Glioblastoma peritumoral tumor infiltration segmentation model

Abstract

The invention discloses a glioblastoma peritumoral tumor infiltration segmentation model, which comprises the steps of clinical case acquisition, image information acquisition, pathological information acquisition, image data preprocessing, self-learning image feature learning on a Gb peritumoral edema area based on a CPC mode, transfer learning by using a 3DUX-Net segmentation model, and segmentation of glioblastoma peritumoral tumor infiltration. And constructing a peritumor infiltration segmentation model and verifying the segmentation model. According to the method, an artificial ROI segmentation feature is learned, the peritumoral infiltration segmentation model is established, the effectiveness of the model is verified through a double-blind test, the peritumoral tumor infiltration is accurately segmented through the peritumoral infiltration segmentation model, and an important reference is provided for clinical multidisciplinary diagnosis and treatment.

Description

Technical Field

[0001] The invention belongs to the technical field of medical imaging and relates to a glioblastoma peritumoral tumor infiltration segmentation model. Background Art

[0002] Glioblastoma (Gb) peritumoral edema is mainly composed of tumor infiltration and vascular edema. Tumor infiltration usually contains abundant tumor cells and is highly invasive. The infiltration boundary cannot be simply determined based on the specific distance of the enhancement edge. It is difficult to accurately identify tumor infiltration through imaging technology before surgery, resulting in incomplete surgical resection, which is the main reason for the poor prognosis and high recurrence rate of Gb.

[0003] Tumor infiltration in the current preoperative segmentation planned treatment range (PTR) can be summarized into two research trends: exploring imaging signs or markers of tumor infiltration based on multimodal imaging, and determining the infiltration boundary of brain glioma based on artificial intelligence models and technologies. However, in clinical practice, it is impossible to directly obtain a large number of accurate tumor cell infiltration markers and pure vascular edema markers from Gb, and thus it is impossible to accurately segment peritumoral tumor infiltration through limited infiltration segmentation models. Summary of the invention

[0004] In view of the above problems, the present invention proposes a glioblastoma peritumoral tumor infiltration segmentation model, which well solves the problems in the prior art.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] Glioblastoma peritumoral tumor infiltration segmentation model, including:

[0007] Clinical case collection: design inclusion and exclusion criteria for glioblastoma, non-enhancing glioma, and meningioma, and collect cases that meet the requirements;

[0008] Image information acquisition: using a 3TMR scanner to acquire case images, including conventional MRI plain scan, multi-b-value diffusion-weighted imaging scan and high-resolution 3DT1WI enhanced scan;

[0009] Pathological information collection, including stereotactic biopsy information of glioblastoma patients and cell morphology, mitotic count, and Ki-67 labeling index under pathological examination;

[0010] Image data preprocessing, parameter map calculation and registration of images, and manual segmentation of regions of interest, followed by image data standardization;

[0011] Image feature learning: using contrast prediction coding to self-learn the Gb peritumoral edema area and extract complete features from the Gb dataset;

[0012] Establish an invasion segmentation model, use the 3DUX-Net segmentation model for transfer learning, and build a peritumoral invasion segmentation model;

[0013] Segmentation model validation: double-blind analysis and prediction of cases were performed using manual and peritumoral infiltration segmentation models, and the analysis results were compared to evaluate the effectiveness of the model.

[0014] Optionally, the image data standardization includes:

[0015] Image normalization, converting the pixel values ​​of image data to a specific range so that different images have similar pixel value ranges;

[0016] Image cropping: extracting the region of interest from the 3D image data based on the image center and cropping it to a uniform size;

[0017] Resampling, calculate the size ratio or resolution ratio of the original image and the target image, in each dimension, calculate the position of the pixel position in the target image in the original image by interpolation method, and use the corresponding interpolation method to calculate the pixel value.

[0018] Optionally, the image feature learning also includes supervised learning based on an attention mechanism and a convolutional network.

