A method for constructing an intracranial tumor diagnosis model based on a double-path parallel hierarchical network

Abstract

The application discloses a method for constructing an intracranial tumor diagnosis model based on a double-path parallel hierarchical network. First, according to the pathological characteristics of intracranial tumors, prior tumor characteristics of solid tumors and intratumoral necrosis are extracted. Then, the data set is re-divided through multi-feature entropy weight to optimize the feature learning ability. Finally, the double-path structure of the double-path parallel hierarchical network (ME-DPNet) is used to realize multi-modal input, and multiple features are superimposed and fused to realize grading. The constructed diagnosis model has high grading accuracy and can automatically grade intracranial tumors to assist doctors in reducing subjective diagnostic errors.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation, and in particular to a method for constructing an intracranial tumor diagnostic model based on a dual-path parallel hierarchical network. Background Technology

[0002] Intracranial tumors originate from abnormally proliferating cells in the brain. They occupy brain space, replace tissues needed for vital bodily functions, affect the central nervous system, and most are accompanied by complications. The number of patients is increasing globally, with persistently high morbidity and mortality rates. Intracranial tumors are typically diagnosed using MRI imaging. MRI's multimodal nature provides multi-angle imaging results, making it the most sensitive non-invasive imaging method for diagnosing intracranial tumors. Certain specialized sequences can provide further information about the tumor's internal structure; these include commonly used sequences such as T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and T1WI contrast-enhanced imaging (T1CE). The grading and diagnosis of intracranial tumors are crucial for developing treatment plans, prognostic assessments, and establishing preoperative clinical guidelines.

[0003] Currently, deep learning classification methods for intracranial tumors exist, such as traditional classification models like AlexNet, VGG, and ResNet. These models improve network performance by continuously increasing network depth, thereby refining the extraction of lesion features. However, as the severity of tumor lesions increases, local details of the lesion area become difficult to capture, and traditional models cannot accurately identify the lesion level. Lu et al. improved the traditional AlexNet by replacing the last three layers of the network with an extreme learning machine, which improved the model's generalization ability, but the accuracy of multi-class classification in clinical practice still needs improvement. Naser et al. built a fully connected classifier for tumor grading to overcome overfitting; Khan et al. fine-tuned the VGG19 model, but both used synthetic data augmentation techniques to obtain samples to increase the available data for classifier training. Although they can correctly classify benign and malignant tumors, and even grade II and III tumors, abnormal errors exist when facing multi-level classifications above grade IV, resulting in unsatisfactory model performance. Because intracranial tumors need to be classified into four levels during clinical diagnosis, and the pathological characteristics of tumors at different levels are different, the existing methods for grading intracranial tumors have problems such as identification errors and low accuracy. Summary of the Invention

[0004] The present invention aims to solve the aforementioned technical problems existing in the prior art by providing a method for constructing an intracranial tumor diagnostic model based on a dual-path parallel hierarchical network.

[0005] The technical solution of this invention is: a method for constructing an intracranial tumor diagnostic model based on a dual-path parallel hierarchical network, which is carried out according to the following steps:

[0006] Step 1. Extract features of solid masses;

[0007] Step 2. Extract intratumoral necrosis features;

[0008] Step 3. Optimize the dataset using multi-feature entropy weighting;

[0009] Step 4. Use a dual-path parallel grading diagnostic model to identify and grade intracranial tumors.

[0010] Step 1 involves inputting an intracranial tumor image I and using an extraction algorithm based on slope difference distribution to extract the solid mass segmentation results, which are then used as prior-learned solid mass features.

[0011] Step 2 involves inputting an intracranial tumor image I and using a whale-based maximum entropy optimization segmentation method to extract the segmentation results of intratumoral necrosis. These results are then used as prior learning features of intratumoral necrosis, as detailed below:

[0012] An intracranial tumor image I has O×Q pixels and contains G gray levels. The average gray value of the pixel is calculated based on pixel I(x,y) and its eight neighbors. ij P represents the number of pixels with pixel value i and average gray value j. ij The probability of a pixel meeting this condition appearing:

[0013]

[0014] An intracranial tumor image I is divided into four regions by a threshold (s,w): target region A, background region B, boundary region C, and noise region D. In target region A and boundary region B, the optimal threshold for image segmentation is calculated using two-dimensional maximum entropy, where the entropy function is defined as:

