Non-invasive prediction model training system and prediction system for glioma IDH mutation status

Abstract

The present invention discloses a non-invasive prediction model training system and prediction system for the IDH mutation status of glioma, relating to the field of medical image processing technology. The training system comprises a data acquisition module, a data processing module, a first-stage neural network model training module, a second-stage neural network model training module, a third-stage neural network model training module, and a model evaluation module. The training system operates on conventional nuclear magnetic resonance (NMR) sequence image data and FW sequence image data, FA sequence image data, and MD sequence image data of the peritumoral edema area obtained by processing DTI sequence image data, thereby making feature extraction and analysis more reliable and accurate, solving the limitations of DWI sequence image data in analyzing the IDH mutation status, and ensuring the accuracy of the established prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a non-invasive prediction model training system and prediction system for glioma IDH mutation status based on multiple nuclear magnetic resonance imaging feature information of peritumoral edema areas. Background Art

[0002] Currently, preoperative invasive testing for IDH (isocitrate dehydrogenase) mutation status in gliomas requires obtaining tumor tissue specimens, which places high demands on experimental conditions. Non-invasive preoperative testing for IDH mutation status in gliomas requires imaging. For example, conventional MRI sequence data can identify glioma IDH mutation status based on information such as tumor location, diameter, degree of cystic necrosis, midline structural shift, degree of edema, and enhancement. While combining conventional MRI imaging features can achieve non-invasive prediction, overall diagnostic efficacy is relatively low, with a positive predictive value of only 39%, resulting in moderate diagnostic efficacy. A more reliable and objective method is needed. Summary of the Invention

[0003] The purpose of the present invention is to provide a non-invasive prediction model training system and prediction system for glioma IDH mutation status, so as to improve the accuracy of non-invasive prediction of glioma IDH mutation status.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] In a first aspect, the present invention provides a non-invasive prediction model training system for glioma IDH mutation status, comprising:

[0006] A data acquisition module is used to obtain various nuclear magnetic resonance imaging data of the glioma region and corresponding IDH mutation status information; wherein the nuclear magnetic resonance imaging data obtained by the data acquisition module include T1 sequence imaging data, T2 sequence imaging data, T2-FLAIR sequence imaging data, T1+C sequence imaging data and DTI sequence imaging data;

[0007] a data processing module, configured to process each nuclear magnetic resonance imaging sequence data to obtain nuclear magnetic resonance imaging sequence data of each peritumoral edema area and IDH mutation status information corresponding to the nuclear magnetic resonance imaging sequence data of each peritumoral edema area; wherein the nuclear magnetic resonance imaging sequence data of the peritumoral edema area output by the data processing module includes T1 sequence imaging data of the peritumoral edema area, T2 sequence imaging data of the peritumoral edema area, T2-FLAIR sequence imaging data of the peritumoral edema area, T1+C sequence imaging data of the peritumoral edema area, FW sequence imaging data of the peritumoral edema area, FA sequence imaging data of the peritumoral edema area, and MD sequence imaging data of the peritumoral edema area;

[0008] The first-stage neural network model training module is used to: use the MRI sequence image data of a peritumoral edema area as input data and the corresponding IDH mutation status information as label data to train the first-stage neural network model, thereby obtaining a trained first-stage neural network model corresponding to the MRI sequence image data of each peritumoral edema area;

[0009] The second stage neural network model training module is used to:

[0010] Using the structural phase feature as input data and the IDH mutation status information corresponding to the structural phase feature as label data, the first-stage neural network model is trained to obtain a trained second-stage neural network model corresponding to the structural phase feature; the structural phase feature includes the feature of the fully connected layer output of the first-stage neural network model trained with the first label; the first-stage neural network model trained with the first label includes a trained first-stage neural network model corresponding to the T1 sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the T2 sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the T2-FLAIR sequence image data of the peritumoral edema area, and a trained first-stage neural network model corresponding to the T1+C sequence image data of the peritumoral edema area;

