Glioma mr image segmentation method based on edge loss and knowledge fusion decision

Abstract

This invention provides a glioma MR image segmentation method based on edge loss and knowledge fusion decision-making. The segmentation model includes: a grading module for providing grading information after grading; a segmentation module for providing area ratio information of different regions after initial segmentation; a fuzzy decision module for receiving the grading information from the grading module and the area ratio information from the segmentation module, performing fuzzy decision-making, and outputting control coefficients for edema regions and enhanced tumor regions; an adaptive post-processing module for multiplying the control coefficients by the predicted values ​​of the corresponding initially segmented edema regions and enhanced tumor regions to obtain the probability distribution of edema regions and enhanced tumor regions; extracting all voxels that were initially segmented into edema regions and enhanced tumor regions respectively, and using voxels with a probability greater than a threshold of 0.5 as new edema regions and enhanced tumor regions, with the tumor core region being the result of subtracting the new edema region from the overall tumor region.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and in particular relates to a method for glioma MR image segmentation based on edge loss and knowledge fusion decision-making. Background Technology

[0002] Medical image analysis is widely used in clinical diagnosis and medical research, such as computer-aided diagnosis, medical robotics, and survival analysis. Among the many tasks in medical image analysis, glioma MR image segmentation has always been a research hotspot. Despite researchers' tireless efforts, accurate glioma MR image segmentation remains a key challenge due to issues such as location uncertainty, morphological uncertainty, data imbalance, and low image contrast. Gliomas can be further classified into high-grade gliomas (HGG) and low-grade gliomas (LGG). As a currently popular non-invasive imaging technique, MR can provide high-quality, non-invasive imaging of gliomas and eliminate skull artifacts, making it a primary tool for radiologists in diagnosing and treating gliomas. Multimodal MRI tumor tissue type localization aids in the diagnosis of glioma patients, delineating viable tumors from edema and normal brain tissue. Furthermore, glioma segmentation on MR images is an important first step in cancer diagnosis, providing quantitative and intuitive assistance for disease assessment. The ultimate goal of glioma segmentation is to divide gliomas into three regions: the whole tumor region (WT), the tumor core region (TC), and the enhanced tumor region (ET). However, due to the significant differences in shape, size, texture, and location of brain tumors, and the unclear boundaries between them and normal brain tissue, manually segmenting brain tumors from MR images is subjective, tedious, and time-consuming. To address these issues, researchers urgently need to develop solutions. Therefore, designing an automated and robust glioma segmentation method for accurate and effective glioma segmentation is essential.

[0003] Current glioma segmentation primarily relies on well-designed deep neural networks to automatically extract high-dimensional discriminative features. Among these, designing effective modules and network architectures has become a crucial factor in achieving accurate segmentation performance. Network architecture design mainly involves shifts from single-channel to multi-channel networks, from fully connected layer networks to fully convolutional networks, and from simple networks to deep cascaded networks. The aim is to enhance the network's feature learning ability and achieve more precise segmentation. While network structure design improves segmentation accuracy to some extent, the limited input knowledge still leads to missegmentation in certain brain regions. Summary of the Invention

[0004] To address the aforementioned issues, this invention first crops and standardizes MR image data of gliomas, then trains a grading module and a segmentation module separately. In the segmentation module, a novel edge region voxel-weighted loss function is designed. This loss function first extracts the edges of the predicted image and the ground truth label as the calculation region for loss, and then assigns new weights to voxels using the distance from the voxel to the edge. The grading information and the area ratio information of the initially segmented image are used as inputs to the fuzzy decision module. After fuzzification, fuzzy inference, and defuzzification, the control coefficients of the ED and ET segmentation regions are finally output, and then multiplied by the ED and ET segmentation maps output by the segmentation module to obtain new ED and ET segmentation maps.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The first aspect of the present invention provides a method for glioma MR image segmentation based on edge loss and knowledge fusion decision-making, comprising: a hierarchical module, a segmentation module, a fuzzy decision-making module and an adaptive post-processing module;

[0007] The grading module is used to classify the input image into grades and provide the graded grade information to the fuzzy decision module.

