marker in glioma. IDH mutation positive indicates that glioma is at least WHO grade II. Therefore, accurate IDH genotyping is of great significance for individualized treatment and clinical prognosis evaluation of glioma.

[0003] Medical image classification is one of the important means in medical image processing. Through medical image classification, the disease category to which each picture belongs is obtained, medical personnel can quickly and efficiently and accurately perform clinical disease diagnosis, and further improve the medical efficiency. In the field of medical image analysis, one of the earliest applications of deep learning is image classification. Medical image classification refers to taking medical images as input, and using a relatively best trained model to predict whether a disease is present or the severity of the disease is classified as output. The earliest deep learning model used for image classification tasks is a self-encoding algorithm (SAE), a deep belief network (DBN), and other unsupervised learning networks. From the early AlexNet network, VGGNet network, GoogLeNet network, ResNet network, DenseNet network to the current top-ranked Swim Transform network, all have left a deep and colorful stroke in the field of medical image classification. At present, the convolutional neural network (CNN) is very popular in the medical field, and its medical applications are very extensive. However, the current deep learning model has non-interpretable parameters and randomization, and the model prediction has uncertainty.

[0004] Medical image processing is of great significance for medical auxiliary diagnosis. However, with the development of science and technology, medical imaging equipment and technology are constantly updated, and artificial intelligence AI (Artificial Intelligence) is leading the development of the medical field. The traditional manual classification method cannot meet the increasing demand of medical images, and cannot obtain the subtle information in the medical images. In recent years, the development of artificial intelligence has helped us find a suitable way to solve this problem. Deep learning is a new field in machine learning, which can continuously obtain the characteristics of things by simulating the neural network of the human brain, and can fully describe the information expressed by the data.

[0005] With the development of science and technology, the application of deep learning in daily life is more and more. Although the current deep neural network has excellent performance, the uncertainty related to the learning process cannot be ignored. In the process of calibrating the deep learning model in the deep convolutional neural network, the calibration is often not good, resulting in a calibration effect that does not conform to the result. Therefore, in this patent, the model is corrected after the geometric optimization algorithm, and the result of the model is corrected.

[0006] In addition, in the traditional data preprocessing, only the labeled medical image dataset can be cropped, rotated, scaled, flipped and translated on the original medical image data to achieve the effect of data set expansion. However, medical image dataset often has labeled and unlabeled data. The labeled dataset can use traditional data augmentation method, and the unlabeled dataset is often not used due to no label. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art and design a glioma image intelligent diagnosis method based on geometric optimization uncertainty correction. The method improves K-means clustering through pre-training model, finds true value uncertainty correction, and adds pseudo label, solves the problem of too few medical image datasets, and realizes the classification of medical image of glioma molecular diagnosis, thereby assisting clinical diagnosis.

[0008] A glioma image intelligent diagnosis method based on geometric optimization uncertainty correction, comprising the following steps:

[0009] Step one: collect a plurality of glioma images, extract features of the glioma images through a deep convolutional neural network, and classify the extracted glioma image features;

[0010] Step two: considering the joint influence of multiple glioma images, improving the Venn diagram algorithm, and correcting the class uncertainty of the glioma images based on the improved Venn diagram algorithm;

[0011] Step three: based on the geometric optimization algorithm, the glioma image classification model is corrected to improve the accuracy of the correction model in the glioma image.

[0012] Step four: obtain the final category information of each glioma medical image, and realize intelligent diagnosis of glioma image.

[0013] Further, in step one:

[0014] Firstly, the size of the glioma image is adjusted to the input layer size of the model, and then input into the pre-trained deep convolutional neural network model to learn and train the glioma image data, and the image features are extracted to improve the classification accuracy and generalization ability of the model to the glioma image data.

[0015] Then, the extracted glioma image features are clustered to obtain the category information of the classified glioma images.

[0016] Finally, the true value discovery algorithm is used to correct the uncertainty of the glioma image category information obtained by clustering, so as to achieve the purpose of data augmentation, reducing experimental error and improving the generalization ability of the model.

[0017] Preferably, the improved K-means algorithm is used to cluster the extracted glioma image features, which effectively considers the influence of prior knowledge of medical images on sample classification, solves the problem that the traditional K-means algorithm is greatly affected by randomness based on Euclidean distance calculation of sample category distance, and improves the accuracy of glioma image feature clustering by weighting calculation of sample category distance based on prior knowledge.

[0018] Further, the improved K-means algorithm for clustering the extracted glioma image features includes the following steps:

[0019] S01: input the glioma image features extracted by the pre-trained convolutional neural network, and assign weights to the features corresponding to each glioma image according to the prior knowledge;

[0020] S02: according to the prior knowledge, select one of the most representative glioma images in each category as the focus;

[0021] S03: calculate the distance between each weighted glioma image sample and the focus, and classify the glioma image into the nearest focus category;

[0022] S04: repeat the operation until all glioma images are classified.