[0015]

[0016]

[0017] in,

[0018] The total entropy of the image is:

[0019] E(s,w)=E(A)+E(B) (4)

[0020] The optimal segmentation threshold selected is:

[0021] E(s * ,w * = max(E(s,w)) (5)

[0022] To fully integrate the maximum entropy algorithm with the whale optimization algorithm, the entropy function of the dead region is used when searching for the optimal segmentation threshold. As the fitness function for optimization, the fitness value is calculated, and the minimum value obtained is W. b Then update the parameters; to further improve local search capabilities, add an adaptive weight ω. μ is a constant that accelerates the convergence rate;

[0023] Perform local and global searches using the following formulas:

[0024]

[0025] W i (t+1)=W r (t)-AD' (7)

[0026] After iteration T, select the optimal segmentation threshold (s) that satisfies the conditions. * ,w * To divide the necrotic region, in formula (6), W i (t+1) is a candidate position for the next transformation, W b (t) represents the current optimal position of the whale, t is the current iteration number, A and C are coefficients, A = a(2r-1), C = 2r, and a is a control parameter that decreases linearly from 2 to 0. r is a random number in [0,1], T is the maximum number of iterations, and D = |C*W b (t)-W i (t)|,W i (t) represents the current position of the whale, P w To determine the probability of selecting the contraction encirclement mechanism, 1-P w To select the probability of updating the whale's position using the spiral model, b is a constant used to constrain the spiral shape, and l is a random number in the range [-1, 1], representing the current distance between the whale and the target. In formula (7), the distance D' = |C*W r (t)-W i (t)|,W r (t) represents the randomly selected whale location;

[0027] Step 3 involves using the Maximum Information Coefficient (MIC) to calculate the correlation between multiple features and the label, as detailed below:

[0028] The mutual information between the input intracranial tumor image I and the label L is represented by information entropy, and there are a total of L. I L II L III L IV Four labels, starting with L IFor example, its mutual information is defined as:

[0029] R(I,L I )=E(I)-E(I|L I (8)

[0030] In formula (8), E(I) is the information entropy of the input image, E(I|L I ) indicates that in the known label L I Under the given conditions, the information entropy of the input image is used to calculate the MIC value:

[0031]

[0032] U and V represent the size of the data graph, DS is the number of images in the corresponding dataset, and R... * (I,L I () represents the maximum mutual information value for a given value;

[0033] After calculating the MIC values ​​of the four types of labels, a vector is formed, representing the four types of images (I, II, III, IV) contained in the intracranial tumor image data. w I c I s I n These represent T2WI, T1CE, solid mass, and intratumoral necrosis images, respectively, and can form four vectors: o, p, q, and f, defined as follows:

[0034]

[0035] Find the entropy E of o, p, q, f. o The definition is as follows:

[0036]

[0037] Among them, entropy value o I This corresponds to the I-level feature vector; other entropy values ​​can be obtained similarly.

[0038] Based on this, the weights of o, p, q, and f are as follows:

[0039]

[0040] Step 4 is a dual-path parallel hierarchical network based on multi-feature entropy weighting of intracranial tumors (ME-DPNet), which uses T2WI and T1CE to achieve dual-path input and then performs multi-feature fusion.

[0041] The dual-path parallel hierarchical network has multiple input layers, with two paths inputting T2WI image set X1 and T1CE image set X2 respectively. The output layer is responsible for outputting the target result P. g, g represents the lesion grade of different intracranial tumors, g={1,2,3,4,5} representing normal, grade I, grade II, grade III and grade IV respectively; it includes convolutional modules, pooling layers and fully connected layers. The nonlinear relationship between each layer is deepened by adding a nonlinear activation function ReLU(x,y)=max(0,(x,y)) after the convolutional layer, and the training of the network is accelerated by reducing the internal variable bias through batch normalization layers.

[0042]

[0043] In formula (13), ε is a constant to prevent erroneous calculations, N is the number of input images, and k is the ordinal number. Normalization for channel z;

[0044] The image features of T2WI and T1CE, as well as the extracted features of the solid mass and intratumoral necrosis, are fused at multiple scales:

[0045]

[0046] Where c represents the number of network channels, K represents the convolution kernel, * represents convolution, + represents the superposition and fusion of information from multiple features, and Y1, Y2, Y3, and Y4 represent T2WI, T1CE, solid mass features, and intratumoral necrosis features, respectively.