[0008] The segmentation module is used to initially segment the input image into the overall tumor region WT, the tumor core region TC, and the enhanced tumor region ET, and to provide the area ratio information of ET / TC and TC / WT after the initial segmentation to the fuzzy decision module.

[0009] The fuzzy decision-making module receives the hierarchical information from the grading module and the area ratio information from the segmentation module, performs fuzzy decision-making, and outputs the control coefficient λ of the edema region ED. ED and the control coefficient λ for enhancing ET in the tumor region ET Among them, the edema region ED is the difference between the overall tumor region WT and the tumor core region TC;

[0010] Adaptive post-processing module:

[0011] The control coefficient λ output by the fuzzy decision module ED and λ ET Multiplying the predicted values ​​of the edema region ED and the enhanced tumor region ET in the corresponding initial segmentation yields the probability distribution of the edema region ED and the enhanced tumor region ET.

[0012] Voxels of all regions initially segmented into edema region (ED) and enhanced tumor region (ET) are extracted separately. Voxels whose probability of being edema region ED and enhanced tumor region ET is greater than the threshold of 0.5 are used as new edema region ED and new enhanced tumor region ET. The tumor core region TC is the result of subtracting the total tumor region WT from the new edema region ED.

[0013] The second aspect of the present invention provides a method for segmenting MR images of gliomas. When segmenting MR images of gliomas, the deep neural network model used is the aforementioned method for segmenting MR images of gliomas based on edge loss and knowledge fusion decision-making.

[0014] A third aspect of the present invention provides a glioma MR image segmentation device, comprising:

[0015] The memory and a processor coupled to the memory, the processor being configured to execute the program of the glioma MR image segmentation method based on edge loss and knowledge fusion decision based on instructions stored in the memory.

[0016] A fourth aspect of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned glioma MR image segmentation method based on edge loss and knowledge fusion decision-making.

[0017] This invention has outstanding substantive features and significant progress compared to the prior art, specifically:

[0018] 1. The MR image segmentation model for glioma proposed in this invention, based on edge loss and knowledge fusion decision-making, combines a segmentation module, a grading module, a fuzzy decision-making module, and an adaptive post-processing module to improve segmentation performance on limited training data.

[0019] 2. The segmentation model of the present invention designs a fuzzy decision module that can combine hierarchical information and initial segmentation information to make fuzzy decisions based on prior knowledge of gliomas.

[0020] 3. The segmentation model of this invention also proposes a novel edge region voxel-weighted loss function. This loss function achieves the goal of balancing the ratio of positive and negative samples by extracting the edges of the predicted image and the real label as the calculation region of the loss. Furthermore, since the edge region of glioma is very blurry and difficult to distinguish, the distance weighting method is used to separate the difficult-to-distinguish samples, which solves the extreme class imbalance problem in the segmentation network and significantly improves the segmentation accuracy. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this invention, the accompanying drawings used in the embodiments will be briefly described below. Obviously, the drawings shown below are only individual examples of this invention. Researchers in the art can apply this method to glioma MR images through simple reproduction of this invention.

[0022] Figure 1 This is an overall framework diagram of the model in Embodiment 1 of the present invention.

[0023] Figure 2 This is the structure diagram of the edge region voxel weighted loss function mentioned in Embodiment 1 of the present invention.

[0024] Figure 3 This is the training curve of the edge region voxel-weighted loss function mentioned in Embodiment 1 of the present invention.

[0025] Figure 4 This is an application effect diagram of the segmentation method in Embodiment 1 of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0027] Example 1

[0028] The dataset used in this embodiment is BraTS2019, created by the Center for Biomedical Image Computing and Analysis (CBICA) team. The BraTS2019 training set contains 259 high-level (HGG) and 76 low-level (LGG) sequences. Each case consists of four sequences: T1, T2, T1ce, and FLAIR, with each sequence measuring 150×240×240 pixels.