[0011] Using the functional phase feature as input data and the IDH mutation status information corresponding to the functional phase feature as label data, a second-stage neural network model is trained to obtain a trained second-stage neural network model corresponding to the functional phase feature; the functional phase feature includes features output by a fully connected layer of the first-stage neural network model trained with the second label; the first-stage neural network model trained with the second label includes a trained first-stage neural network model corresponding to the FW sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the FA sequence image data of the peritumoral edema area, and a trained first-stage neural network model corresponding to the MD sequence image data of the peritumoral edema area;

[0012] A third-stage neural network model training module is configured to: use comprehensive features as input data and IDH mutation status information corresponding to the comprehensive features as label data to train the third-stage neural network model to obtain a trained third-stage neural network model; the comprehensive features include features output by the fully connected layer of the trained second-stage neural network model corresponding to the structural phase features and features output by the fully connected layer of the trained second-stage neural network model corresponding to the functional phase features;

[0013] The model evaluation module is used to evaluate all trained first-stage neural network models, trained second-stage neural network models, and trained third-stage neural network models to obtain evaluation results, and based on the evaluation results, determine the trained neural network model with the best model performance as the non-invasive prediction model for glioma IDH mutation status.

[0014] Optionally, the data processing module specifically includes:

[0015] An extraction and conversion processing unit is used to extract and convert the DTI sequence image data to obtain FW sequence image data, FA sequence image data and MD sequence image data;

[0016] A correction and registration unit is used to perform correction and registration processing on T1 sequence image data, T2 sequence image data, T2-FLAIR sequence image data, T1+C sequence image data, FW sequence image data, FA sequence image data and MD sequence image data;

[0017] an image segmentation unit, configured to perform image segmentation on the corrected and registered T2-FLAIR sequence image data and the corrected and registered T1+C sequence image data to obtain T2-FLAIR sequence image data and T1+C sequence image data of the peritumoral edema area, and perform feature extraction on the corrected and registered T1 sequence image data, the corrected and registered T2 sequence image data, the corrected and registered FW sequence image data, the corrected and registered FA sequence image data, and the corrected and registered MD sequence image data, respectively, based on the T2-FLAIR sequence image data and the T1+C sequence image data of the peritumoral edema area, to obtain T1 sequence image data, T2 sequence image data, FW sequence image data, FA sequence image data, and MD sequence image data of the peritumoral edema area;

[0018] The IDH mutation status information matching unit is used to perform IDH mutation status information matching operations on the T1 sequence image data of the peritumoral edema area, the T2 sequence image data of the peritumoral edema area, the T2-FLAIR sequence image data of the peritumoral edema area, the T1+C sequence image data of the peritumoral edema area, the FW sequence image data of the peritumoral edema area, the FA sequence image data of the peritumoral edema area, and the MD sequence image data of the peritumoral edema area.

[0019] Optionally, the correction and registration unit is further configured to:

[0020] Adjusting T1 sequence image data, T2 sequence image data, T2-FLAIR sequence image data, T1+C sequence image data, FW sequence image data, FA sequence image data, and MD sequence image data to sequence image data with the same slice thickness and the same voxel;

[0021] The T2-FLAIR sequence image data was used as the reference sequence image data, and other sequence image data with the same slice thickness and the same voxel were registered.

[0022] Optionally, the first-stage neural network model includes an input layer, a feature extraction layer, a fully connected layer and a classifier connected in sequence; the feature extraction layer is composed of multiple 3D convolution blocks and multiple separable convolution blocks.

[0023] Optionally, the feature extraction layer includes a first 3D convolution block, a second 3D convolution block, a first separable convolution block, a second separable convolution block and a 3D convolution block group connected in sequence; wherein the 3D convolution block group includes two branches, namely a first branch and a second branch; the first branch and the second branch each include two connected 3D convolution blocks, and the input ends of the first branch and the second branch are both connected to the output end of the second separable convolution block, and the output ends of the first branch and the second branch are both connected to the fully connected layer; the fully connected layer includes two connected fully connected blocks.

[0024] Optionally, the non-invasive prediction model training system for glioma IDH mutation status is characterized in that the second-stage neural network model includes a feature mapping fusion layer, a first convolution block, a second convolution block, a fully connected layer and a classifier connected in sequence.