[0019] Optionally, the CPC self-learning includes:

[0020] Feature encoding, the encoder network E maps the original data to a low-dimensional feature space and outputs the encoded feature Z:

[0021] Z=E(X)

[0022] Among them, X is the original 3D data, and Z is the encoded feature representation;

[0023] Predictive modeling: The prediction model P predicts future features based on historical features. Assuming that the features of the next k steps are to be predicted, then:

[0024]

[0025] in, is the predicted future feature, Z t：t+k-1 It is a sequence of historical characteristics;

[0026] Contrastive loss construction,Contrastive loss is constructed based on InfoNCE loss, which is used to maximize the similarity between positive samples and minimize the similarity between negative samples. The contrastive loss can be defined as:

[0027]

[0028] Among them, sim(x) is the similarity measurement function, and τ is used to adjust the parameters of the contrast loss scale;

[0029] Model training, model training uses the optimizer and contrast loss function to update the parameters of the encoder E and the prediction model P:

[0030]

[0031] Among them, θ is the model parameter, η is the learning rate, is the gradient operator.

[0032] Optionally, the 3DUX-Net segmentation model includes,

[0033] Feature extraction, input the preprocessed data into the model respectively, let the multimodal data be X, satisfy Represents data with N samples, C channels, depth Z, height H, and width W, and the extracted feature is F. Similarly, Extract features:

[0034] F=E(X)

[0035] Among them, E represents feature extractor, 3D convolutional neural network;

[0036] Feature fusion, let the fused feature be G, satisfying Fusion features:

[0037]

[0038] in, Represents the fusion method, which attempts element-by-element addition, element-by-element multiplication, or channel concatenation, etc.

[0039] Feature decoding, decode the fused feature G, and set the decoded result to Y, satisfying Feature decoding:

[0040] Y=D(G)

[0041] Where D stands for decoder, 3D transposed convolutional neural network;

[0042] Loss calculation and optimization, let the loss value be L, the true label be T, and calculate the loss:

[0043] L = LossFunction(Y, T);

[0044] Model training, based on the training data, feature extractor E, fusion mechanism, decoder D, loss function L, use the optimizer; and loss function to update the model parameters to minimize the loss and perform training:

[0045]

[0046] Among them, θ is the model parameter, η is the learning rate, is the gradient operator.

[0047] Optionally, the feature extraction includes:

[0048] Extract spatial features from multimodal MRI via 3D convolutional layers.

[0049]

[0050] in, represents the output feature map of position (i, j, k) in layer l, represents a 3D convolution kernel of size (m, n, o) in the lth layer, the symbol * represents the convolution operation, and b l is the bias term of the lth layer, and f is the activation function;

[0051] Through the hierarchical Transformer structure of the self-attention mechanism,

[0052]

[0053] Among them, Q, K, and V represent query, key, and value, respectively, which are usually obtained from the input feature map through linear transformation. k is the dimension of the key vector, used to scale the attention scores, and softmax is the normalization function.

[0054] Optionally, the segmentation accuracy judgment method in the segmentation model verification includes Dice similarity coefficient, Hausdorff distance and intersection-over-union ratio.

[0055] Optionally, the manual segmentation of the region of interest uses at least two different manual segmentations, and the difference in the overlapping area of ​​the two ROIs should not be greater than 5%.

[0056] Compared with the prior art, the present invention has the following beneficial effects: by learning the artificial ROI segmentation features, a peritumoral infiltration segmentation model is established, and the effectiveness of the model is verified through double-blind testing. The peritumoral infiltration segmentation model is used to accurately segment peritumoral tumor infiltration, providing an important reference for clinical multidisciplinary diagnosis and treatment. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The glioblastoma peritumoral tumor infiltration segmentation model disclosed in the embodiment of the present invention includes:

[0059] Clinical case collection: design inclusion and exclusion criteria for glioblastoma, non-enhancing glioma, and meningioma, and collect cases that meet the requirements;

[0060] Image information acquisition: 3TMR scanner is used to acquire case images, including conventional magnetic resonance imaging (MRI) plain scan, multi-b-value diffusion-weighted imaging (DWI) scan and high-resolution 3DT1WI enhanced scan;

[0061] Pathological information collection, including stereotactic biopsy information of glioblastoma patients, cell morphology, mitotic count, and Ki-67 labeling index under pathological examination. Pathological information is also used to verify the segmentation model;

