Brain glioma grade detection method, device and equipment based on brain MRI (Magnetic Resonance Imaging), and medium

Abstract

The invention belongs to the technical field of computer vision, and discloses a brain glioma grade detection method, device and equipment based on brain MRI, and a medium, and the method comprises the steps: obtaining a plurality of sequences of brain MRI images based on a target case; performing a registration operation to align azimuth information of the brain MR I images of all sequences; for each brain MRI image, segmenting a brain area, and obtaining skull-removed MRI images of each sequence; performing feature fusion on the selected skull-removed MRI image, and then combining the selected skull-removed MRI image with a target skull-removed MRI image to obtain an MR I image set; the MRI image set is input into a brain glioma segmentation model, brain glioma high-order features are obtained, and the brain glioma segmentation model obtains the brain glioma high-order features when segmenting a brain glioma area; and inputting the brain glioma high-order features into a brain glioma grade classification model to obtain brain glioma grade information. The accuracy of brain glioma segmentation and detection can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and specifically relates to a method, device, equipment and medium for detecting the grade of glioma based on brain MRI. Background Art

[0002] Gliomas are the most common primary intracranial malignant tumors, which can be divided into grades I-IV according to the degree of malignancy of the tumor. The malignancy increases with the grade. The treatment and prognosis of gliomas of different grades and different gene mutations are different. Therefore, if the tumor area and tumor grade can be accurately segmented and judged before surgical treatment, it will help guide the treatment plan and the selection of surgical resection area, which is of great value in improving the treatment effect and prognosis of patients.

[0003] The human eye's observation of MRI (magnetic resonance imaging) stays at the visual surface, resulting in MRI files containing a lot of information, but difficult to extract; and accurate pathological diagnosis requires surgical collection of patient samples, which are determined by pathologists based on the morphology of tissue sections, morphological characteristics of cells, cell proliferation activity, and other specific markers. Therefore, the accuracy of non-invasive examination methods is insufficient to support preoperative diagnosis of brain glioma grade; and invasive detection methods have a lag, and a surgical sampling before surgery not only increases the degree of patient trauma, but also requires more medical resources.

[0004] Although deep learning is widely used in medical images and can learn deep features between medical images to achieve tasks such as image classification and target detection, the existing network models are still not accurate enough for the segmentation and detection of brain gliomas, making it difficult to meet the requirements of preoperative non-invasive diagnosis. Summary of the invention

[0005] The object of the present invention is to provide a method, device, equipment and storage medium for detecting the grade of glioma based on brain MRI, which can improve the accuracy of glioma segmentation and detection.

[0006] The first aspect of the present invention discloses a method for detecting the grade of glioma based on brain MRI, comprising:

[0007] Based on the target case, multiple sequences of brain MRI images are acquired;

[0008] Performing a registration operation to align the position information of the brain MRI images of all sequences;

[0009] For each of the brain MRI images, the brain region is segmented to obtain a post-skull removal MRI image of each sequence;

[0010] Performing feature fusion on the selected MRI image after skull removal, and then combining it with the target MRI image after skull removal to obtain an MRI image set;

[0011] Inputting the MRI image set into a glioma segmentation model to obtain high-order features of the glioma, wherein the glioma segmentation model obtains the high-order features of the glioma when segmenting the glioma region;

[0012] The high-order features of the glioma are input into a glioma grade classification model to obtain glioma grade information.

[0013] In some embodiments, the glioma segmentation model is constructed based on a 2D segmentation network, and the MRI image set is input into the glioma segmentation model to obtain high-order features of the glioma, including:

[0014] In the glioma segmentation model, each MRI image in the MRI image set is sliced ​​in 2D to obtain multiple 2D images corresponding to each MRI image, extract features of each 2D image, and stack the features of all 2D images according to the hierarchical structure of the MRI image to which the 2D image belongs to obtain a glioma region; the feature extracted from the last layer of each 2D image during downsampling is set as the first glioma feature, the first glioma features corresponding to all 2D images of each MRI image are combined into the second glioma features corresponding to each MRI image, and all the second glioma features in the MRI image set constitute the glioma high-order feature.