[0025] Optionally, the third stage neural network model includes a feature map fusion layer, a fully connected layer and a classifier connected in sequence.

[0026] In a second aspect, the present invention provides a non-invasive prediction system for the IDH mutation status of glioma, comprising:

[0027] The data acquisition module is used to: collect the nuclear magnetic resonance imaging data of the glioma area input by the non-invasive prediction model of the glioma IDH mutation status;

[0028] The data screening module is used to screen the nuclear magnetic resonance imaging data of the glioma area to obtain the nuclear magnetic resonance imaging data of the peritumoral edema area;

[0029] The prediction module is used to input the nuclear magnetic resonance imaging data of the peritumoral edema area into the non-invasive prediction model of the glioma IDH mutation status to predict the patient's glioma IDH mutation status information;

[0030] Among them, the non-invasive prediction model of glioma IDH mutation status is determined based on the non-invasive prediction model training system of glioma IDH mutation status described in the first aspect.

[0031] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0032] This study uses FW, FA, and MD sequence image data obtained after processing DTI sequence image data to explore the imaging characteristics of different segmented regions of gliomas, particularly the peritumoral edema area. This method can, to a certain extent, avoid interference with DTI sequence image data caused by cystic degeneration and necrosis in the tumor core, reduce the partial volume effect, and improve the actual feature quality of DTI sequence image data, thereby increasing diagnostic efficacy, which is relatively rare in previous studies. This study takes into account the partial volume effect of DTI sequence image data and the characteristics of glioma tumor heterogeneity, making feature extraction and analysis more reliable and accurate, addressing the limitations of DWI sequence image data in analyzing IDH mutation status, while ensuring the accuracy of the established prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 A structural block diagram of a non-invasive prediction model training system for glioma IDH mutation status provided by an embodiment of the present invention;

[0035] Figure 2 This is a structural block diagram of a non-invasive prediction system for IDH mutation status in gliomas provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0037] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Example 1

[0039] This embodiment provides a model for predicting the isocitrate dehydrogenase mutation status of gliomas by utilizing multiple nuclear magnetic resonance imaging feature information of the peritumoral edema area.

[0040] like Figure 1 As shown, this embodiment provides a non-invasive prediction model training system for glioma IDH mutation status, including:

[0041] The data acquisition module 101 is used to obtain the nuclear magnetic resonance imaging data of the glioma area and the corresponding IDH mutation status information; wherein the nuclear magnetic resonance imaging data obtained by the data acquisition module include T1 sequence imaging data, T2 sequence imaging data, T2-FLAIR sequence imaging data, T1+C (T1 enhancement) sequence imaging data and DTI sequence imaging data.

[0042] The data processing module 102 is used to process each MRI sequence image data to obtain MRI sequence image data of each peritumoral edema area and IDH mutation status information corresponding to the MRI sequence image data of each peritumoral edema area; wherein the MRI sequence image data of the peritumoral edema area output by the data processing module includes T1 sequence image data of the peritumoral edema area, T2 sequence image data of the peritumoral edema area, T2-FLAIR sequence image data of the peritumoral edema area, T1+C (T1 enhancement) sequence image data of the peritumoral edema area, FW sequence image data of the peritumoral edema area, FA sequence image data of the peritumoral edema area, and MD sequence image data of the peritumoral edema area.

[0043] The first-stage neural network model training module 103 is configured to: use the MRI sequence image data of a peritumoral edema area as input data and the corresponding IDH mutation status information as label data to train the first-stage neural network model, thereby obtaining a trained first-stage neural network model corresponding to the MRI sequence image data of each peritumoral edema area;

[0044] The second stage neural network model training module 104 is used to:

[0045] Using the structural phase feature as input data and the IDH mutation status information corresponding to the structural phase feature as label data, the first-stage neural network model is trained to obtain a trained second-stage neural network model corresponding to the structural phase feature; the structural phase feature includes the feature of the fully connected layer output of the first-stage neural network model trained with the first label; the first-stage neural network model trained with the first label includes a trained first-stage neural network model corresponding to the T1 sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the T2 sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the T2-FLAIR sequence image data of the peritumoral edema area, and a trained first-stage neural network model corresponding to the T1+C sequence image data of the peritumoral edema area;