[0023] Preferably, the distance between each weighted glioma image sample and the focus in S03 is calculated according to the following formula:

[0024]

[0025] Wherein, D represents the distance between the glioma sample to be classified and the cluster center of the glioma image cluster, x and y are the coordinates of the cluster center of the glioma image cluster respectively, F is the weight of the cluster center, F (i) is the weight based on prior knowledge of the glioma sample to be classified, x i , y i are the coordinates of the glioma image to be classified respectively.

[0026] Preferably, step one uses fruit fly algorithm to update the classification result weight of the deep convolutional neural network classification result, so that the result is more accurate.

[0027] Further, the specific operation steps of step two are:

[0028] S21: considering the joint influence of multiple glioma images, the space is divided into several cells, and the glioma image feature set extracted by the deep convolutional neural network is input;

[0029] S22: constructing glioma image feature point cluster, glioma image feature box tree, and glioma image feature point cluster after division;

[0030] S23: judging whether the feature box tree is empty, if the box tree is not empty, the first group of glioma feature point clusters after division is valued, if the box tree is empty, the feature box tree is constructed, and the valuation operation is repeated;

[0031] S24: judging whether the glioma image is two-classification or multi-classification, if the glioma image classification is two-classification, the second group of glioma image feature point clusters after division is error evaluated, and the valuation operation is performed; if the glioma image classification is multi-classification, the glioma image feature point clusters are sorted, and the feature point cluster division is re-performed, and the valuation operation is repeated.

[0032] Further, in order to keep the error controllable, the following operations are taken:

[0033] Firstly, input a group of points P, error error rate β>0 and box tree structure data, secondly, construct feature box tree for point P, then construct the nodes of feature box tree, and finally output the simplified feature box tree;

[0034] Obtain a subset of glioma image data set, then construct feature box tree, and finally perform species division, i.e. the final glioma image belongs to which category;

[0035] For glioma image unit p∈P, and any second category of glioma image unit c, let D(c) be the diameter of unit c, and r be the shortest distance from unit c to p, then,

[0036]

[0037] For p∈P, and the first type of unit c with distance node v, then for q∈c and p'∈P\P v Then satisfies,

[0038]

[0039] Where, P\P v Indicates the known subset P v of P, l(v) indicates the input point with distance node v, P(|P|) indicates the value obtained by the point cluster P in the multinomial bounded function.

[0040] Further, step three includes the following process:

[0041] S31: input the high-dimensional vector data d of the glioma image obtained by the improved Vino graph algorithm and the integrated data d of the glioma image ens ;

[0042] S32: obtain the index value I of the maximum data in the current integrated data d ens and obtain the index set S of the sorting value in , wherein d i indicates i high-dimensional vector data, and L indicates the number of glioma images;

[0043] S33: if the first value of the index value set S of the sorting value is equal to the index value I of the maximum data in the current integrated data d ens , return the high-dimensional vector data d.

[0044] S34: if the first value of the index value set S of the sorting value is not equal to the index value I of the maximum data in the current integrated data d ens , calculate the corrected high-dimensional vector data d c .

[0045] S35: if the corrected high-dimensional vector data d c is greater than the high-dimensional vector data d, update the corrected high-dimensional vector data d c .

[0046] S36: repeat S34, S35 until all high-dimensional vectors are updated.

[0047] Preferably, the geometric optimization precision maintaining formula is:

[0048]

[0049]

[0050]

[0051] where d * represents the real probability vector, ω s represents the credibility, s represents the jth classifier model, S represents the number of classifiers, l represents the dimension of the high-dimensional vector, and the constraint specification introduces the subspace Ω: d l > d m , The above formula is used to realize that the real probability vector d * falls in Δ α = Δ L∩Ω , wherein Δ α represents the accuracy maintaining area, and Δ L∩Ω represents the intersection area of the original probability and the predicted probability. If the real probability vector d * falls outside Δ α , the projection value of the real probability vector d * in Δ α is found to pull the real probability vector d α outside Δ * back to Δ α , so that the accuracy is not affected.

[0052] Compared with the prior art, the present application has the following beneficial effects:

[0053] 1. The present application designs a data augmentation method of improved K-means clustering of pre-trained models based on true value discovery. The improved K-means clustering is performed on the unlabeled data set, the pseudo label is added, and the final pseudo label of the unlabeled data image is obtained based on the uncertainty correction of the true value discovery algorithm, thereby solving the problem of insufficient medical image data set. Through experimental simulation, it is shown that the algorithm has certain improvement in experimental accuracy and experimental effect. The improved K-means algorithm is based on prior knowledge weighting to calculate the sample category distance, effectively considers the influence of medical image prior knowledge on sample classification, selects the cluster center based on prior knowledge, effectively considers the influence of glioma image cluster center selection on sample classification, while the original K-means algorithm is based on Euclidean distance to calculate the sample category distance, and selects the cluster center based on the randomization method.