[0015] In some embodiments, combining the first brain glioma features corresponding to all 2D images of the MRI image into the second brain glioma features corresponding to the MRI image includes:

[0016] All of the first brain glioma features are combined by using a numerical calculation method to fuse the first brain glioma features at different levels to obtain the second brain glioma features.

[0017] In some embodiments, the multiple sequences of brain MRI images include brain MRI images of FLAIR sequence, T1 sequence, T1 GD sequence and T2 sequence, feature fusion is performed on the selected post-skull removal MRI image, and then combined with the target post-skull removal MRI image to obtain an MRI image set, including:

[0018] Inputting brain MRI images of T1 GD sequence, FLAIR sequence, T1 sequence and T2 sequence into a densely connected convolutional network, and extracting features of each sequence at the position of the first feature extraction unit of the densely connected convolutional network;

[0019] The features of the T1 GD sequence are respectively fused with the features of the FLAIR sequence, the features of the T1 sequence, and the features of the T2 sequence to obtain a fused T1 sequence, a fused T2 sequence, and a fused FLAIR sequence;

[0020] The fused T1 sequence, the fused T2 sequence, the fused FLAIR sequence and the FLAIR sequence are combined to obtain the MRI image set.

[0021] In some embodiments, when performing a registration operation, the registration index is the mean square error, the optimizer uses a gradient descent method, and the interpolation method is a linear interpolation method.

[0022] The second aspect of the present invention discloses a brain glioma grade detection device based on brain MRI, comprising:

[0023] The MRI processing module is used to perform a registration operation to align the orientation information of all sequences of brain MRI images; for each of the brain MRI images, the brain region is segmented to obtain each sequence of MRI images after skull removal; feature fusion is performed on the selected MRI image after skull removal, and then combined with the target MRI image after skull removal to obtain an MRI image set;

[0024] A glioma segmentation module, used for inputting the MRI image set into a glioma segmentation model to obtain high-order features of the glioma, wherein the glioma segmentation model obtains the high-order features of the glioma when segmenting the glioma region;

[0025] The glioma grade classification module is used to input the glioma high-order features into the glioma grade classification model to obtain glioma grade information.

[0026] In some embodiments, the glioma segmentation model is constructed based on a 2D segmentation network, and the glioma segmentation model includes a slicing unit and a feature extraction unit. The slicing unit is used to perform 2D slicing on each MRI image in the MRI image set, obtain multiple 2D images corresponding to each MRI image, extract features of each 2D image, and stack the features of all 2D images according to the hierarchical structure of the MRI image to which the 2D image belongs to obtain a glioma area; the feature extraction unit is used to set the feature extracted from the last layer of each 2D image during downsampling as the first glioma feature, and combine the first glioma features corresponding to all 2D images of each MRI image into second glioma features corresponding to each MRI image, and all second glioma features in the MRI image set constitute the glioma high-order feature.

[0027] In some embodiments, it also includes a data collection module and a training module, wherein the data collection module is used to collect multiple sequences of brain MRI images of the case, and the training module is used to train the glioma segmentation model and the glioma grade classification model.

[0028] The third aspect of the present invention discloses an electronic device, comprising a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the brain glioma grade detection method based on brain MRI disclosed in the first aspect.

[0029] The fourth aspect of the present invention discloses a computer-readable storage medium, which stores a computer program, wherein the computer program enables a computer to execute the brain glioma grade detection method based on brain MRI disclosed in the first aspect.

[0030] The beneficial effect of the present invention is that by registering the brain MRI image and removing the skull, the interference in the MRI image is greatly removed. By performing a more detailed segmentation of the MRI image through the glioma segmentation model, the tumor position, size and shape can be located, and the precise glioma area can be obtained, which provides accurate features for the subsequent glioma grading, and at the same time outputs the high-order features of the glioma area. The glioma grade classification model uses these high-order features to obtain accurate glioma grade information. Accurate segmentation and detection of gliomas are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings herein show specific examples of the technical solutions described in the present invention, and together with the specific implementation methods, constitute a part of the specification, and are used to explain the technical solutions, principles and effects of the present invention.