[0062] Image data preprocessing, parameter map calculation and registration of images, and manual region of interest (ROI) segmentation, followed by image data standardization;

[0063] Image feature learning: using contrast prediction coding (CPC) to self-learn the Gb peritumoral edema area and extract complete features from the Gb dataset;

[0064] Establish an invasion segmentation model, use the 3DUX-Net segmentation model for transfer learning, and build a peritumoral invasion segmentation model;

[0065] Segmentation model validation: double-blind analysis and prediction of cases were performed using manual and peritumoral infiltration segmentation models, and the analysis results were compared to evaluate the effectiveness of the model.

[0066] Specifically, clinical cases are collected and suitable cases are screened for imaging and pathological information collection. After manual segmentation of the region of interest (ROI), feature extraction is performed through relevant software, and the peritumoral infiltration segmentation model is trained. Finally, the results of the peritumoral infiltration segmentation model analysis and prediction are compared with those of manual analysis and prediction under double-blind conditions to evaluate the effectiveness of the model.

[0067] In this way, by learning the artificial ROI segmentation features, a peritumoral infiltration segmentation model is established, and the effectiveness of the model is verified through double-blind testing. The peritumoral infiltration segmentation model is used to accurately segment peritumoral tumor infiltration, providing an important reference for clinical multidisciplinary diagnosis and treatment.

[0068] In some feasible methods, when clinical cases are collected, they are included and screened. The inclusion criteria include age group, patients who have not received radiotherapy or chemotherapy before surgery, clear images and postoperative pathological confirmation of this type of disease, etc. The exclusion criteria include the lack of necessary magnetic resonance imaging sequences, the use of unconventional imaging parameters, and low-quality images.

[0069] A 3TMR scanner with a 64-channel head and neck coil can be used for image information acquisition. Conventional MRI plain scan sequences include T1WI, T2WI, and FLAIR. During high-resolution 3DT1WI enhanced scanning, gadopentetate meglumine magnetic resonance enhancement agent is injected through the cubital vein. For ease of implementation, this embodiment provides scanning parameters for image information acquisition:

[0070] The scanning parameters of axial TiWI were as follows: repetition time (TR) 250 ms, echo time (TE) 2.46 ms, matrix 314 × 314, field of view 220 mm × 220 mm, slice thickness 5 mm, number of slices 20, and scanning time 37 s.

[0071] The scanning parameters of axial T2WI were: TR 4090 milliseconds, TE 99 milliseconds, matrix 733×733, field of view 220 mm×220 mm, slice thickness 5 mm, 20 slices, and scanning time 34 seconds.

[0072] The scanning parameters of axial FLAIR were: TR 8000 milliseconds, TE 81 milliseconds, matrix 314×314, field of view 220 mm×220 mm, slice thickness 5 mm, number of slices 20, and scanning time 1 minute 38 seconds. The scanning parameters of 3D-T1-MPRAGE scanning were: TR 2300 milliseconds, TE 2.32 milliseconds, matrix 266×266, field of view 240 mm×240 mm, slice thickness 0.9 mm, number of slices 176, and scanning time 5 minutes 21 seconds.

[0073] Diffusion-weighted magnetic resonance imaging (DWI) was performed using a single-shot echo-planar sequence and 6 b-values ​​(0, 500, 1000, 1500, 2000, and 2500 s / mm2) for axial scanning. Each non-zero b-value was performed in 30 encoding directions, with TR 3500 milliseconds, TE 81 milliseconds, matrix 383×383, field of view 230 mm×230 mm, slice thickness 5 mm, 20 slices, and scanning time 54 seconds.

[0074] In some feasible methods, image data preprocessing includes parameter map calculation, image registration and segmentation, and image data standardization. In this embodiment, the sequence and parameter map are provided as shown in Table 1, where the sagittal 3D-T1-MPRAGE image is reconstructed into 20 layers of images in axial, sagittal and coronal positions.