[0046] The second-stage neural network model is trained using the functional phase feature as input data and the IDH mutation status information corresponding to the functional phase feature as label data to obtain a trained second-stage neural network model corresponding to the functional phase feature; the functional phase feature includes the features of the fully connected layer output of the first-stage neural network model trained with the second label; the first-stage neural network model trained with the second label includes the trained first-stage neural network model corresponding to the FW sequence image data of the peritumoral edema area, the trained first-stage neural network model corresponding to the FA sequence image data of the peritumoral edema area, and the trained first-stage neural network model corresponding to the MD sequence image data of the peritumoral edema area.

[0047] The third-stage neural network model training module 105 is used to: use the comprehensive features as input data and the IDH mutation status information corresponding to the comprehensive features as label data to train the third-stage neural network model to obtain a trained third-stage neural network model; the comprehensive features include the features of the fully connected layer output of the trained second-stage neural network model corresponding to the structural phase features and the features of the fully connected layer output of the trained second-stage neural network model corresponding to the functional phase features.

[0048] The model evaluation module 106 is used to evaluate all trained first-stage neural network models, trained second-stage neural network models, and trained third-stage neural network models to obtain evaluation results, and determine the trained neural network model with the best model performance as the non-invasive prediction model for glioma IDH mutation status based on the evaluation results.

[0049] Since conventional MRI sequence image data cannot fully reflect the molecular pathological changes of tumors, DTI sequence image data is added in this embodiment. Among them, each MRI sequence image data and the corresponding glioma IDH mutation status information are directly obtained from the hospital's medical system, and the data format is Dicom format.

[0050] In this embodiment, the data processing module specifically includes:

[0051] The extraction and conversion processing unit is used to extract and convert the DTI sequence image data to obtain FW sequence image data, FA sequence image data and MD sequence image data.

[0052] For example, medical data in Dicom format were imported into 3Dslicer software (open source and free) and converted to nrrd format. The DWIConvert (github.com / BRAINSia / BRAINSTools) module in 3Dslicer then converted the multi-frame DTI file format from the MRI scan data into a single image, named a DWI sequence. The newly generated DWI sequence was then processed using the default DWIProtocol settings in DTIPrep (www.nitrc.org / projects / dtiprep), a DTI quality control tool, using the IMAGE_bCheck, DIFFUSION_bCheck, SLICE_bCheck, INTERLACE_bCheck, BASELINE_bAverage, EDDYMOTION_bCorrect, GRADIENT_bCheck, and DTI_bCompute methods, generating the DWI_QCed_Baseline and DWI_QCed files. The BRAINS module of 3DSlicer software was used to align and correct the DWI_QCed_Baseline file with other sequences of the same patient. The Freewater transformation formula, FA transformation formula, and MD transformation formula were used to calculate the DWI_QCed file to obtain the corresponding FW sequence image data, FA sequence image data, and MD sequence image data.

[0053] The correction and registration unit is used to perform correction and registration processing on T1 sequence image data, T2 sequence image data, T2-FLAIR sequence image data, T1+C (T1 enhancement) sequence image data, FW sequence image data, FA sequence image data and MD sequence image data.

[0054] An example: Before registering each MRI sequence image data, the ResampleScalarVolume module in the Slicer software is used to adjust the existing sequence to a sequence with the same layer thickness and the same voxel. The reference sequence is the T2-FLAIR sequence image data (because the T2-FLAIR sequence image data has a better tumor display effect and is conducive to tumor segmentation), and the voxel unit is changed to 1mm×1mm×5.235mm. Then, the BRAINS module in the Slicer software is used to perform sequence registration. Because in this embodiment, tumor segmentation mainly relies on the T2-FLAIR sequence (which has a better tumor display), the Fixed Image Volume (reference sequence) in the BRAINS module is the T2-FLAIR sequence image data, and the Moving Imaging Volume (registration sequence) is other MRI sequence image data. Rigid linear registration is performed on the MRI sequence image data, and the generated registration model is used to register similar data.