[0032] Unless otherwise specified or defined, the same reference numerals in different drawings represent the same or similar technical features, and the same or similar technical features may also be represented by different reference numerals.

[0033] Figure 1 It is an architecture diagram of a brain glioma grade detection system based on brain MRI disclosed in an embodiment of the present invention;

[0034] Figure 2 It is a flow chart of a method for detecting the grade of glioma based on brain MRI disclosed in an embodiment of the present invention;

[0035] Figure 3 yes Figure 2 A schematic diagram of the functional architecture of an embodiment;

[0036] Figure 4 yes Figure 2Flow chart of step S400 in the embodiment;

[0037] Figure 5 is a structural schematic diagram of a brain glioma grade detection device based on brain MRI according to an embodiment of the present invention;

[0038] Figure 6 It is a structural schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] Unless otherwise specified or defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art. In the case of combining the technical solution of the present invention with realistic scenarios, all technical and scientific terms used herein may also have meanings corresponding to the purpose of implementing the technical solution of the present invention. The "first, second..." used herein is only used to distinguish the names and does not represent a specific quantity or order. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0040] It should be noted that when a component is considered to be "fixed to" another component, it can be directly fixed to the other component or there can be a central component; when an component is considered to be "connected to" another component, it can be directly connected to the other component or there can be a central component at the same time; when an component is considered to be "installed on" another component, it can be directly installed on the other component or there can be a central component at the same time. When an component is considered to be "set on" another component, it can be directly set on the other component or there can be a central component at the same time.

[0041] Unless otherwise specified or defined, the "said" and "the" used in this document refer to the technical features or technical contents mentioned or described before the corresponding position, and the technical features or technical contents may be the same as or similar to the technical features or technical contents mentioned therein. In addition, the terms "including" and "having" and any variations thereof used in this document are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices.

[0042] At present, the accuracy of using MRI visual features to judge the grade of brain glioma is not enough, and it is difficult to meet the requirements of preoperative non-invasive diagnosis. The surgical pathology results are still the gold standard. However, using the patient's multi-sequence MRI images to perform non-invasive examinations before surgery and accurately obtain the patient's brain glioma tumor grade is of great significance for guiding doctors to diagnose and treat cases.

[0043] In order to achieve non-invasive and accurate determination of the grade of glioma, the present invention first aligns the brain MRI images and removes the skull, which greatly removes the interference in MRI. Then, feature fusion is performed to obtain the distribution of enhanced tumor areas from the serial MRI images, better highlighting the features and pixel contrast of the tumor area, and then the high-order features of glioma are extracted through the glioma segmentation model. Using these high-order features, the glioma grade classification model can obtain accurate glioma grade classification.

[0044] like Figure 1 As shown, this embodiment deploys a brain glioma grade detection system based on brain MRI, including a user terminal, an algorithm layer, a hardware layer and a data collection terminal. The data collection terminal is deployed on each hospital platform, collects MRI image data of glioma patients in each hospital, and uploads the data to the data center to form a multi-center glioma data set. The algorithm program and model in the algorithm layer are deployed on the computing server, and the public data set is used for initial pre-training to save the training weight. Then, the multi-center glioma data set collected by the data collection terminal is used for testing and retraining to further improve the generalization ability of the algorithm and model. A test port is opened on the computing server. The user terminal can view the results of the algorithm layer by using the test port, and decide whether to retrain until the user terminal usage requirements are met. The weights of the algorithms and models that meet the user terminal usage requirements are frozen, saved in the storage space of this user, and directly called by the user terminal. When in use, the user terminal only needs to upload the patient data and call the port to derive the glioma grade based on the MRI image. It is easy to use, can meet the usage requirements of each user, and can be personalized according to the data characteristics of each user terminal. Training.