[0029] This embodiment provides a glioma MR image segmentation method based on edge loss and knowledge fusion decision-making. When segmenting glioma MR images, the deep neural network model used is a glioma MR image segmentation model based on edge loss and knowledge fusion decision-making. Wherein, as... Figure 1 As shown, the MR image segmentation model for gliomas based on boundary loss function and knowledge fusion decision-making includes a grading module, a segmentation module, a fuzzy decision-making module, and an adaptive post-processing module.

[0030] Pre-constructed hierarchical task dataset I for model training classWith the segmentation task dataset I seg ;in,

[0031] Segmentation task dataset I seg Construct it as follows:

[0032] Use T1ce, T2, and FLAIR images from the BraTS2019 dataset;

[0033] 1) Crop each image to a size of 155×160×160, then standardize the data. The mean of the standardized data is 0 and the variance is 1.

[0034] 2) Along the Axial direction, slice each standardized image with a step size of 16 and a block length of 32;

[0035] 3) Merge the sliced ​​images, and the image size will become 3×32×160×160;

[0036] Hierarchical Task Dataset I class Construction:

[0037] Use T1ce, T2, and FLAIR images from the BraTS2019 dataset;

[0038] 1) Fill each image with 5 layers of zero values ​​along the Axial direction to change the image size to 160×240×240;

[0039] 2) Crop the image along the Coronal and Sagittal directions with the image center as the origin. The image size will then be 160×160×160.

[0040] 3) Standardize the cropped image. The mean of the standardized data is 0 and the variance is 1.

[0041] 4) Merge the standardized images, and the image size will become 3×160×160×160.

[0042] Hierarchical module

[0043] Used to classify the input image into levels and provide the classified level information to the fuzzy decision module;

[0044] The hierarchical module consists of three 3D encoding modules and one fully connected module connected in a cascaded manner; each 3D encoding module consists of a convolutional layer, a group normalization layer, an activation layer, and a pooling operation; during model training, the constructed dataset I... class Input the data into the hierarchical module, and set the parameters in the hierarchical module as follows:

[0045] Batch size is set to 1; training iterations are set to 100; SGD is used as the optimization method; weight decay is set to 0.0001; momentum parameter is set to 0.99; initial learning rate is set to 1.5 × 10⁻⁶. -3 The learning rate changes as the number of training iterations increases, as shown in the following formula:

[0046]

[0047] Where lr0 is the initial learning rate, and epoch represents the number of training iterations.

[0048] Segmentation Module

[0049] The segmentation module is used to initially segment the input image into the overall tumor region WT, the tumor core region TC, and the enhanced tumor region ET, and to provide the area ratio information of ET / TC and TC / WT after the initial segmentation to the fuzzy decision module.

[0050] The segmentation module uses UNet as the network framework and adds a gated attention mechanism (AG) to the decoder to highlight the organizational structure of the target region and suppress the image features of irrelevant regions, so that when using convolutional neural networks (CNNs), there is no need to use a target localization module.

[0051] When training the model, the constructed dataset I seg Input the data into the segmentation module and set the parameters in the segmentation module as follows:

[0052] Batch size is set to 2; training iterations are set to 100; SGD is used as the optimization method; weight decay is set to 0.0001; momentum parameter is set to 0.99; initial learning rate is set to 1.5 × 10⁻⁶. -3 The learning rate changes as the number of training iterations increases, as shown in the following formula:

[0053]

[0054] Where lr0 is the initial learning rate, and epoch represents the number of training iterations.

[0055] Fuzzy decision module

[0056] It receives the grading information from the grading module and the area ratio information from the segmentation module, performs fuzzy decision-making, and outputs the control coefficient λ of the edema region ED. ED and the control coefficient λ for enhancing ET in the tumor region ETAmong them, the edema region ED is the difference between the overall tumor region WT and the tumor core region TC;

[0057] The fuzzy decision-making module includes fuzzification, a fuzzy rule base, fuzzy inference, and defuzzification. Fuzzification involves processing the input conditional information using membership functions. Fuzzy rules are a knowledge representation method used to represent knowledge in the medical field. Fuzzy inference is the process of deriving a new result from one or more known conditions based on knowledge represented by fuzzy rules. Defuzzification essentially involves finding a precise value that can represent the fuzzy set.