[0055] The image segmentation unit is used to perform image segmentation on the T2-FLAIR sequence image data after correction and registration and the T1+C sequence image data after correction and registration to obtain T2-FLAIR sequence image data and T1+C sequence image data of the peritumoral edema area, and perform feature extraction on the T1 sequence image data after correction and registration, the T2 sequence image data after correction and registration, the FW sequence image data after correction and registration, the FA sequence image data after correction and registration, and the MD sequence image data after correction and registration based on the T2-FLAIR sequence image data and the T1+C sequence image data of the peritumoral edema area to obtain T1 sequence image data, T2 sequence image data, FW sequence image data, FA sequence image data, and MD sequence image data of the peritumoral edema area.

[0056] An example: After the MRI sequence image data are corrected and registered, the segmented labels can be accurately superimposed on the MRI sequence image data. For example, the peritumoral edema zone is segmented based on the T2-FLAIR sequence image data, and the segmented area itself becomes a label. In this way, the image information characteristics of the peritumoral edema area on the T1+C sequence image data can be extracted by simply superimposing the label on the T1+C sequence image data.

[0057] The MRI sequence image data of glioma patients in nrrd format were imported into 3D Slicer software. The General Registration BRIANS module was used for image registration. After completion, the Segmentation module was used to segment the tumor, mainly segmenting the overall tumor area on the T2-FLAIR sequence image data and the tumor enhancement area on the T1+C sequence image data. The Subtract ScalarVolumes module was then used to increase or decrease the segmented area. The tumor edema area was obtained by subtracting the core enhancement area from the overall tumor area, and the final image processing segmentation area, namely the peritumoral edema area, was obtained.

[0058] The IDH mutation status information matching unit is used to perform IDH mutation status information matching operations on the T1 sequence image data of the peritumoral edema area, the T2 sequence image data of the peritumoral edema area, the T2-FLAIR sequence image data of the peritumoral edema area, the T1+C sequence image data of the peritumoral edema area, the FW sequence image data of the peritumoral edema area, the FA sequence image data of the peritumoral edema area, and the MD sequence image data of the peritumoral edema area.

[0059] Since the sizes of peritumoral edema areas vary among patients, the segmented MRI sequence image data of each peritumoral edema area were adjusted to a uniform size (e.g., 64*64*16) to facilitate subsequent model training.

[0060] In this embodiment, the deep learning network model uses MRI sequence data from the peritumoral edema area as input, including conventional sequences (T1, T1+C, T2, and T2-FLAIR) and advanced sequences (FA, FW, and MD) derived from DTI sequence data processing. The entire model architecture comprises a three-stage network, each with its own complete feature extractor and classifier, each expected to output relevant prediction results. Furthermore, the data is divided into a validation set and a training set in a certain proportion. The network model is trained based on the training set, and then the networks at each stage of the model are evaluated and screened based on the validation set.

[0061] The first-stage neural network model is a network whose input is single nuclear magnetic resonance imaging data, that is, each nuclear magnetic resonance imaging data has its own independent neural network.

[0062] The first-stage neural network model includes an input layer, a feature extraction layer, a fully connected layer and a classifier connected in sequence; the feature extraction layer is composed of multiple 3D convolution blocks and multiple separable convolution blocks.

[0063] The feature extraction layer includes a first 3D convolution block, a second 3D convolution block, a first separable convolution block, a second separable convolution block and a 3D convolution block group connected in sequence; the 3D convolution block group includes two branches, namely a first branch and a second branch; the first branch and the second branch each include two connected 3D convolution blocks, and the input ends of the first branch and the second branch are both connected to the output end of the second separable convolution block, and the output ends of the first branch and the second branch are both connected to the fully connected layer; the fully connected layer includes two connected fully connected blocks.

[0064] The convolution block (convblock) mainly includes a convolution layer (including activation function), batch normalization, a max pooling layer, and dropout. In addition to extracting deep features, the convolution block can also downsample the image by adjusting the filter size of the max pooling layer.