[0045] Among them, when the algorithm program and model in the algorithm layer run on the computing server, a brain glioma grade detection method based on brain MRI is implemented. Figure 2 and Figure 3 , the specific steps include:

[0046] Step S100: acquiring multiple sequences of brain MRI images based on a target case;

[0047] Perform an MRI scan on the target case to obtain a brain MRI image. Since the anatomical structure and pathological changes of different patients may be different, in order to meet different clinical needs and imaging purposes, multiple sequences of brain MRI images are generally obtained during MRI scanning. Therefore, it is necessary to select a suitable sequence according to the specific situation. In this embodiment, four sequences of MRI images are selected according to the correspondence between the tumor area and the sequence and the pixel value level to be distinguished: T1-weighted sequence (T1), T2-weighted sequence (T2), FLAIR sequence (FLAIR) and T1-weighted enhanced sequence (T1 GD).

[0048] Step S200: performing a registration operation to align the orientation information of all sequences of brain MRI images;

[0049] The registration operation is used to align the orientation information of multiple selected sequences of brain MRI images, which can improve the comparability of MRI images. In this embodiment, the T1 GD sequence is selected as the registration reference, and the other three sequences are all corresponding to the position information of the T1 GD sequence. During the registration calculation, the mean square error is selected as the registration index, the optimizer selects the gradient descent method, and the interpolation method selects the linear interpolation with a faster calculation speed. After the registration is completed, the orientation information between different sequences is completely unified. Among them, the registration algorithm of MRI data is a conventional technical means in this field, which will not be repeated here.

[0050] Step S300: for each brain MRI image, segment the brain region to obtain the MRI images after skull removal of each sequence;

[0051] Brain MRI images are three-dimensional data, containing images of multiple detection surfaces, and each image has a high resolution and strong three-dimensionality, so the number of pixels is extremely large. When detecting brain MRI, not only the human brain area is covered, but also the entire head is scanned, so brain MRI also includes the skull and the area below the brain. On the one hand, these areas will bring noise interference to the automatic analysis of the brain, and on the other hand, they will increase the consumption of computing resources. Therefore, it is necessary to segment the brain area in the brain MRI image.

[0052] In MRI images, the brain area is large and has obvious boundaries; and there are only two types of segmentation in MRI images: brain area and background area, so the segmentation complexity is not high. A convolutional neural network can be used as the backbone to build a deep neural network as a brain segmentation model, and the brain area is marked as a segmentation label on the registered MRI image to form a training data set to train the brain segmentation model. Then, the trained brain segmentation model is used to segment the brain area in the brain MRI image, remove irrelevant tissues such as the skull and mandible, and obtain the MRI images after de-skulling of each sequence.

[0053] Since HD-BET is an efficient brain segmentation tool specially designed for brain segmentation of multi-sequence MRI data, this embodiment directly uses the built-in 3DCNN segmentation model and its model weights of HD-BET to efficiently segment the brain area, remove the skull, and obtain the MRI images of each sequence after the skull is removed.

[0054] Step S400: performing feature fusion on the selected MRI image after skull removal, and then combining it with the target MRI image after skull removal to obtain an MRI image set;

[0055] Since the enhanced tumor area is obviously clear in the T1GD sequence, the features of the T1GD sequence are fused on the three sequences of T1, T2, and FLAIR to complete the feature fusion between different sequences. That is, the selected post-skull removal MRI images are the post-skull removal MRI images of the T1 sequence, T2 sequence, and FLAIR sequence. The post-skull removal MRI images of the T1 GD sequence are feature fused with the post-skull removal MRI images of the other three sequences to obtain T1-T1GD images, T2-T1 GD images, and FLAIR R-T1GD images, so that the distribution of the enhanced tumor area can also be obtained in the other three sequences, and the features and pixel contrast of the tumor area can be better highlighted. The target post-skull removal MRI image is the post-skull removal MRI image of the FLAIR sequence. The fused T1-T1 GD images, T2-T1GD images, and FLAIR-T1GD images, together with the original post-skull removal MRI images of the FLAIR sequence, form an MRI image set.