[0075] Table 1 Sequence parameter diagram

[0076]

[0077]

[0078] DMRI parameter map calculation, the DWI image is imported into the post-processing software NeuDilab (based on DIPY), and the various parameter maps of DKI, DTI, NODDI, and MAP are calculated from the DWI image. The meaning of each parameter is prior art and will not be repeated here.

[0079] During image registration, the open source software ITK-SNAP's Automatic registration function mode was used, and FLAIR was used as a template to register the parameter maps of other conventional sequences and diffusion sequences. After registration, the images of each sequence had the same voxel size, initial image position, and resolution.

[0080] In ROI segmentation, experienced practitioners in related disciplines should manually draw through ITK-SNAP to outline the peritumoral edema area on glioblastoma, including the infiltration area and vasogenic edema area, the peritumoral infiltration area on non-enhancing glioma, and the edema area on meningioma, which is generally considered to be the vasogenic edema area. When outlining the ROI, refer to the corresponding conventional sequence.

[0081] In some feasible ways, in order to compensate for manual segmentation deviation, if the difference between the two manually segmented ROI images is less than 5%, the final ROI is determined as the overlapping area of ​​the two ROIs. If the difference is greater than 5%, a third person will make a judgment or re-segment and compare. 5% is a more appropriate empirical value in the experimental process.

[0082] As a specific implementation of the glioblastoma peritumoral tumor infiltration segmentation model provided in the application, image data standardization includes:

[0083] Image normalization, converting the pixel values ​​of image data to a specific range so that different images have similar pixel value ranges;

[0084] Image cropping: extracting the region of interest (ROI) from the 3D image data based on the image center and cropping it to a uniform size;

[0085] Resampling, calculate the size ratio or resolution ratio of the original image and the target image, in each dimension, calculate the position of the pixel position in the target image in the original image by interpolation method, and use the corresponding interpolation method to calculate the pixel value.

[0086] In general, standardizing image data so that different images have similar pixel value ranges, regulating the image size, and focusing on displaying the region of interest (ROI) in the center of the image can facilitate subsequent processing and improve the accuracy of feature extraction.

[0087] In some feasible ways, the interpolation method of this embodiment can be selected as spline interpolation. High-order spline interpolation (Spline Interpolation) is a commonly used interpolation method. Due to its smoothness and high precision, it is suitable for scenes that require precise resampling or transformation, such as image registration, reconstruction, segmentation post-processing, etc. Spline interpolation performs interpolation calculations on the neighborhood of the target pixel based on the spline function. In a two-dimensional image, high-order spline interpolation uses multiple pixel values ​​around the target pixel position to calculate the value of the target pixel by fitting the spline function. In this embodiment, cubic spline interpolation is used. In a two-dimensional image, cubic spline interpolation is a process of performing one-dimensional cubic interpolation on rows and columns respectively (i.e., separating axis interpolation). Assuming that the coordinates of the target point are (x, y), the calculation steps are as follows:

[0088] (1) Determine the neighborhood of the target point and find the nearest 4 rows and 4 columns of integer coordinates around the target point (x, y), denoted as: (i1, i2, i3, i4) and (j1, j2, j3, j4), where:

[0089] i2≤x<i3, that is, i2 and i3 are the two integer rows closest to the target point x;

[0090] j2≤y<j3, that is, j2 and j3 are the two integer sequences closest to the target point y;

[0091] (2) Perform one-dimensional cubic spline interpolation on each column, and for each column j k (k=1,2,3,4), use cubic spline interpolation in the row direction to calculate the intermediate values:

[0092]

[0093] Among them, a m is the coefficient of cubic spline interpolation, which depends on the pixel value I(i1, j k ), I(i2,j k ), I(i3,j k ), I(i4,j k ).

[0094] Get 4 intermediate values

[0095] (3) Perform another cubic spline interpolation on the middle value in the row direction and on the middle value in the column direction. Use cubic spline interpolation again in the column direction to calculate the final target value:

[0096]

[0097] Among them, b n are the coefficients of the cubic spline interpolation, depending on

[0098] As a specific implementation of the glioblastoma peritumoral tumor infiltration segmentation model provided in the application, image feature learning also includes supervised learning based on an attention mechanism and a convolutional network (3DUX-Net).