[0065] The separable convolution block (sep conv block) mainly includes a separable convolution layer (including activation function), batch normalization, and dropout, and is mainly used for training deeper modules. Because deep convolution modules require more parameters to train, using separable convolution blocks can speed up training and save time.

[0066] The fully connected block (FC block) mainly includes the fully connected layer (including activation function), batch normalization and Dropout, and is mainly used in the classifier part of the latter part of the network.

[0067] An example: (1) The input is the peritumoral edema image data of three groups of patients (IDH wild type or mutant) after single sequence preprocessing. (2) Deep feature extraction: First, two to three convolution blocks are used to perform deep feature extraction and downsampling on the image. In order to facilitate subsequent modality fusion, the size of the convolution block is adjusted so that the feature map size between different modalities after training at this stage is consistent, and the number of maps can be appropriately increased. Then, multiple separable convolution blocks are used to further deepen the model for training. Next, the feature map is divided into two groups before and after, and multiple group convolutions are performed. Group convolution can also achieve a certain effect of improving training efficiency. More importantly, considering that the input data comes from three groups of patients, group convolution learning can discover "individual features" in "common features", which may achieve unexpected results. Group convolution is the tail of the feature extraction part, so the three-dimensional feature map will gradually become a one-dimensional feature map, and then be reconnected into a group. (3) Task-based classification learning: The feature map is followed by two to three fully connected blocks and a normalized exponential function as the classifier of the first-stage neural network model.

[0068] The second-stage neural network model includes a feature map fusion layer, a first convolutional block, a second convolutional block, a fully connected layer, and a classifier, which are connected in sequence.

[0069] The second-stage neural network model is a co-phase sequence fusion network model. After training the first-stage neural network model, the second-stage neural network model performs preliminary fusion between different modalities. It is expected to fuse deep feature maps from conventional imaging sequences, such as T1+C, T2, and T2FLAIR, to form a structural neural network. Deep feature maps from DIT imaging sequences, such as FA, FW, and MD, will be fused to form a functional neural network. The output of the first fully connected layer of the first-stage neural network model serves as the input to the second-stage neural network model. The fusion weights of the different modal input values ​​are determined by the classification results of the first-stage neural network model for each modality. Furthermore, when performing multimodal feature fusion, some samples may be missing from some modalities. Therefore, the maximal set of intermodal samples is selected during fusion. Since the input is one-dimensional data, the second-stage neural network model primarily utilizes convolutional blocks, fully connected blocks, and a final normalized exponential function.

[0070] The third stage neural network model includes a feature map fusion layer, a fully connected layer and a classifier connected in sequence.

[0071] The third-stage neural network model is a multi-sequence fully fused network model, similar to the second-stage neural network model, but this one incorporates clinical information modalities. Due to the relatively small dimensionality of clinical information data, it is expected that after two rounds of fully connected block training, it will be integrated with conventional and advanced sequence networks to form the final multimodal neural network model.

[0072] Taking into account the overfitting problem of deep network models, the classification performance of the neural network models at each stage is evaluated based on the reserved validation set, as well as indicators such as sensitivity, specificity, accuracy and AUC (Area Under Curve).

[0073] The selection priority of the stage neural network model is: if the classification evaluation result of the third stage neural network model is the best, the network models of all three stages are retained; if the evaluation result of the structural phase or functional phase neural network model of the second stage is the best, the network models of the corresponding two stages are retained; similarly, whichever sequence of the neural network model in the first stage has the best evaluation result, the corresponding first stage neural network model is retained.

[0074] Example 2

[0075] like Figure 2 As shown, this embodiment provides a non-invasive prediction system for the IDH mutation status of glioma, including:

[0076] The data acquisition module 201 is used to collect the nuclear magnetic resonance imaging data of the glioma region input into the non-invasive prediction model of the glioma IDH mutation status.

[0077] The data screening module 202 is used to screen the nuclear magnetic resonance imaging data of the glioma area to obtain the nuclear magnetic resonance imaging data of the peritumoral edema area.