[0058] The specific implementation method of the fuzzy decision-making module:

[0059] 1) Define the input and output fuzzy sets of the fuzzy decision module:

[0060] (1) The hierarchical task dataset I class The image levels are divided into two fuzzy sets: low level (LGG) and high level (HGG);

[0061] (2) Divide the area ratio of TC / WT into three fuzzy sets: small, moderate, and large;

[0062] (3) Divide the area ratio of ET / TC into three fuzzy sets: small, moderate, and large;

[0063] (4) The control coefficients of the output ED and ET are divided into five fuzzy sets: small, less, moderate, more and large;

[0064] 2) Triangular membership functions used in fuzzification:

[0065]

[0066] Where parameters a and c represent the left and right vertices of the lower part of the triangle, respectively, and parameter b represents the upper vertex of the triangle;

[0067] 3) The fuzzy rules used in constructing the fuzzy rule base for fuzzy inference:

[0068] (1) If it is a low level and the TC / WT ratio is small, then the control coefficient of ED is appropriate;

[0069] (2) If it is a low level and the TC / WT ratio is moderate, the control coefficient of ED is small;

[0070] (3) If it is a low level and the TC / WT ratio is large, then the control coefficient of ED is small;

[0071] (4) If it is a high level and the TC / WT ratio is small, then the control coefficient of ED is large;

[0072] (5) If it is a high level and the TC / WT ratio is moderate, then the control of ED is greater;

[0073] (6) If it is a high level and the TC / WT ratio is large, then the control coefficient of ED is appropriate;

[0074] (7) If it is a low level and the ET / TC ratio is small, then the control coefficient of ET is small;

[0075] (8) If it is a low level and the ET / TC ratio is moderate, the control coefficient of ET is small;

[0076] (9) If it is a low level and the ET / TC ratio is large, then the control coefficient of ET is appropriate;

[0077] (10) If it is a high level and the ET / TC ratio is small, then the control coefficient of ET is appropriate;

[0078] (11) If it is a high level and the ET / TC ratio is moderate, the control coefficient of ET is relatively large;

[0079] (12) If it is a high level and the ET / TC ratio is large, then the control coefficient of ET is large;

[0080] 4) Convert the constructed fuzzy rules into a fuzzy rule table, as shown in the table below:

[0081]

[0082] 5) Since the result derived from fuzzy inference still belongs to a fuzzy subset, defuzzification is required to obtain the output control coefficients. This embodiment uses the centroid method for defuzzification. The centroid method provides smoother output inference control; even small changes in the input signal will affect the output. The formula for centroid defuzzification is as follows:

[0083]

[0084] Where, μ v (v k ) is the output quantity v k The membership function of the fuzzy set, where v0 is the precise quantity obtained from fuzzy inference.

[0085] Adaptive post-processing module

[0086] The control coefficient λ output by the fuzzy decision module ED and λ ETMultiplying the predicted values ​​of the edema region ED and the enhanced tumor region ET in the corresponding initial segmentation yields the probability distribution of the edema region ED and the enhanced tumor region ET.

[0087] Voxels of all regions initially segmented into edema region (ED) and enhanced tumor region (ET) are extracted separately. Voxels whose probability of being edema region ED and enhanced tumor region ET is greater than the threshold of 0.5 are used as new edema region ED and new enhanced tumor region ET. The tumor core region TC is the result of subtracting the total tumor region WT from the new edema region ED.

[0088] Custom loss function

[0089] During network training, due to the use of an imbalanced dataset, the gradient may be dominated by excessively large samples. Therefore, this embodiment applies a custom loss function in the segmentation module to adjust the gradient during the segmentation model training process, thereby reducing the weight of easily segmented samples in the loss function and increasing the weight of difficult-to-segment samples, making the model pay more attention to the smaller subset of samples and reducing the impact of the imbalanced dataset.