[0047] By using fully connected layers to map multi-level lesion features to the label space, and employing a Softmax classifier to calculate the probability values ​​of the fully connected layers, multi-level classification of intracranial tumor lesion severity is achieved.

[0048] This invention presents a method for constructing an intracranial tumor diagnostic model based on a dual-path parallel hierarchical network. First, prior tumor features, including solid masses and intratumoral necrosis, are extracted based on the pathological characteristics of intracranial tumors. Then, the dataset is re-divided using multi-feature entropy weighting to optimize feature learning capabilities. Finally, the dual-path structure of the dual-path parallel hierarchical network (ME-DPNet) enables multimodal input, and multiple features are superimposed and fused to achieve hierarchical classification. The constructed diagnostic model exhibits high classification accuracy and can automatically classify intracranial tumors, assisting physicians in reducing subjective diagnostic errors. Attached Figure Description

[0049] Figure 1 This is a flowchart of an embodiment of the present invention.

[0050] Figure 2 This is the feature extraction result of a solid intracranial tumor in an embodiment of the present invention.

[0051] Figure 3This is the result of extracting intracranial tumor necrosis features according to an embodiment of the present invention.

[0052] Figure 4 This is a flowchart of the multi-feature entropy weight optimization process according to an embodiment of the present invention.

[0053] Figure 5 This is a schematic diagram of the dual-path parallel hierarchical network (ME-DPNet) according to an embodiment of the present invention.

[0054] Figure 6 This is the intracranial tumor grading result of an embodiment of the present invention.

[0055] Figure 7 This is a graph showing the model accuracy results of an embodiment of the present invention.

[0056] Figure 8 This is an evaluation table for different Epoch results in embodiments of the present invention.

[0057] Figure 9 This is an evaluation table for different learning rates in embodiments of the present invention.

[0058] Figure 10 This is a comparison of the accuracy of different convolutional layers in embodiments of the present invention.

[0059] Figure 11 This is an evaluation table for the grading results of existing algorithms in this embodiment of the invention. Detailed Implementation

[0060] The present invention provides a method for constructing an intracranial tumor diagnostic model based on a dual-path parallel hierarchical network, as follows: Figure 1 As shown, proceed as follows:

[0061] Step 1. Extract features of solid masses;

[0062] Step 2. Extract intratumoral necrosis features;

[0063] Step 3. Optimize the dataset using multi-feature entropy weighting;

[0064] Step 4. Use a dual-path parallel grading diagnostic model to identify and grade intracranial tumors.

[0065] Step 1 involves inputting an intracranial tumor image I and using an extraction algorithm based on slope difference distribution to extract the solid mass segmentation results, which are then used as prior-learned solid mass features. The solid mass feature extraction results are as follows: Figure 2 As shown;

[0066] Step 2 involves inputting an intracranial tumor image I and using a whale-based maximum entropy optimization segmentation method to extract the segmentation results of intratumoral necrosis as prior learning of intratumoral necrosis features.

[0067] Specifically as follows:

[0068] An intracranial tumor image I has O×Q pixels and contains G gray levels. The average gray value of the pixel is calculated based on pixel I(x,y) and its eight neighbors. ij P represents the number of pixels with pixel value i and average gray value j. ij The probability of a pixel meeting this condition appearing:

[0069]

[0070] An intracranial tumor image I is divided into four regions by a threshold (s,w): target region A, background region B, boundary region C, and noise region D. In target region A and boundary region B, the optimal threshold for image segmentation is calculated using two-dimensional maximum entropy, where the entropy function is defined as:

[0071]

[0072]

[0073] in,

[0074] The total entropy of the image is:

[0075] E(s,w)=E(A)+E(B) (4)

[0076] The optimal segmentation threshold selected is:

[0077] E(s * ,w * = max(E(s,w)) (5)

[0078] To fully integrate the maximum entropy algorithm with the whale optimization algorithm, the entropy function of the dead region is used when searching for the optimal segmentation threshold. As the fitness function for optimization, the fitness value is calculated, and the minimum value obtained is W. b Then update the parameters; to further improve local search capabilities, add an adaptive weight ω. μ is a constant that accelerates the convergence rate;