[0054] 2. The improved Venn diagram is used, the input point is a given point set, the Venn diagram is constructed by the joint influence of all Venn point pairs on any Venn point Q, the input point in the traditional Venn diagram is a given point or object, and the Venn diagram is constructed by calculating the distance between a single Venn point and any Venn point Q, the problems of space division and space allocation in the traditional Venn diagram are solved, and in the traditional deep convolutional neural network, a high-dimensional vector is mapped to a low-dimensional vector to obtain the classification category corresponding to the input.

[0055] 3. The fly algorithm is applied to the glioma image true / false label, the weight of the glioma image true / false label based on the deep convolutional neural network is updated, the optimal weight value of the false label is obtained, and the actual influence of the glioma image true / false label on the experimental precision is improved.

[0056] 4. The model post-calibration based on the geometric optimization algorithm is realized, the glioma image data calibrated based on the improved Venn diagram is calibrated after the deep convolutional neural network classifier, and the problem that the traditional experimental calibration result is limited is solved. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a true value discovery algorithm flowchart of the present application;

[0058] Figure 2 is a fly optimization algorithm flowchart of the present application;

[0059] Figure 3 is an improved Venn diagram algorithm flowchart of the present application;

[0060] Figure 4 is a geometric optimization algorithm flowchart of the present application;

[0061] Figure 5 is a glioma image intelligent diagnosis flowchart based on geometric optimization uncertainty correction of the present application. DETAILED DESCRIPTION

[0062] The glioma image intelligent diagnosis method based on geometric optimization uncertainty correction of the present application will be further described in detail in combination with the drawings and specific implementation methods.

[0063] Since there are defects in using the traditional clustering algorithm to cluster medical images: 1) when randomly selecting a clustering center point, randomness often affects the final clustering effect; 2) prior knowledge has no effect on the data set. In view of the above problems, the improved K-means algorithm is used, which better solves the problems of random clustering points and slow convergence leading to poor clustering effect.

[0064] Because the traditional network cannot distinguish the image brought by true / false label due to the uniform weight assignment for image classification, the weight of each glioma image is initialized, the fruit fly algorithm used in the foregoing is used to update the weight of the clustering result of the medical image, the number of clustering centers is represented by using variable quantization, the fixed step in the fruit fly algorithm is replaced by using random step, the optimal odor concentration value is obtained by iteratively calculating the image entropy, the optimal selection of the cut-off distance and the clustering center is optimized, the optimal weight value is obtained by training the network, and the true / false label glioma images are divided. Each real label image is assigned a high weight, and each false label image is assigned a low weight, so that the medical image dataset is better utilized.

[0065] In the traditional Venn diagram, a plurality of points on a plane are divided according to the nearest neighbor principle, each point is associated with its nearest neighbor region, and the input is often a given point or object. The Venn region is divided by calculating the distance between a single Venn point or any Venn point and other points or objects using Euclidean distance or Hausdorff distance. Each Venn region is composed of samples belonging to the same type. However, the traditional Venn diagram has the problems of uneven space division and inconsistent space allocation. To solve the above problems, an improved Venn diagram algorithm is used for optimization. The improved Venn diagram algorithm uses point clusters instead of single points for input; the glioma image joint influence is used instead of the traditional distance to divide the Venn region. The improved Venn diagram algorithm inputs the processed glioma image obtained by true value discovery calibration and optimal weight assignment using fruit fly algorithm into the deep convolutional neural network, constructs an approximate influence Venn diagram, judges whether the feature box tree state is empty, then judges whether the glioma image belongs to class 0 or 1, and finally outputs the category information corresponding to the glioma image.

[0066] Because the model has uncertainty in the learning process of the traditional algorithm, and there are defects in the calibration of the classifier class: the calibration effect does not conform to the result. To solve the above problems, the improved K-means algorithm is used to calibrate the image class after clustering; the geometric optimization algorithm is used to calibrate the glioma image data after the improved Venn diagram calibration of the deep convolutional neural network classifier. First, all glioma images are input for pre-clustering, the label information obtained after clustering is assigned to the unlabeled glioma images to obtain pseudo-labeled glioma images, the pseudo-labeled dataset is calibrated using the true value discovery algorithm to achieve the effect of pseudo-label calibration and improve the experimental precision; then the data calibrated by the improved Venn diagram algorithm is input into the geometric optimization algorithm, so that the model post-calibration method keeps the model accuracy unaffected, and the problem of model uncertainty is better solved by using the algorithm.