[0090] The custom loss function is the edge region voxel weighted loss function L. BRD It is designed according to the following steps, such as Figure 2 As shown:

[0091] 1) Using the edge of the Ground Truth as the baseline, obtain the edge of a label by performing edge extraction; calculate the Euclidean distance from each voxel to the edge using the following formula:

[0092]

[0093] Among them, W i Let (x0, y0, z0) represent the distance weight of the i-th voxel, and (x0, y0, z0) represent the points on the label edges. i y i , z i ) represents the i-th voxel in the entire image;

[0094] The final result is a distance-weighted graph;

[0095] 2) Obtaining the voxel-weighted loss function for the edge region: The computational region is calculated using the edge of the Ground Truth as the baseline. An edge region R is obtained by performing edge extraction operations. targets Similarly, the predicted image also obtains an edge region R through this operation. loqits The union of the two regions is taken as the region R for calculating the loss. The edge region is obtained by subtracting the eroded image from the original image, as shown in the following formula:

[0096] Rlogits =I SourceLogits -I ErosionLogits

[0097] R targets =I SourceTargets -I ErosionTargets

[0098] R = R logits ∪R targets

[0099] Among them, I SourceLogits I ErosionLogits These represent the original predicted value and the predicted value after corrosion, respectively; I SourceTargets I ErosionTargets Represent the original true value and the true value after erosion, respectively; R logits R targets R represents the edge region of the predicted value, the edge region of the true value, and the union of the edge regions of the predicted value and the true value, respectively.

[0100] 3) Use the computational region R as a mask to re-filter the predicted values, actual values, and distance-weighted maps within the computational region R;

[0101] 4) The selected predicted values, actual values, and distance-weighted maps are used to calculate the voxel-weighted loss function for the edge region;

[0102] The final edge region voxel-weighted loss function L BRD Written as:

[0103]

[0104] Among them, y ic ∈[0,1] represents the one-hot encoding of voxel i when it is of class c; p ic ∈[0,1] is the probability that voxel i is predicted to be of class c; N is the number of voxels in the region of interest R, C is the total number of classes (including background), and w i It is the dynamic adaptive weight of each voxel;

[0105] 5) When the region-based Dice loss function is used for medical image segmentation, it achieves better segmentation results. To simultaneously utilize L... BRD And Dice's loss, will L BRD It is integrated with the Dice loss function to form a joint loss L DiceBRD , written as:

[0106]

[0107] Where voxels represent the total number of voxels, and ε is a very small number; y ic∈[0,1] represents the one-hot encoding of voxel i when it is of class c; p ic ∈[0,1] is the probability that voxel i is predicted to be of class c; N is the number of voxels in the region of interest R, C is the total number of classes (including background), and w i It is the dynamic adaptive weight of each voxel.

[0108] like Figure 3 As shown, the edge region voxel-weighted loss function proposed in this embodiment exhibits smaller fluctuations in the later stages of training and eventually tends to stabilize; and compared with the Dice loss function and the Dice+WCE loss function, the edge voxel-weighted loss function achieves better results.

[0109] Experimental results

[0110] The experimental platform in this embodiment is an Ubuntu system, an Nvidia RTX 3090Ti graphics card with 24GB of video memory, and an Intel(R) Xeon(R) Silver CPU E5-2678 v3@2.50GHz processor. The experimental model is implemented based on the PyTorch framework.

[0111] The result of the experimental segmentation is shown in the figure below. Figure 4 As shown, the yellow and red areas are closer to the true labels. This is because the yellow area gradually shrinks, eliminating false positives that enhance the tumor, while the red area gradually expands, with some voxels in the edema region being replaced by the tumor core, which increases the number of true positives in the tumor core. Ultimately, this makes the segmentation result closer to the true labels.

[0112] Example 2

[0113] This embodiment provides a glioma MR image segmentation device, including:

[0114] The memory and a processor coupled to the memory, the processor being configured to execute the program of the glioma MR image segmentation model based on edge loss and knowledge fusion decision as described in Embodiment 1, based on instructions stored in the memory.

[0115] The memory may include, for example, system memory, fixed non-volatile storage media, etc. System memory may store, for example, the operating system, application programs, the boot loader, and other programs.