[0078] Prediction module 203 is used to input the peritumoral edema area MRI sequence image data into the glioma IDH mutation status non-invasive prediction model to predict the patient's glioma IDH mutation status information.

[0079] Among them, the non-invasive prediction model for glioma IDH mutation status is determined according to Example 1.

[0080] This study compared and analyzed different existing diagnostic models for brain tumors, finding that a combined diagnostic prediction model combining these different sequences has a high accuracy of 90%. Analysis of imaging features of gliomas in different segmented regions (the peritumoral edema zone, the core area, and the tumor as a whole) suggests that the peritumoral edema zone prediction model also has a higher accuracy than previous methods using only the peritumoral edema zone.

[0081] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0082] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A non-invasive prediction model training system for glioma IDH mutation status, characterized by: include: A data acquisition module is used to obtain various nuclear magnetic resonance imaging data of the glioma region and corresponding IDH mutation status information; wherein the nuclear magnetic resonance imaging data obtained by the data acquisition module include T1 sequence imaging data, T2 sequence imaging data, T2-FLAIR sequence imaging data, T1+C sequence imaging data and DTI sequence imaging data; a data processing module, configured to process each nuclear magnetic resonance imaging sequence data to obtain nuclear magnetic resonance imaging sequence data of each peritumoral edema area and IDH mutation status information corresponding to the nuclear magnetic resonance imaging sequence data of each peritumoral edema area; wherein the nuclear magnetic resonance imaging sequence data of the peritumoral edema area output by the data processing module includes T1 sequence imaging data of the peritumoral edema area, T2 sequence imaging data of the peritumoral edema area, T2-FLAIR sequence imaging data of the peritumoral edema area, T1+C sequence imaging data of the peritumoral edema area, FW sequence imaging data of the peritumoral edema area, FA sequence imaging data of the peritumoral edema area, and MD sequence imaging data of the peritumoral edema area; The first-stage neural network model training module is used to: use the MRI sequence image data of a peritumoral edema area as input data and the corresponding IDH mutation status information as label data to train the first-stage neural network model, thereby obtaining a trained first-stage neural network model corresponding to the MRI sequence image data of each peritumoral edema area; The second stage neural network model training module is used to: Using the structural phase feature as input data and the IDH mutation status information corresponding to the structural phase feature as label data, the first-stage neural network model is trained to obtain a trained second-stage neural network model corresponding to the structural phase feature; the structural phase feature includes the feature of the fully connected layer output of the first-stage neural network model trained with the first label; the first-stage neural network model trained with the first label includes a trained first-stage neural network model corresponding to the T1 sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the T2 sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the T2-FLAIR sequence image data of the peritumoral edema area, and a trained first-stage neural network model corresponding to the T1+C sequence image data of the peritumoral edema area; Using the functional phase feature as input data and the IDH mutation status information corresponding to the functional phase feature as label data, a second-stage neural network model is trained to obtain a trained second-stage neural network model corresponding to the functional phase feature; the functional phase feature includes features output by a fully connected layer of the first-stage neural network model trained with the second label; the first-stage neural network model trained with the second label includes a trained first-stage neural network model corresponding to the FW sequence image data of the peritumoral edema area, a trained first-stage neural network model corresponding to the FA sequence image data of the peritumoral edema area, and a trained first-stage neural network model corresponding to the MD sequence image data of the peritumoral edema area; A third-stage neural network model training module is configured to: use comprehensive features as input data and IDH mutation status information corresponding to the comprehensive features as label data to train the third-stage neural network model to obtain a trained third-stage neural network model; the comprehensive features include features output by the fully connected layer of the trained second-stage neural network model corresponding to the structural phase features and features output by the fully connected layer of the trained second-stage neural network model corresponding to the functional phase features; The model evaluation module is used to evaluate all trained first-stage neural network models, trained second-stage neural network models, and trained third-stage neural network models to obtain evaluation results, and based on the evaluation results, determine the trained neural network model with the best model performance as the non-invasive prediction model for glioma IDH mutation status.