[0056] like Figure 4 As shown, the specific steps include:

[0057] Step S410: inputting the MRI images of the T1 GD sequence, FLAIR sequence, T1 sequence and T2 sequence after skull removal into a densely connected convolutional network, and extracting the features of each sequence at the position of the first feature extraction unit of the densely connected convolutional network;

[0058] Step S420: fusing the features of the T1GD sequence with the features of the FLAIR sequence, the features of the T1 sequence, and the features of the T2 sequence, respectively, to obtain a fused T1 sequence, a fused T2 sequence, and a fused FLAIR sequence;

[0059] Step S430: combining the fused T1 sequence, the fused T2 sequence, the fused FLAIR sequence, and the MRI image after skull removal of the FLAIR sequence to obtain an MRI image set.

[0060] Specifically, Torchvision is a PyTorch library for computer vision tasks. This embodiment uses the built-in DenseNet network model (densely connected convolutional network) of the Torchvision library to extract the features of each sequence at the position of the first feature extraction unit of the densely connected convolutional network, such as the shape features of the lesion area, such as area, perimeter, shape factor, etc.

[0061] The features of the T1GD sequence are fused on the T1 sequence, T2 sequence and FLAIR sequence, for example, the features of the T1 sequence are spliced ​​with the features of the T1GD sequence, the features of the T2 sequence are spliced ​​with the features of the T1GD sequence, and the features of the FLAIR sequence are spliced ​​with the features of the T1 GD sequence, to complete the feature fusion between different sequences, and obtain the fused T1 sequence T1-T1GD, the fused T2 sequence T2-T1 GD, and the fused FLAIR sequence FLAIR R-T1 GD, plus the original FLAIR R sequence MRI image after skull removal, to form an MRI image set. It should be noted that the method of feature fusion is not limited, and various existing methods can be used for feature fusion, such as weighted averaging or building a fusion model for feature fusion.

[0062] Step S500: inputting the MRI image set into the glioma segmentation model to obtain high-order features of the glioma, wherein the glioma segmentation model obtains the high-order features of the glioma when segmenting the glioma region;

[0063] The MRI image set is input into the glioma segmentation model, which can accurately segment the shape and size of the tumor area, indicating that the features extracted by the glioma segmentation model are accurate descriptions of the glioma area features. Therefore, this embodiment directly uses the features extracted by the glioma segmentation model without creating a model specifically for feature extraction.

[0064] Specifically, this embodiment regards the MRI image as a stereoscopic image composed of multiple plane stereos, and each 2D image is called a slice. The glioma segmentation model does not directly use a 3D network, but is built based on a 2D segmentation network (such as nnU-Net), which includes a downsampling stage and an upsampling stage. Among them, nnU-Net is a deep learning model for image segmentation tasks, and its structural design is intended to accurately segment different areas in the image. Moreover, by using the U-Net framework, the problems of insufficient computing power and low precision of shallow models can be solved.

[0065] After the MRI image set is input into the glioma segmentation model, each MRI image in the MRI image set is sliced ​​in 2D in the glioma segmentation model to obtain multiple 2D images corresponding to each MRI image. The U-Net framework network extracts the features of each 2D image and stacks the features of all 2D images according to the hierarchical structure of the MRI image to which the 2D image belongs, thereby restoring the original 3D structure and obtaining the segmented glioma area.

[0066] The glioma segmentation model can complete the segmentation of the tumor area, proving that it focuses on the tumor area during the downsampling process, and the extracted features can accurately express the tumor area and its different tumor types. Therefore, the features extracted from each 2D image at the last layer of downsampling (i.e., the high-order features of the tumor extracted at the last layer of downsampling during forward calculation) are set as the first glioma features, and the first glioma features corresponding to all 2D images of each MRI image are respectively processed to obtain the second glioma features corresponding to each MRI image. All the second glioma features in the MRI image set constitute the high-order features of gliomas.