[0116] The device may also include input / output interfaces, network interfaces, and storage interfaces. These interfaces, as well as the memory and processor, can be connected via, for example, a bus. The input / output interfaces provide connection interfaces for input / output devices such as monitors, mice, keyboards, and touchscreens. The network interfaces provide connection interfaces for various networked devices. The storage interfaces provide connection interfaces for external storage devices such as SD cards and USB flash drives.

[0117] Example 3

[0118] This embodiment provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the glioma MR image segmentation model based on edge loss and knowledge fusion decision-making as described in Embodiment 1.

[0119] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-non-transitory readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer program code.

[0120] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] The above-described embodiments merely illustrate some implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for glioma MR image segmentation based on edge loss and knowledge fusion decision-making, characterized in that, include: The module includes a hierarchical module, a segmentation module, a fuzzy decision-making module, and an adaptive post-processing module. The grading module is used to classify the input image into grades and provide the graded grade information to the fuzzy decision module. The segmentation module is used to initially segment the input image into the overall tumor region WT, the tumor core region TC, and the enhanced tumor region ET, and to provide the area ratio information of ET / TC and TC / WT after the initial segmentation to the fuzzy decision module. The fuzzy decision-making module receives the hierarchical information from the grading module and the area ratio information from the segmentation module, performs fuzzy decision-making, and outputs the control coefficient λ of the edema region ED. ED and the control coefficient λ for enhancing ET in the tumor region ET Among them, the edema region ED is the difference between the overall tumor region WT and the tumor core region TC; The fuzzy decision-making module includes fuzzification, a fuzzy rule base, fuzzy reasoning, and defuzzification. Define the input and output fuzzy sets of the fuzzy decision module: 1) The hierarchical task dataset I class The image levels are divided into two fuzzy sets: low level and high level; 2) Divide the area ratio of TC / WT into three fuzzy sets: small, moderate, and large; 3) Divide the area ratio of ET / TC into three fuzzy sets: small, moderate, and large; 4) The control coefficients of the outputs ED and ET are divided into five fuzzy sets: small, relatively small, moderate, relatively large, and large; In fuzzification, triangular membership functions are used for fuzzification; In deblurring, the centroid method is used. Adaptive post-processing module: The control coefficient λ output by the fuzzy decision module ED and λ ET Multiplying the predicted values ​​of the edema region ED and the enhanced tumor region ET in the corresponding initial segmentation yields the probability distribution of the edema region ED and the enhanced tumor region ET. Voxels of all regions initially segmented into edema region (ED) and enhanced tumor region (ET) are extracted separately. Voxels whose probability of being edema region ED and enhanced tumor region ET is greater than the threshold of 0.5 are used as new edema region ED and new enhanced tumor region ET. The tumor core region TC is the result of subtracting the total tumor region WT from the new edema region ED.

2. The glioma MR image segmentation method based on edge loss and knowledge fusion decision-making according to claim 1, characterized in that: The hierarchical module consists of three 3D encoding modules and one fully connected module connected in a cascaded manner; each 3D encoding module consists of a convolutional layer, a group normalization layer, an activation layer, and a pooling operation. When training the model, set the batch size in the hierarchical module to 1; The number of training iterations was set to 100; the optimization method used was SGD; the weight decay was set to 0.0001; the momentum parameter was set to 0.99; and the initial learning rate was set to 1.5 × 10⁻⁶. -3 The learning rate changes as the number of training iterations increases, as shown in the following formula: lr = lr 0×(1-epoch / 100) 0.9 in, lr 0 represents the initial learning rate, and epoch represents the number of training iterations completed.

3. The glioma MR image segmentation method based on edge loss and knowledge fusion decision-making according to claim 1, characterized in that: The segmentation module uses UNet as the network framework and adds a gated attention mechanism (AG) to the decoder to highlight the organizational structure of the target region and suppress image features of irrelevant regions. When training the model, set the batch size in the segmentation module to 2; The number of training iterations was set to 100; the optimization method used was SGD; the weight decay was set to 0.0001; the momentum parameter was set to 0.99; and the initial learning rate was set to 1.5 × 10⁻⁶. -3 The learning rate changes as the number of training iterations increases, as shown in the following formula: lr = lr 0×(1-epoch / 100) 0.9 in, lr 0 represents the initial learning rate, and epoch represents the number of training iterations completed.