[0067] Specifically, as Figures 1-5As shown, the present application proposes a glioma image intelligent diagnosis method based on geometric optimization uncertainty correction, comprising the following steps:

[0068] Step one: collect a number of glioma images, extract features from them through deep convolutional neural network, and classify the extracted glioma image features;

[0069] Step two: consider the joint influence of multiple glioma images, improve the Venn diagram algorithm, and correct the class uncertainty of glioma images based on the improved Venn diagram algorithm;

[0070] Step three: correct the glioma image classification model based on the geometric optimization algorithm, and improve the problem of accuracy reduction of the correction model in glioma image;

[0071] Step four: get the final class information of each glioma medical image, and realize intelligent diagnosis of glioma image.

[0072] In step one, first, use the improved K-means clustering algorithm to pre-cluster all glioma images, solve the problem of small sample size in the original medical image set, and label the unlabeled glioma image dataset to obtain the pseudo-labeled glioma image dataset. However, since the pseudo-labels may be incorrect, they need to be corrected using a true value discovery algorithm to obtain the final pseudo-labeled dataset, so as to achieve the purposes of data augmentation, reducing experimental error and improving model generalization ability. Figure 1 As shown, first input the glioma class information obtained by the improved K-means algorithm, initialize the true value vector, assign the confidence and uncertainty true value vector, start iterative search, and if it is within the iteration times, output the true value vector; if it is not within the iteration times, alternately update the confidence and uncertainty true value vector, and iterate continuously until all class information, i.e. label information, is corrected by the true value discovery algorithm. Figure 1 is the class correction of the clustered dataset. The image format of the glioma medical image is DICOM; the mutation state is divided into 2 categories.

[0073] The specific operation steps of step one are as follows:

[0074] First, adjust the size of the glioma image to the input layer size of the model, and input it into the pre-trained deep convolutional neural network model for feature extraction;

[0075] Second, use the improved K-means algorithm to cluster the extracted image features to obtain the label of each glioma image;

[0076] S11, input the glioma image features extracted by the pre-trained convolutional neural network, and assign weights to the features corresponding to each glioma image according to prior knowledge;

[0077] S12, select the most representative glioma image in each category as the centroid according to prior knowledge;

[0078] S13, calculate the distance between each weighted glioma image sample and the centroid, and classify it into the nearest centroid category, the distance calculation formula is as follows: D is the distance between the glioma sample to be classified and the glioma image cluster centroid, x, y are the coordinates of the glioma image cluster centroid, F is the weight of the centroid, F (i) is the weight of the glioma sample to be classified based on prior knowledge, x i , y i are the coordinates of the glioma image to be classified.

[0079]

[0080] S14, repeat the operation until all glioma images are classified.

[0081] Finally, the true value discovery algorithm is used to correct the glioma image category information obtained by the improved K-means algorithm, and the final glioma image category information, i.e. glioma image label information, is obtained. Pseudo-labels are added to glioma images without label information to achieve data augmentation.

[0082] Here, the improved K-means algorithm is based on prior knowledge to assign weights to sample category distance calculation, while the original K-means algorithm is based on Euclidean distance to calculate sample category distance. The improved K-means algorithm effectively considers the influence of medical image prior knowledge on sample classification. The improved K-means algorithm is based on prior knowledge to select the centroid, while the original K-means algorithm is based on randomization to select the centroid. The improved K-means algorithm effectively considers the influence of glioma image cluster centroid selection on sample classification.

[0083] After the glioma images are pre-clustered by the K-means algorithm, glioma images with true / pseudo labels are obtained and sent to the true value discovery algorithm for uncertainty correction.

[0084] For any multiple classifiers It can make L glioma image category predictions for L glioma images, and its output can be converted into a probability vector z∈Z at any scale, such that

[0085]

[0086] Assume random variables X, Y represent input sample data and sample labels, respectively, and random variable Z = f θ (X) represents a probability vector, if z l represents the actual probability of class l, then

[0087] P(Y = l | Z = z) = z l

[0088] However, since the input X (i) is misclassified as l ≠ y(i), it is difficult to estimate the probability P(Y = l | Z = z), so the equation is modified to the following equation, which only considers the probability that the sample is correctly predicted.

[0089] P(Y = y (i) | Z = z (i) ) = v (i)

[0090] where v (i) is the winning score, which is the only value that needs to be considered to evaluate the top-1 accuracy ACC (Accuracy).