[0099] Combined with specific usage scenarios, supervised learning is performed through attention mechanism and convolutional network (3DUX-Net), so that the model can more clearly distinguish between tumor infiltration and vasogenic edema.

[0100] Among some possible approaches, the supervised learning process based on the attention mechanism and convolutional network (3DUX-Net) includes:

[0101] (1) Data preparation: Acquire three-dimensional magnetic resonance imaging data of non-enhancing gliomas and meningiomas. Non-enhancing glioma lesions are marked as tumor infiltration labels and meningioma lesions are marked as vasogenic edema labels.

[0102] Input image: Where H, W, and D are the height, width, and depth of the image, respectively;

[0103] Annotation label: Y∈(0, 1, 2) H×W×D , where 0 represents normal tissue, 1 represents tumor infiltration area, and 2 represents vasogenic edema area;

[0104] Normalize image pixels to a standard normal distribution: Among them, μ is the image mean and σ is the standard deviation;

[0105] (2) Network design: 3DUX-Net is used as the basic network structure, and the self-attention mechanism is combined to enable the model to better learn the characteristics of tumor infiltration and vasogenic edema. The encoder uses 3D convolution layers to extract multi-scale features and gradually reduce the spatial resolution through downsampling. After each layer of feature extraction, an attention mechanism is added to enhance the focus on tumor and edema areas. The decoder uses an upsampling module to gradually restore the resolution. Through skip connections, the low-level features of the encoder are fused with the high-level features of the decoder to retain detailed information. The output layer generates prediction results through 3D convolution. The value of each pixel represents the probability of belonging to category 0, 1, or 2, and is transformed through a hierarchical Transformer structure with a self-attention mechanism.

[0106] (3) Loss function setting: Dice loss is used to measure the overlap between the segmentation result and the true label. The formula is as follows:

[0107]

[0108] The cross entropy (CE) loss is used to measure the difference between the predicted probability and the true label. The formula is as follows:

[0109]

[0110] The total loss is expressed as:

[0111]

[0112] Among them, α and β represent weight coefficients.

[0113] As a specific implementation of the glioblastoma peritumoral tumor infiltration segmentation model provided in the application, CPC self-learning includes:

[0114] Feature encoding, the encoder network E maps the original data to a low-dimensional feature space and outputs the encoded feature Z:

[0115] Z=E(X)

[0116] Among them, X is the original 3D data, and Z is the encoded feature representation;

[0117] Predictive modeling: The prediction model P predicts future features based on historical features. Assuming that the features of the next k steps are to be predicted, then:

[0118]

[0119] in, is the predicted future feature, Zt：t+k-1 It is a sequence of historical characteristics;

[0120] Contrastive loss construction,Contrastive loss is constructed based on InfoNCE loss, which is used to maximize the similarity between positive samples and minimize the similarity between negative samples. The contrastive loss can be defined as:

[0121]

[0122] Among them, sim(x) is the similarity measurement function, and T is used to adjust the parameters of the contrast loss scale;

[0123] Model training, model training uses the optimizer and contrast loss function to update the parameters of the encoder E and the prediction model P:

[0124]

[0125] Among them, θ is the model parameter, η is the learning rate, is the gradient operator.

[0126] It should be understood that the CPC self-learning part can perform complete feature extraction on the Gb dataset from the beginning, and has a certain ability to distinguish Gb. As a self-supervised learning method, CPC learns the representation of data by maximizing the mutual information between the context (current voxel block) and the target (future voxel block). This method encourages the model to learn feature representations that can predict future voxel blocks, which helps to better obtain and distinguish feature information. Generally speaking, if you train on similar learning tasks from the beginning and then transfer them, the model extraction ability of the network actually depends more on the similar tasks themselves, and the strength of transfer learning is difficult to control, which makes the model's performance in the final target task limited. The introduced CPC self-learning part can build a complete feature representation extraction capability for the target dataset from the beginning.