[0067] When combining the first glioma features into the second glioma features, all the first glioma features are merged by numerical calculation (such as weighted mean, data standardization, etc.) to obtain the second glioma features. In this way, the size of the first glioma features can be kept unchanged, the fusion of the first glioma features at different levels can be completed, and the size consistency of the second glioma features can be ensured.

[0068] By using the 3D segmentation technology of 2D slicing of 3D images, we can focus on the tumor area of ​​glioma, obtain the segmented tumor area map, and extract the high-order tumor features of the tumor area map. Among them, the high-order tumor features are the data distribution obtained after the MRI underlying feature data is calculated by the deep learning model. In the form of a tensor of fixed dimension, it contains the visual information and orientation information of the MRI image.

[0069] Step S600: inputting the high-order features of glioma into the glioma grade classification model to obtain glioma grade information.

[0070] In the present embodiment, the main body of the glioma grade classification model is EfficientNet model (a kind of efficient convolutional neural network), and there are four feature extraction units. After the glioma high-order features are input into the glioma grade classification model, the glioma grade classification model continues to reduce the dimension of the glioma high-order features input, extracts features, obtains more refined data distribution, and finally uses the classification layer to complete the classification, and obtains the judgment result of the glioma grade, i.e., one of the three classifications of WHO grades such as tumor level 2, tumor level 3, and tumor level 4. It is easy to understand that the glioma grade classification model can also be constructed based on other multi-classification models, such as support vector machine, decision tree, random forest, etc.

[0071] The above-mentioned glioma segmentation model and glioma grade classification model need to be trained before use. The training process of this embodiment is: collect MRI image data of patients who have undergone surgical pathology diagnosis in the past to form a multi-center glioma data set. After the MRI images in the glioma data set are registered, skull removed and feature fusion operations in steps S200 to S400, the sequence registration of each patient is completed and the skull is removed. The obtained MRI image set is selected according to the ratio of 4:1, and the case data is annotated according to the glioma tumor grade obtained by surgical pathology diagnosis, and the tumor area included is annotated and refined. The whole tumor (WT), enhanced tumor (ET) and tumor core (TC) are annotated in order, and the three annotations are merged into a new MRI file. The annotated case data is divided into a training set, a validation set and a test set. The training set and the validation set are used to train the glioma segmentation model to obtain the trained glioma segmentation model, and the performance of the glioma segmentation model is verified by the test set. During training and testing, the average IOU (Intersection over Union) value of the three categories and the Dice coefficient are used as the most important indicators for evaluating the glioma segmentation model. IOU is an indicator to measure the degree of overlap between two regions, which is usually used to evaluate the quality of segmentation results. The Dice coefficient is an indicator to measure the similarity between two samples, especially for image segmentation.

[0072] The high-order features of each case obtained in the glioma segmentation model are used as training data for glioma grade classification, and the annotated glioma grade information is used as the classification label to train the glioma grade classification model. The trained glioma grade classification model can accurately obtain three classification results after calculating the high-order features of any case: tumor grade 2, tumor grade 3, and tumor grade 4.

[0073] Using the above-mentioned trained model, the tumor grade can be accurately obtained before surgery without surgical trauma, providing important support for doctors to specify subsequent diagnostic plans.

[0074] In summary, by automatically processing brain MRI images, we can obtain MRI images with registration completed and skull removed, which greatly removes interference in MRI images. The processed MRI images are input into the glioma segmentation model for more detailed segmentation to obtain the tumor area in the brain, which can locate the tumor position, size and shape, and provide accurate features for subsequent tumor grading. At the same time, the high-order features of the tumor area are output. The glioma grade classification model uses these high-order features to obtain accurate glioma grade classification. Compared with the current operation process that requires surgical sampling and slicing, manual film reading, collection of molecular information, and then judgment, it greatly reduces the operation time and greatly improves efficiency. While having batch operation capabilities, it also has the same accuracy and reliability, and has good clinical application value; it can also accurately determine the grade of glioma, transforming from invasive examination methods to non-invasive examination methods, saving medical resources, and for patients, it can reduce the surgical burden of patients and protect their health.