[0091] For a classifier f θ with parameter distribution q(θ), the prediction of sample X (i) is:

[0092]

[0093] which can be approximated by S independent classifiers S classifiers can be obtained by independent random initialization or by learning rate decay to converge to a local minimum. Because each Therefore

[0094] In deep learning, most existing techniques are to make predictions by averaging multiple sets of classifiers, which come from: (1) random initialization of parameters; (2) learning rate decay; (3) connected optimization of loss function, and rarely consider the squared error between samples. Therefore, a common truth discovery algorithm is introduced.

[0095] In the truth discovery algorithm, for each sample (X (i) , y (i) ) in the validation set, S independent predictions come from all S independently trained models (denoted by z s in the following). Since the classifier is trained using stochastic gradient descent, there is an error in predicting the glioma image for each X (i) .

[0096] Truth discovery algorithm: Given a set of S classifiers In particular, the goal of truth discovery is to find the true probability vector z * ∈Δ L and find the confidence of ω s in S classifiers, so that the objective function is minimized:

[0097]

[0098]

[0099] From the above equation, we can clearly see the direct relationship between truth discovery algorithm and model uncertainty. Define the uncertainty v s of S independent training models as the replacement of the confidence ω s , that is So the above equation can be written as the following equation. Therefore, the truth discovery algorithm can ensure the calibration of model uncertainty by finding the true value vector.

[0100]

[0101]

[0102] After step one, the fruit fly algorithm is used to update the classification result weight of the deep convolutional neural network classification result, so that the result is more accurate.

[0103] The specific steps of the fruit fly algorithm are as follows:

[0104] S111: input the glioma image label weight, input the population size Popsize, the maximum number of iterations Maxgen of fruit fly, the flight radius R of fruit fly, the variable optimization number N, and give the fruit fly random initial position X_axis and Y_axis, the distance Dist, the smell concentration judgment value Smell_t, the smell concentration array Smell, and the smell concentration function Fitness;

[0105] S112: give the fruit fly population flight radius, that is, Xa=X_axis+randi(); Ya=Y_axis+randi(), update the parameters distance Dist, concentration judgment value S, smell concentration value Smell, and start the olfactory search;

[0106] Wherein, Xa represents the position horizontal coordinate after random motion in fruit fly olfactory search; Ya represents the position vertical coordinate after random motion in fruit fly olfactory search; X_axis represents the initial position horizontal coordinate given to fruit fly population; Y_axis represents the initial position vertical coordinate given to fruit fly population; randi() represents the random motion distance of fruit fly population;

[0107] S113: Find the extreme value Smellbest according to the initial smell concentration, retain the current fruit fly optimal individual and the optimal position; enter the visual search, and the fruit fly iteratively optimizes according to the population size Popsize and the iteration number Maxgen, finds the multivariate extreme value, calculates the smell concentration function fit by using the image entropy, and records the test result to the smell concentration array Smell; when the optimal smell concentration value Bestsmell is greater than the extreme value Smellbest, the fruit fly individual and the fruit fly position information X_axis and Y_axis at this time are retained, that is, the parameter information; the parameter information is used for updating the weight of the glioma medical image data set.

[0108] Wherein, the meaning of the initial smell concentration is that the image entropy calculated by the first generation of fruit fly individuals is calculated according to S113 above to find the maximum value.

[0109] As shown in Figure 2 , first, taking a medical image as an example, a pituitary tumor medical image is input, the population size Sizepop is initialized to 30, the iteration number Maxgen is 300, the fruit fly is given a random initial position X_axis and Y_axis, then the fruit fly individual motion is given a random direction and distance, the olfactory search is started, the smell concentration function fit is calculated by using the image entropy, and the test result is recorded to the smell concentration array Smell, the extreme value Smellbest is found according to the initial smell concentration, and the current fruit fly optimal individual and the optimal position are retained; enter the visual search, and the fruit fly iteratively optimizes according to the population size Sizepop and the iteration number Maxgen, finds the multivariate extreme value, when the optimal smell concentration value Bestsmell is greater than the extreme value Smellbest, the fruit fly individual, the fruit fly position information and the optimal smell concentration value X_axis, Y_axis, Bestsmell at this time are retained.

[0110] The glioma image data set is divided, the glioma image data set with a real label is placed in front of the total glioma image data set, the pseudo-label data set is placed behind the total glioma image data set, and the optimal weight of the real label and the optimal weight of the pseudo-label obtained by the fruit fly algorithm are respectively given, the glioma image data set with the given weight is sent into the network for learning and training and the following experiments.