[0127] As a specific implementation of the glioblastoma peritumoral tumor infiltration segmentation model provided in the application, the 3DUX-Net segmentation model includes:

[0128] Feature extraction, input the preprocessed data into the model respectively, let the multimodal data be X, satisfy Represents data with N samples, C channels, depth Z, height H, and width W, and the extracted feature is F. Similarly, Extract features:

[0129] F=E(X)

[0130] Among them, E represents feature extractor, 3D convolutional neural network;

[0131] Feature fusion, let the fused feature be G, satisfying Fusion features:

[0132]

[0133] in, Represents the fusion method, which attempts element-by-element addition, element-by-element multiplication, or channel concatenation, etc.

[0134] Feature decoding, decode the fused feature G, and set the decoded result to Y, satisfying Feature decoding:

[0135] Y=D(G)

[0136] Where D stands for decoder, 3D transposed convolutional neural network;

[0137] Loss calculation and optimization, let the loss value be L, the true label be T, and calculate the loss:

[0138] L = LossFunction(Y, T)

[0139] Among them, the loss function can use Dice Similarity Coefficient (DSC), cross entropy loss, etc.

[0140] Model training, based on the training data, feature extractor E, fusion mechanism, decoder D, loss function L, use the optimizer; and loss function to update the model parameters to minimize the loss and perform training:

[0141]

[0142] Among them, θ is the model parameter, η is the learning rate, is the gradient operator.

[0143] Furthermore, the feature extraction of the 3DUX-Net segmentation model includes:

[0144] Extract spatial features from multimodal magnetic resonance imaging (MRI) through 3D convolutional layers,

[0145]

[0146] in, represents the output feature map of position (i, j, k) in layer l, represents a 3D convolution kernel of size (m, n, o) in the lth layer, the symbol * represents the convolution operation, and b l is the bias term of the lth layer, and f is the activation function;

[0147] Through the hierarchical Transformer structure of the self-attention mechanism,

[0148]

[0149] Among them, Q, K, and V represent query, key, and value, respectively, which are usually obtained from the input feature map through linear transformation. k is the dimension of the key vector, used to scale the attention scores, and softmax is the normalization function.

[0150] As a specific implementation of the glioblastoma peritumoral tumor infiltration segmentation model provided in the application, the segmentation accuracy judgment method in the segmentation model verification includes Dice similarity coefficient, Hausdorff distance and intersection-over-union ratio.

[0151] Combined with specific usage scenarios, the results of the trained segmentation model are evaluated based on pathological validation, which is based on stereotactic biopsy specimens in pathological information acquisition, including biopsy points under neuronavigation. A peritumoral predictive spatial probability map is created based on the segmentation model, and the location of the biopsy point is recorded using neuronavigation. The prediction result is estimated by the average probability of the voxels corresponding to the biopsy location. The pathological diagnosis is a double-blind parallel test, and the biopsy tissue is measured without the relevant professionals knowing the prediction results of the infiltration segmentation model. A comprehensive evaluation is performed by observing cell morphology, mitotic count, and Ki-67 labeling index, which are divided into normal tissue, suspected tumor cell infiltration, and clear tumor cell infiltration. Normal tissue or suspected tumor cell infiltration is regarded as pure vascular edema. The effectiveness of the segmentation model is evaluated by comparing the prediction results with the pathological diagnosis.

[0152] In some feasible ways, T is used to represent the true label and B is used to represent the predicted value. The range of Dice similarity coefficient DSC is [0, 1], which effectively measures the overlap between the segmentation result and the true label, and is defined as follows:

[0153]

[0154] The Hausdorff distance is used to measure the distance between any two sets in space. It is sensitive to the boundaries of the segmentation results and is defined as follows:

[0155] d H (X, Y) = max{max x∈X min y∈Y d(x,y),max y∈Y min x∈X d(x,y)}

[0156] Among them, d H(X, Y) represents the bidirectional Hausdorff distance, and d(x, y) represents the distance between any two points in X and Y.