[0079] Perform local and global searches using the following formulas:

[0080]

[0081] W i (t+1)=W r (t)-AD' (7)

[0082] After iteration T, select the optimal segmentation threshold (s) that satisfies the conditions. * ,w* To divide the necrotic region, in formula (6), W i (t+1) is a candidate position for the next transformation, W b (t) represents the current optimal position of the whale, t is the current iteration number, A and C are coefficients, A = a(2r-1), C = 2r, and a is a control parameter that decreases linearly from 2 to 0. r is a random number in [0,1], T is the maximum number of iterations, and D = |C*W b (t)-W i (t)|,W i (t) represents the current position of the whale, P w To determine the probability of selecting the contraction encirclement mechanism, 1-P w To select the probability of updating the whale's position using the spiral model, b is a constant used to constrain the spiral shape, and l is a random number in the range [-1, 1], representing the current distance between the whale and the target. In formula (7), the distance D' = |C*W r (t)-W i (t)|,W r (t) represents the randomly selected whale location; the results of tumor necrosis feature extraction are as follows: Figure 3 As shown.

[0083] Step 3 involves using the Maximum Information Coefficient (MIC) to calculate the correlation between multiple features and the label, specifically as follows: Figure 4 As shown:

[0084] The mutual information between the input intracranial tumor image I and the label L is represented by information entropy, and there are a total of L. I L II L III L IV Four labels, starting with L I For example, its mutual information is defined as:

[0085] R(I,L I )=E(I)-E(I|L I (8)

[0086] In formula (8), E(I) is the information entropy of the input image, E(I|L I ) indicates that in the known label L I Under the given conditions, the information entropy of the input image is used to calculate the MIC value:

[0087]

[0088] U and V represent the size of the data graph, DS is the number of images in the corresponding dataset, and R... * (I,L I() represents the maximum mutual information value for a given value;

[0089] After calculating the MIC values ​​of the four types of labels, a vector is formed, representing the four types of images (I, II, III, IV) contained in the intracranial tumor image data. w I c I s I n These represent T2WI, T1CE, solid mass, and intratumoral necrosis images, respectively, and can form four vectors: o, p, q, and f, defined as follows:

[0090]

[0091] Find the entropy E of o, p, q, f. o The definition is as follows:

[0092]

[0093] Among them, entropy value o I This corresponds to the I-level feature vector; other entropy values ​​can be obtained similarly.

[0094] Based on this, the weights of o, p, q, and f are as follows:

[0095]

[0096] Step 4 is based on, for example Figure 5 The dual-path parallel hierarchical network (ME-DPNet) for intracranial tumor multi-feature entropy weighting, shown in the figure, can intuitively highlight the advantages of T2WI and T1CE images, realize dual-path input, and then perform multi-feature fusion to clearly reflect the multiple details of the lesion.

[0097] The dual-path parallel hierarchical network model has multiple input layers, with two paths inputting T2WI image set X1 and T1CE image set X2 respectively. The output layer is responsible for outputting the target result P. g , g represents the lesion grade of different intracranial tumors, g={1,2,3,4,5} representing normal, grade I, grade II, grade III and grade IV respectively; it includes convolutional modules, pooling layers and fully connected layers. The nonlinear relationship between each layer is deepened by adding a nonlinear activation function ReLU(x,y)=max(0,(x,y)) after the convolutional layer, and the training of the network is accelerated by reducing the internal variable bias through batch normalization layers.

[0098]

[0099] In formula (13), ε is a constant to prevent erroneous calculations, N is the number of input images, k is the ordinal number, and F kz Normalization for channel z;

[0100] The ME-DPNet model achieves multi-feature fusion, integrating T2WI and T1CE image features with extracted solid mass features and intratumoral necrosis features at multiple scales:

[0101]

[0102] Where c represents the number of network channels, K represents the convolution kernel, * represents convolution, + represents the superposition and fusion of multiple features, and Y1, Y2, Y3, and Y4 represent T2WI, T1CE, solid mass features, and intratumoral necrosis features, respectively. The ME-DPNet model reduces the coupling and correlation between layers and removes redundant information by adding a max pooling layer.