[0111] In step two, the improved Venn diagram algorithm is used to calibrate the glioma image category weight obtained by the fruit fly algorithm. As shown in Figure 3As shown, first, the updated glioma image category weight obtained by the fruit fly algorithm is input, the error range is set, and the feature box tree is constructed using the auxiliary approximate decomposition algorithm; second, the approximate influence voronoi diagram is constructed, and the voronoi region is divided; then, the clustering center to be divided of the feature box tree of the glioma classification category is determined, if the feature box tree is empty, the approximate influence voronoi diagram algorithm is used for division; if the feature box tree is not empty, the auxiliary approximate influence voronoi diagram algorithm is used for division; finally, each voronoi unit is assigned a value, which is the class with the greatest influence on the distance clustering point, that is, the class of the glioma.

[0112] The specific operation steps of step two are:

[0113] S21, considering the joint influence of multiple glioma images, the space is decomposed into several cells, and the glioma image feature set extracted by the deep convolutional neural network is input;

[0114] S22, constructing glioma image feature point clusters, glioma image feature box trees, and divided glioma image feature point clusters;

[0115] S23, judging whether the feature box tree is empty, if the box tree is not empty, the first group of glioma feature point clusters after division is assigned a value, if the box tree is empty, the feature box tree is constructed, and the assignment operation is repeated;

[0116] S24, judging whether the glioma image is two-class or multi-class. If the glioma image classification is two-class, the error evaluation function is used to evaluate the error of the second group of glioma image feature point clusters after division, and the assignment operation is performed; if the glioma image classification is multi-class, the glioma image feature point clusters are sorted, and the feature point cluster division is performed again, and the assignment operation is repeated.

[0117] For the traditional voronoi diagram, given a set of points or objects P in space, the voronoi diagram obtained by P is a cell division of the space, where each cell of the voronoi diagram is the union of points whose distance to a particular point in the cell is less than the distance to any other point. The voronoi diagram can be regarded as the result of competition between certain points in the set P, for any point Q, the final merged voronoi diagram is the particular point that has the greatest influence on Q. However, for the glioma image classification category uncertainty correction experiment based on the improved voronoi diagram algorithm, it is desired to consider the joint influence of multiple glioma images for the experiment. In order to meet this expectation, the joint influence of multiple objects is considered in this patent, and a non-negative influence function F(C,Q) is used to measure it. F(C i ,Q), C iC is called the maximum influence cluster of Q. There are two main differences between the improved V-Graph and the traditional V-Graph:

[0118] Each V-Point of the improved V-Graph is a given point set (i.e., the features of multiple glioma images), while each V-Point of the traditional V-Graph is an input point (i.e., the features of a single glioma image);

[0119] The improved V-Graph is based on the joint influence of glioma images, while the traditional V-Graph is based on distance. To effectively solve the spatial partitioning and assignment problem of the traditional V-Graph, the improved V-Graph algorithm uses the divide-and-conquer method, combining the approximate influence V-Graph and the pruning method.

[0120] To construct the improved V-Graph algorithm, the main task is to divide the space into several cells and assign each glioma image to the approximate maximum influence cluster. The glioma image is divided into two subsets, and the approximate influence V-Graph of the glioma image dataset is constructed recursively and merged into the approximate influence V-Graph of the glioma image dataset with assignment. To keep the error within a controllable range, the approximate influence AI decomposition algorithm is extended to the assist approximate influence AAI decomposition algorithm.

[0121] AI decomposition steps: First, input a set of points P, error rate β > 0, and box tree structure data, then construct the feature box tree for point P, then construct the nodes of the feature box tree, and finally output the simplified feature box tree.

[0122] AI decomposition steps: First, obtain a subset of the glioma image dataset, then construct the feature box tree, and finally perform category division, i.e., the final glioma image belongs to which category.

[0123] For glioma image unit p ∈ P and any second category glioma image unit c, let D(c) be the diameter of unit c, and r be the shortest distance from unit c to p, then,

[0124]

[0125] For p ∈ P and the first category unit c with distance node v, then for q ∈ c and p' ∈ P\P v then satisfies,

[0126]

[0127] where P\P v represents the known subset P v, l(v) represents the input point with a distance from the node v, and P(|P|) represents the value obtained by the point cluster P in the multinomial bounded function.

[0128] In step three, the geometric optimization algorithm is used to post-correct the class probability corrected in step two based on the improved Venn diagram algorithm deep convolutional neural network classifier. As shown in Figure 4 , first input the high-dimensional vector data d, and the integrated vector data d ens is calculated. The index value I of the maximum value in the integrated vector data d ens is calculated, and the sorting algorithm calculates the index set S of . If the assignment parameters are equal, the high-dimensional vector d is directly output; if the assignment parameters are not equal, the mean value is calculated, and if the mean value is greater than the next parameter, the parameter is updated, and the latest mean value true vector d c is output; if the mean value is less than the next parameter, the mean value is continuously calculated until it is greater than the next parameter, and the latest mean value true vector d c is output. The loop is iterated until the maximum number of iterations is reached.