[0075] like Figure 5 As shown, based on the above-mentioned brain glioma grade detection method based on brain MRI, the embodiment of the present invention discloses a brain glioma grade detection device based on brain MRI, which runs on the above-mentioned computing server, including:

[0076] The MRI processing module 600 is used to perform a registration operation to align the orientation information of all sequences of brain MRI images; for each of the brain MRI images, segment the brain region to obtain each sequence of MRI images after skull removal; perform feature fusion on the selected MRI image after skull removal, and then combine it with the target MRI image after skull removal to obtain an MRI image set;

[0077] A glioma segmentation module 610 is used to input the MRI image set into a glioma segmentation model to obtain high-order features of the glioma. The glioma segmentation model obtains the high-order features of the glioma when segmenting the glioma region.

[0078] The glioma grade classification module 620 is used to input the glioma high-order features into the glioma grade classification model to obtain glioma grade information.

[0079] In some embodiments, the glioma segmentation model is constructed based on a 2D segmentation network, and the glioma segmentation model includes a slicing unit and a feature extraction unit, wherein the slicing unit is used to perform 2D slicing on each MRI image in the MRI image set, obtain multiple 2D images corresponding to each MRI image, extract features of each 2D image, and stack the features of all 2D images according to the hierarchical structure of the MRI image to which the 2D image belongs to obtain a glioma region; the feature extraction unit is used to set the feature extracted from the last layer of each 2D image during downsampling as the first glioma feature, and respectively combine the first glioma features corresponding to all 2D images of each MRI image into second glioma features corresponding to each MRI image, and all second glioma features in the MRI image set constitute the glioma high-order feature;

[0080] In some embodiments, it also includes a data collection module and a training module, wherein the data collection module is used to collect multiple sequences of brain MRI images of the case, and the training module is used to train the glioma segmentation model and the glioma grade classification model.

[0081] like Figure 6 As shown, an embodiment of the present invention discloses an electronic device, including a memory 401 storing executable program codes and a processor 402 coupled to the memory 401;

[0082] The processor 402 calls the executable program code stored in the memory 401 to execute the brain glioma grade detection method based on brain MRI described in the above embodiments.

[0083] An embodiment of the present invention further discloses a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the brain glioma grade detection method based on brain MRI described in the above embodiments.

[0084] The purpose of the above embodiments is to exemplarily reproduce and deduce the technical solution of the present invention, and to fully describe the technical solution, purpose and effect of the present invention. Its purpose is to make the public understand the disclosed content of the present invention more thoroughly and comprehensively, and it does not limit the scope of protection of the present invention.

[0085] The above embodiments are not exhaustive enumerations of the present invention, and there may be multiple other implementations not listed. Any replacement and improvement made without violating the concept of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for detecting the grade of glioma based on brain MRI, characterized in that: include: Based on the target case, multiple sequences of brain MRI images are acquired; Performing a registration operation to align the position information of the brain MRI images of all sequences; For each of the brain MRI images, the brain region is segmented to obtain a post-skull removal MRI image of each sequence; Performing feature fusion on the selected MRI image after skull removal, and then combining it with the target MRI image after skull removal to obtain an MRI image set; Inputting the MRI image set into a glioma segmentation model to obtain high-order features of the glioma, wherein the glioma segmentation model obtains the high-order features of the glioma when segmenting the glioma region; The high-order features of the glioma are input into a glioma grade classification model to obtain glioma grade information.