[0103] By using fully connected layers to map multi-level lesion features to the label space, and employing a Softmax classifier to calculate the probability values ​​of the fully connected layers, multi-level classification of intracranial tumor lesion severity is achieved.

[0104] The intracranial tumor grading results of this invention are as follows: Figure 6 As shown, the model accuracy results of this embodiment of the invention are as follows: Figure 7 As shown.

[0105] experiment:

[0106] 1. Comparison experiment of different Epochs in the embodiments of the present invention

[0107] By selecting different Epoch parameters, the model was experimentally compared to obtain the most suitable Epoch. Evaluation tables for different evaluation metrics are shown below. Figure 8 As shown.

[0108] 2. Comparison experiment of different learning rates in embodiments of the present invention

[0109] By selecting different learning rates and conducting experimental comparisons of the model, the most suitable learning rate parameter can be obtained. Evaluation tables for different evaluation metrics are shown below. Figure 9 As shown.

[0110] 3. Comparison experiment of different convolutional layers in the embodiments of the present invention

[0111] By setting different numbers of convolutional layers, experiments were conducted to compare the models and determine the number of convolutional layers that achieved the highest accuracy. The accuracy results are compared below. Figure 10 As shown.

[0112] 4. Comparison Experiment of Grading Results of the Embodiment of the Invention (ME-DPNet) with Existing Traditional and Advanced Models

[0113] The present invention (ME-DPNet), the traditional model, and the model methods described in references [1]-[9] were used to classify and diagnose intracranial tumors. The evaluation table of the classification results for different evaluation indicators is shown below. Figure 11 As shown.

[0114] The results showed that the accuracy of the present invention in grading and diagnosing intracranial tumors was superior to that of the control group.

[0115] Comparative literature:

[0116] [1] S.Lu, SHWang and YDZhang, "Detection of abnormal brain in MRIvia improved AlexNet and ELM optimized by chaotic bat algorithm," NeuralComputing and Applications, vol.33, pp.10799-10811, 2021.

[0117] [2]MANaser and MJDeen, "Brain tumor segmentation and grading of lower-grade glioma using deep learning in MRI images," Computers in Biologyand Medicine, vol.121, pp.10375, 2020.

[0118] [3] ARKhan et al., "Brain tumor segmentation using K-means clustering and deep learning with synthetic data augmentation for classification," Microscopy Research and Technique, vol.84, no.7, pp.1389-1399, 2021.

[0119] [4]E.Irmak,“Multi-Classification of Brain Tumor MRI Images Using DeepConvolutional Neural Network with Fully Optimized Framework,”Iranian Journalof Science and Technology,Transactions of Electrical Engineering,vol.45,pp.1015-1036,2021.

[0120] [5]M.Nawaz et al.,“Analysis of Brain MRI Images Using ImprovedCornerNet Approach,”Diagnostics,vol.11,no.10,pp.1856,2021.

[0121] [6]H.H.Sultan,N.M.Salem and W.Al-Atabany,“Multi-Classification ofBrain Tumor Images Using Deep Neural Network,”IEEE Access,vol.7,pp.69215-69225,2019.

[0122] [7]G.Gilanie et al.,“Risk-free WHO grading of astrocytoma usingconvolutional neural networks from MRI images,”Multimedia Tools andApplications,vol.80,pp.4295-4306,2021.

[0123] [8]Latif G et al.,“Glioma Tumors’Classification Using Deep-Neural-Network-Based Features with SVM Classifier,”Diagnostics,vol.12,no.4,pp.1018,2022.

[0124] [9]R.Singh,A.Goel and D.K.Raghuvanshi,“Computer-aided diagnosticnetwork for braintumor classification employing modulated Gabor filterbanks,”The Visual Computer,vol.37,pp.2157-2171,2021。