4. The glioma MR image segmentation method based on edge loss and knowledge fusion decision-making according to claim 1, characterized in that, The fuzzy rules used in constructing the fuzzy rule base for fuzzy inference: 1) If it is a low level and the TC / WT ratio is small, then the control factor of ED is appropriate; 2) If it is a low level and the TC / WT ratio is moderate, the control coefficient of ED will be small; 3) If it is a low level and the TC / WT ratio is large, then the control coefficient of ED is small; 4) If it is a high level and the TC / WT ratio is small, then the control coefficient of ED is large; 5) If it is a high level and the TC / WT ratio is moderate, then the control of ED is greater; 6) If it is a high level and the TC / WT ratio is large, then the control factor of ED is appropriate; 7) If it is a low level and the ET / TC ratio is small, then the control coefficient of ET is small; 8) If it is a low level and the ET / TC ratio is moderate, then the control coefficient of ET is relatively small; 9) If it is a low level and the ET / TC ratio is large, then the control factor for ET is appropriate; 10) If it is a high level and the ET / TC ratio is small, then the control factor for ET is appropriate; 11) If it is a high level and the ET / TC ratio is moderate, then the control coefficient of ET is relatively large; 12) If it is a high level and the ET / TC ratio is large, then the control coefficient of ET is large.

5. The glioma MR image segmentation method based on edge loss and knowledge fusion decision-making according to claim 1, characterized in that, The segmentation module uses the edge region voxel-weighted loss function L BRD With Dice loss function L Dice Perform gradient optimization; where, Voxel-weighted loss function L for edge regions BRD for: Among them, y ic ∈[0,1] represents the one-hot encoding of voxel i when it is of class c; p ic ∈[0,1] is the probability that voxel i is predicted to be of class c; N is the number of voxels in the region of interest R, C is the total number of classes, and w i It is the dynamic adaptive weight of each voxel; Voxel-weighted loss function L for edge regions BRD Joint loss L with Dice loss function DiceBRD for: Where voxels represent the total number of voxels, and ε is a very small number; y ic ∈[0,1] represents the one-hot encoding of voxel i when it is of class c; p ic ∈[0,1] is the probability that voxel i is predicted to be of class c; N is the number of voxels in the region of interest R, C is the total number of classes, and w i It is the dynamic adaptive weight of each voxel.

6. The glioma MR image segmentation method based on edge loss and knowledge fusion decision-making according to claim 1, characterized in that: Pre-constructed hierarchical task dataset I for model training class With the segmentation task dataset I seg ;in, Segmentation task dataset I seg Construct it as follows: Use T1ce, T2, and FLAIR images from the BraTS2019 dataset; Each image is cropped to 155×160×160, and then standardized. The mean of the standardized data is 0 and the variance is 1. Along the Axial direction, slice the standardized image with a step size of 16 and a block length of 32. Merging the sliced ​​images will result in a size of 3×32×160×160. Hierarchical Task Dataset I class Construction: Use T1ce, T2, and FLAIR images from the BraTS2019 dataset; Each image is filled with 5 layers of zero values ​​along the Axial direction, and the size will become 160×240×240; By cropping along the Coronal and Sagittal directions with the center of each image as the origin, the size will become 160×160×160; The cropped image is standardized, and the mean of the standardized data is 0 and the variance is 1. The standardized images are then merged, and the image size becomes 3×160×160×160.

7. A method for segmenting MR images of gliomas, characterized in that: When segmenting MR images of gliomas, the deep neural network model used is the MR image segmentation method for gliomas based on edge loss and knowledge fusion decision-making as described in any one of claims 1-6.

8. A glioma MR image segmentation device, comprising: The memory and a processor coupled to the memory, the processor being configured to execute a program of the glioma MR image segmentation method based on edge loss and knowledge fusion decision as described in any one of claims 1-6, based on instructions stored in the memory.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the glioma MR image segmentation method based on edge loss and knowledge fusion decision-making as described in any one of claims 1-6.

Images