2. The method for detecting the grade of glioma based on brain MRI according to claim 1, characterized in that: The glioma segmentation model is constructed based on a 2D segmentation network. The MRI image set is input into the glioma segmentation model to obtain high-order features of glioma, including: In the glioma segmentation model, each MRI image in the MRI image set is sliced ​​in 2D to obtain multiple 2D images corresponding to each MRI image, and features of each 2D image are extracted. The features of all 2D images are stacked according to the hierarchical structure of the MRI image to which the 2D image belongs to obtain the glioma region; the feature extracted from the last layer of each 2D image during downsampling is set as the first glioma feature, and the first glioma features corresponding to all 2D images of each MRI image are respectively combined into the second glioma features corresponding to each MRI image, and all the second glioma features in the MRI image set constitute the glioma high-order feature.

3. The method for detecting the grade of glioma based on brain MRI according to claim 2, characterized in that: Combining the first brain glioma features corresponding to all 2D images of the MRI image into the second brain glioma features corresponding to the MRI image, including: All of the first brain glioma features are combined by using a numerical calculation method to fuse the first brain glioma features at different levels to obtain the second brain glioma features.

4. The method for detecting the grade of glioma based on brain MRI according to claim 1, characterized in that: The brain MRI images of the plurality of sequences include FLAIR sequence, T1 sequence, T1 GD sequence and T2 sequence brain MRI images, feature fusion is performed on the selected post-skull removal MRI image, and then combined with the target post-skull removal MRI image to obtain an MRI image set, including: Inputting brain MRI images of T1 GD sequence, FLAIR sequence, T1 sequence and T2 sequence into a densely connected convolutional network, and extracting features of each sequence at the position of the first feature extraction unit of the densely connected convolutional network; The features of the T1 GD sequence are respectively fused with the features of the FLAIR sequence, the features of the T1 sequence, and the features of the T2 sequence to obtain a fused T1 sequence, a fused T2 sequence, and a fused FLAIR sequence; The fused T1 sequence, the fused T2 sequence, the fused FLAIR sequence and the FLAIR sequence are combined to obtain the MRI image set.

5. The method for detecting the grade of glioma based on brain MRI according to claim 1, characterized in that: When performing the registration operation, the registration index is the mean square error, the optimizer uses the gradient descent method, and the interpolation method is the linear interpolation method.

6. A device for detecting the grade of glioma based on brain MRI, characterized in that: include: An MRI processing module is used to perform a registration operation to align the orientation information of all sequences of brain MRI images; for each of the brain MRI images, the brain region is segmented to obtain a post-skull removal MRI image of each sequence; Performing feature fusion on the selected MRI image after skull removal, and then combining it with the target MRI image after skull removal to obtain an MRI image set; A glioma segmentation module, used for inputting the MRI image set into a glioma segmentation model to obtain high-order features of the glioma, wherein the glioma segmentation model obtains the high-order features of the glioma when segmenting the glioma region; The glioma grade classification module is used to input the glioma high-order features into the glioma grade classification model to obtain glioma grade information.

7. The device for detecting glioma grade based on brain MRI according to claim 6, characterized in that: The glioma segmentation model is constructed based on a 2D segmentation network. The glioma segmentation model includes a slicing unit and a feature extraction unit. The slicing unit is used to perform 2D slicing on each MRI image in the MRI image set, obtain multiple 2D images corresponding to each MRI image, extract features of each 2D image, and stack the features of all 2D images according to the hierarchical structure of the MRI image to which the 2D image belongs, so as to obtain a glioma region. The feature extraction unit is used to set the feature extracted from the last layer of each 2D image during downsampling as the first glioma feature, and to combine the first glioma features corresponding to all 2D images of each MRI image into second glioma features corresponding to each MRI image, and all the second glioma features in the MRI image set constitute the glioma high-order feature.

8. The device for detecting glioma grade based on brain MRI according to claim 6, characterized in that: It also includes a data collection module and a training module. The data collection module is used to collect multiple sequences of brain MRI images of cases, and the training module is used to train the glioma segmentation model and the glioma grade classification model.

9. An electronic device, characterized in that: It includes a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the brain glioma grade detection method based on brain MRI as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program enables a computer to execute the method for detecting the grade of glioma based on brain MRI according to any one of claims 1 to 5.

Images