Claims

1. A method for constructing an intracranial tumor diagnostic model based on a dual-path parallel hierarchical network, characterized in that... Follow these steps: Step 1. Extract features of solid masses; Step 2. Extract intratumoral necrosis features; Step 3. Optimize the dataset using multi-feature entropy weights; Step 4. Use a dual-path parallel grading diagnostic model to identify and grade intracranial tumors; Step 1 involves inputting an image of the intracranial tumor. An extraction algorithm based on slope difference distribution is used to extract solid mass segmentation results, which are then used as prior-learned solid mass features. Step 2 involves inputting an image of the intracranial tumor. A whale-based maximum entropy optimization segmentation method was used to extract the segmentation results of intratumoral necrosis, which were then used as prior learning features of intratumoral necrosis, as follows: Intracranial tumor images have 1 pixel, containing G Each gray level, based on pixel points The average gray value of the pixel is calculated using its eight neighboring pixels. Indicates pixel value And the average gray value is The number of pixels, The probability of a pixel meeting this condition appearing: （1） Intracranial tumor images Threshold The image is divided into four regions: target region A, background region B, boundary region C, and noise region D. In target region A and boundary region B, the optimal threshold for image segmentation is calculated using two-dimensional maximum entropy, where the entropy function is defined as: （2） （3） in, , ; The total entropy of the image is: （4） The optimal segmentation threshold selected is: （5） To fully integrate the maximum entropy algorithm with the whale optimization algorithm, the entropy function of the dead region is used when searching for the optimal segmentation threshold. As the fitness function for optimization, the fitness value is calculated, and the minimum value obtained is... Then update the parameters; to further improve local search capabilities, add adaptive weights. , , A constant used to accelerate convergence; Perform local and global searches using the following formulas: （6） （7） When the iteration reaches Then, select the optimal segmentation threshold that meets the conditions. To segment the necrotic region, in formula (6), It is a candidate position for the next transformation. This is the optimal position for the whale at present. t This represents the current iteration number. A and C For coefficients, , , As a control parameter, it shows a linear decreasing trend from 2 to 0. , yes 0 1 Random numbers in the data, The maximum number of iterations, , This is the current location of the whale. To determine the probability of selecting the contraction encirclement mechanism, To select a spiral model to update the probability of the whale's position, It is a constant used to constrain the spiral shape. for -1 1 The random number represents the current distance between the whale and the target. In formula (7), the distance is... , This indicates a randomly selected whale location; Step 3 involves calculating the correlation between multiple features and labels using the maximum information coefficient, as detailed below: The input intracranial tumor image is represented by information entropy. With tags Mutual information between them, total , , , Four tags, with For example, its mutual information is defined as: （8） In formula (8), The information entropy of the input image, Indicates that in the known label Under the given conditions, the information entropy of the input image is used to calculate the MIC value: （9） and Indicates the size of the data graph. The number of images in the corresponding dataset. Represents the maximum mutual information value for a given value; After calculating the MIC values ​​of the four categories of labels, a vector is formed, representing the four categories of images contained in the intracranial tumor image data. , , , These represent T2WI, T1CE, solid mass, and intratumoral necrosis images, respectively, and can form... , , , Four vectors, defined as follows: （10） right , , , Calculate its entropy. The definition is as follows: （11） Among them, entropy value , This corresponds to the I-level feature vector; other entropy values ​​can be obtained similarly. Based on this, , , , The weights are as follows: （12） Step 4 is a dual-path parallel hierarchical network based on multi-feature entropy weighting for intracranial tumors. It utilizes T2WI and T1CE to achieve dual-path input, followed by multi-feature fusion. The dual-path parallel hierarchical network has multiple input layers, with the two paths respectively inputting T2WI image sets. and T1CE image set The output layer is responsible for outputting the target result. , This indicates the lesion grade of different intracranial tumors. These represent normal, level I, level II, level III, and level IV, respectively; they include convolutional modules, pooling layers, and fully connected layers, achieved by adding a non-linear activation function after the convolutional layers. To deepen the nonlinear relationships between layers, and to accelerate network training by reducing internal variable bias through batch normalization layers: （13） In formula (13), It is a constant to prevent incorrect calculations. The number of input images, For ordinal numbers, For channel Normalization below; The image features of T2WI and T1CE, as well as the extracted features of the solid mass and intratumoral necrosis, are fused at multiple scales: （14） in The number of network channels, Represents the convolution kernel. For convolution, + indicates the overlay and fusion of information from multiple features. , , , These are T2WI, T1CE, solid mass characteristics, and intratumoral necrosis characteristics, respectively. By using fully connected layers to map multi-level lesion features to the label space, and employing a Softmax classifier to calculate the probability values ​​of the fully connected layers, multi-level classification of intracranial tumor lesion severity is achieved.

Images