[0129] The specific operation steps of step three are as follows:

[0130] S31: input the high-dimensional vector data d of the glioma image obtained by the improved Venn diagram algorithm and the integrated data d ens of the glioma image;

[0131] S32: obtain the index value I of the maximum data in the current integrated data d ens and obtain the index set S of the sorted value in , wherein d i represents i high-dimensional vector data, and L represents the number of glioma images;

[0132] S33: if the first value of the index value set S of the sorted value is equal to the index value I of the maximum data in the current integrated data d ens , the high-dimensional vector data d is returned;

[0133] S34: if the first value of the index value set S of the sorted value is not equal to the index value I of the maximum data in the current integrated data d ens , the corrected high-dimensional vector data d c is calculated;

[0134] S35: if the corrected high-dimensional vector data d c is greater than the high-dimensional vector data d, the corrected high-dimensional vector data d c is updated;

[0135] S36: repeat S34 and S35 until all high-dimensional vectors are updated.

[0136] Since the related algorithm based on post-model correction may affect the accuracy of the model if the probability of each glioma image classification category cannot be kept unchanged. And the common ground truth discovery algorithm may also not meet the experimental requirement of keeping the accuracy unchanged. Because the real probability vector d * may change. Suppose dp is the prediction from the ensemble vector d ens , i.e. dp = argmax l (d ens ) l After the ground truth vector is obtained by the ground truth discovery algorithm, the accuracy of the ensemble vector d ens is expected to remain unchanged. Then the geometric optimization accuracy preservation formula is adopted:

[0137]

[0138]

[0139]

[0140] d * represents the real probability vector, ω s represents the credibility, s represents the jth classifier model, S represents the number of classifiers, and l represents the dimension of the high-dimensional vector. The subspace Ω is introduced in the constraint specification formula d l > d m , Through the above formula, we can make the real probability vector d * fall within Δ α = Δ L∩Ω , Δ α represents the accuracy preservation area, and Δ L∩Ω represents the intersection area of the original probability and the predicted probability. If there is a real probability vector d * outside Δ α , by finding the projection value of the real probability vector d * within Δ α , the real probability vector d α outside Δ * is pulled back to within Δ α , so that the accuracy is not affected.

[0141] In step four, based on the experimental results of step one, step two and step three, the category information of each glioma medical image is obtained, and intelligent diagnosis of glioma images is realized.

[0142] The application provides a glioma image intelligent diagnosis method based on geometric optimization uncertainty correction. The above examples are only used for describing the technical concept and characteristics of the application, and the purpose is to enable those skilled in the art to understand the content of the application and implement it, and cannot limit the protection scope of the application. Any equivalent changes or modifications made according to the spirit and principle of the application shall be covered within the protection scope of the application.

Claims

1. A method for intelligent diagnosis of glioma images based on geometric optimization uncertainty correction, characterized in that, Includes the following steps: Step 1: Collect several glioma images, extract features from them using a deep convolutional neural network, and classify the extracted glioma image features; Step 2: Considering the combined effect of multiple glioma images, improve the Vinaut chart algorithm, and perform category uncertainty correction on glioma images based on the improved Vinaut chart algorithm; Step 3: Post-correction of the glioma image classification model based on geometric optimization algorithm to improve the problem of decreased accuracy of the corrected model in glioma images; Step 4: Obtain the final category information for each glioma medical image to achieve intelligent diagnosis of glioma images; The specific steps for step two are as follows: S21: Considering the combined effect of multiple glioma images, the space is decomposed into several cells, and the glioma image feature set extracted by a deep convolutional neural network is input. S22: Divide the glioma image dataset, place the glioma image dataset with real labels before the total glioma image dataset, and place the pseudo-label dataset after the total glioma image dataset. Assign the optimal weights of the real labels and pseudo labels obtained by the fruit fly algorithm to the glioma image feature point clusters, glioma image feature box trees, and the divided glioma image feature point clusters. S23: Determine if the feature box tree is empty. If the box tree is not empty, assign values ​​to the first group of glioma feature point clusters after partitioning. If the box tree is empty, construct the feature box tree and repeat the assignment operation. S24: Determine whether the glioma image is classified as binary or multi-class. If the glioma image is classified as binary, perform error evaluation on the feature point clusters of the second group of glioma images after division and perform the value assignment operation. If the glioma image is classified as multi-class, sort the feature point clusters of the glioma image, re-divide the feature point clusters, and repeat the value assignment operation. Step three includes the following process: S31: Input the high-dimensional vector data d of multiple glioma images obtained by the improved Vinaurus algorithm and the ensemble data d of the glioma images. ens ; S32: Get the current integration data d ens The index value I of the largest data and how to obtain it A set S of indices of sorted numerical values, where d i Let L represent i high-dimensional vector data, and L represent the number of glioma images; S33: If the first value of the set of index values ​​S of the sorted values ​​is equal to the current integrated data d ens If the index value I of the largest data in the array is given, then the high-dimensional vector data d is returned. S34: If the first value of the set of index values ​​S that does not satisfy the sorting criteria is equal to the current integrated data d ens If the index value I of the largest data in the middle is used, then the corrected high-dimensional vector data d is calculated. c ; S35: If the high-dimensional vector data d is corrected c If the value is greater than the high-dimensional vector data d, then update and correct the high-dimensional vector data d. c ; S36: Repeat S34 and S35 until all high-dimensional vectors have been updated.