2. A non-invasive prediction model training system for glioma IDH mutation status according to claim 1, characterized in that: The data processing module specifically includes: An extraction and conversion processing unit is used to extract and convert the DTI sequence image data to obtain FW sequence image data, FA sequence image data and MD sequence image data; A correction and registration unit is used to perform correction and registration processing on T1 sequence image data, T2 sequence image data, T2-FLAIR sequence image data, T1+C sequence image data, FW sequence image data, FA sequence image data and MD sequence image data; an image segmentation unit, configured to perform image segmentation on the corrected and registered T2-FLAIR sequence image data and the corrected and registered T1+C sequence image data to obtain T2-FLAIR sequence image data and T1+C sequence image data of the peritumoral edema area, and perform feature extraction on the corrected and registered T1 sequence image data, the corrected and registered T2 sequence image data, the corrected and registered FW sequence image data, the corrected and registered FA sequence image data, and the corrected and registered MD sequence image data, respectively, based on the T2-FLAIR sequence image data and the T1+C sequence image data of the peritumoral edema area, to obtain T1 sequence image data, T2 sequence image data, FW sequence image data, FA sequence image data, and MD sequence image data of the peritumoral edema area; The IDH mutation status information matching unit is used to perform IDH mutation status information matching operations on the T1 sequence image data of the peritumoral edema area, the T2 sequence image data of the peritumoral edema area, the T2-FLAIR sequence image data of the peritumoral edema area, the T1+C sequence image data of the peritumoral edema area, the FW sequence image data of the peritumoral edema area, the FA sequence image data of the peritumoral edema area, and the MD sequence image data of the peritumoral edema area.

3. A non-invasive prediction model training system for glioma IDH mutation status according to claim 2, characterized in that: The correction and registration unit is further configured to: Adjusting T1 sequence image data, T2 sequence image data, T2-FLAIR sequence image data, T1+C sequence image data, FW sequence image data, FA sequence image data, and MD sequence image data to sequence image data with the same slice thickness and the same voxel; The T2-FLAIR sequence image data was used as the reference sequence image data, and other sequence image data with the same slice thickness and the same voxel were registered.

4. The non-invasive prediction model training system for glioma IDH mutation status according to claim 1, characterized in that: The first-stage neural network model includes an input layer, a feature extraction layer, a fully connected layer and a classifier connected in sequence; the feature extraction layer is composed of multiple 3D convolution blocks and multiple separable convolution blocks.

5. The non-invasive prediction model training system for glioma IDH mutation status according to claim 4, characterized in that: The feature extraction layer includes a first 3D convolution block, a second 3D convolution block, a first separable convolution block, a second separable convolution block and a 3D convolution block group connected in sequence; wherein the 3D convolution block group includes two branches, namely a first branch and a second branch; the first branch and the second branch each include two connected 3D convolution blocks, and the input ends of the first branch and the second branch are both connected to the output end of the second separable convolution block, and the output ends of the first branch and the second branch are both connected to the fully connected layer; The fully connected layer includes two connected fully connected blocks.

6. The non-invasive prediction model training system for glioma IDH mutation status according to claim 1, characterized in that: The second-stage neural network model includes a feature map fusion layer, a first convolution block, a second convolution block, a fully connected layer and a classifier connected in sequence.

7. The non-invasive prediction model training system for glioma IDH mutation status according to claim 1, characterized in that: The third-stage neural network model includes a feature map fusion layer, a fully connected layer, and a classifier connected in sequence.

8. A non-invasive prediction system for IDH mutation status in glioma, characterized by: include: The data acquisition module is used to: collect the nuclear magnetic resonance imaging data of the glioma area input by the non-invasive prediction model of the glioma IDH mutation status; The data screening module is used to screen the nuclear magnetic resonance imaging data of the glioma area to obtain the nuclear magnetic resonance imaging data of the peritumoral edema area; The prediction module is used to input the nuclear magnetic resonance imaging data of the peritumoral edema area into the non-invasive prediction model of the glioma IDH mutation status to predict the patient's glioma IDH mutation status information; Among them, the non-invasive prediction model of glioma IDH mutation status is determined according to the non-invasive prediction model training system of glioma IDH mutation status according to any one of claims 1-7.

Images