2. The intelligent diagnostic method for glioma images based on geometric optimization uncertainty correction according to claim 1, characterized in that, In step one: First, the size of the glioma image is adjusted to the size of the model input layer and input into the pre-trained deep convolutional neural network model to learn and train on the glioma image data, and to extract glioma image features to improve the model's classification accuracy and generalization ability for glioma image data. Then, the extracted glioma image features are clustered to obtain the clustered glioma image category information, and pseudo-labels are added to glioma images without label information; Finally, a truth discovery algorithm is used to correct the uncertainty of the glioma image category information obtained by clustering, so as to achieve the purpose of data amplification, reducing experimental error and improving the generalization ability of the model.

3. The intelligent diagnostic method for glioma images based on geometric optimization uncertainty correction according to claim 2, characterized in that: An improved K-means algorithm was used to cluster the extracted glioma image features. The improved K-means algorithm for clustering the extracted glioma image features includes the following steps: S01: Input the glioma image features extracted by the pre-trained convolutional neural network, and assign weights to the features corresponding to each glioma image based on prior knowledge; S02: Based on prior knowledge, select the most representative glioma image in each category as the focal point; S03: Calculate the distance between each weighted glioma image sample and the cluster center, and assign the glioma image to the class with the nearest cluster center; S04: Repeat the operation until all glioma images are classified.

4. The intelligent diagnostic method for glioma images based on geometric optimization uncertainty correction according to claim 3, characterized in that: In S03, the distance between each weighted glioma image sample and the center of the convergence is calculated using the following formula: Where D represents the distance between the glioma sample to be classified and the cluster center of the glioma image, x and y are the coordinates of the cluster center of the glioma image, and F is the weight of the cluster center. (i) For the prior knowledge-based weights of the glioma samples to be classified, x i y i These are the coordinates of the image of the glioma to be classified.

5. The intelligent diagnostic method for glioma images based on geometric optimization uncertainty correction according to claim 1, characterized in that: Step 1 uses the fruit fly algorithm to update the weights of the classification results of the deep convolutional neural network, making the results more accurate.

6. The intelligent diagnostic method for glioma images based on geometric optimization uncertainty correction according to claim 1, characterized in that: To keep the error under control, the following steps are taken: First, input a set of points P, with an error rate β > 0 and box tree structure data. Then, construct a feature box tree for point P, construct the nodes of the feature box tree, and finally output the simplified feature box tree. A subset of the glioma image dataset is obtained, a feature box tree is constructed, and finally, the categories to which the glioma images belong are classified. For a box node p∈P with a distance of x1, and for any second-class glioma image unit c, let D(c) be the diameter of unit c, and r be the shortest distance from unit c to p, then... For a cell c of the first class with p∈P and distance v, then for any point q∈c and p'∈P\P v Then it is satisfied. Among them, P\P v P represents a known subset of P. v l(v) represents the input point with a distance from node v, and Ρ(|P|) represents the value obtained by the point cluster P in the polynomial bounded function.

7. The intelligent diagnostic method for glioma images based on geometric optimization uncertainty correction according to claim 1, characterized in that, The accuracy preservation formula for geometric optimization is: Where, d * Let ω represent the true probability vector. s Let s represent the confidence level, s represent the nth classifier model, S represent the number of classifiers, and l represent the dimension of the high-dimensional vector. Under the constraint normalization formula... Introducing subspace Ω: d l >d m , The above formula enables the true probability vector d to be changed in each iteration. * Falling on Δ α =Δ L∩Ω , where Δ α Indicates the precision preservation region, Δ L∩Ω This represents the intersection region of the original probability and the predicted probability, if there is a true probability vector d. * Falling on Δ α In addition, by finding the true probability vector d * In Δ α The projected value on will fall on Δ α The true probability vector d outside * Mapping to Δ α This ensures that the accuracy remains unaffected.

Images