[0157] The intersection over union (IoU) is the most commonly used evaluation indicator for tasks such as target detection, semantic segmentation, and tracking, and is defined as follows:

[0158]

[0159] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a glioblastoma peritumoral infiltration segmentation model, characterized in that: include: Clinical case collection: design inclusion and exclusion criteria for glioblastoma, non-enhancing glioma, and meningioma, and collect cases that meet the requirements; Image information acquisition: using a 3TMR scanner to acquire case images, including conventional MRI plain scan, multi-b-value diffusion-weighted imaging scan and high-resolution 3DT1WI enhanced scan; Pathological information collection, including stereotactic biopsy information of glioblastoma patients and cell morphology, mitotic count, and Ki-67 labeling index under pathological examination; Image data preprocessing, parameter map calculation and registration of images, and manual segmentation of regions of interest, followed by image data standardization; Image feature learning: using contrast prediction coding to self-learn the Gb peritumoral edema area and extract complete features from the Gb dataset; Establish an invasion segmentation model, use the 3DUX-Net segmentation model for transfer learning, and build a peritumoral invasion segmentation model; Segmentation model validation: double-blind analysis and prediction of cases were performed using manual and peritumoral infiltration segmentation models, and the analysis results were compared to evaluate the effectiveness of the model.

2. The glioblastoma peritumoral tumor infiltration segmentation model according to claim 1, characterized in that: The image data standardization includes: Image normalization, converting the pixel values ​​of image data to a specific range so that different images have similar pixel value ranges; Image cropping: extracting the region of interest from the 3D image data based on the image center and cropping it to a uniform size; Resampling, calculate the size ratio or resolution ratio of the original image and the target image, in each dimension, calculate the position of the pixel position in the target image in the original image by interpolation method, and use the corresponding interpolation method to calculate the pixel value.

3. The glioblastoma peritumoral tumor infiltration segmentation model according to claim 1, characterized in that: The image feature learning also includes supervised learning based on attention mechanism and convolutional network.

4. The glioblastoma peritumoral tumor infiltration segmentation model according to claim 1, characterized in that: The CPC self-learning includes: Feature encoding, the encoder network E maps the original data to a low-dimensional feature space and outputs the encoded feature Z: Z=E(X) Among them, X is the original 3D data, and Z is the encoded feature representation; Predictive modeling: The prediction model P predicts future features based on historical features. Assuming that the features of the next k steps are to be predicted, then: in, is the predicted future feature, Z t：t+k-1 It is a sequence of historical characteristics; Contrastive loss construction,Contrastive loss is constructed based on InfoNCE loss, which is used to maximize the similarity between positive samples and minimize the similarity between negative samples. The contrastive loss can be defined as: Among them, sim(x) is the similarity measurement function, and T is used to adjust the parameters of the contrast loss scale; Model training, model training uses the optimizer and contrast loss function to update the parameters of the encoder E and the prediction model P: Among them, θ is the model parameter, η is the learning rate, is the gradient operator.

5. The glioblastoma peritumoral tumor infiltration segmentation model according to claim 1, characterized in that: The 3DUX-Net segmentation model includes, Feature extraction, input the preprocessed data into the model respectively, let the multimodal data be X, satisfy Represents data with N samples, C channels, depth Z, height H, and width W, and the extracted feature is F. Similarly, Extract features: F=E(X) Among them, E represents feature extractor, 3D convolutional neural network; Feature fusion, let the fused feature be G, satisfying Fusion features: in, Represents the fusion method, which attempts element-by-element addition, element-by-element multiplication, or channel concatenation, etc. Feature decoding, decode the fused feature G, and set the decoded result to Y, satisfying Feature decoding: Y=D(G) Where D stands for decoder, 3D transposed convolutional neural network; Loss calculation and optimization, let the loss value be L, the true label be T, and calculate the loss: L = LossFunction(Y, T); Model training, based on the training data, feature extractor E, fusion mechanism, decoder D, loss function L, use the optimizer; and loss function to update the model parameters to minimize the loss and perform training: Among them, θ is the model parameter, η is the learning rate, is the gradient operator.

6. The glioblastoma peritumoral tumor infiltration segmentation model according to claim 5, characterized in that: The feature extraction includes: Extract spatial features from multimodal MRI via 3D convolutional layers. in, represents the output feature map of position (i, j, k) in layer l, represents a 3D convolution kernel of size (m, n, o) in the lth layer, the symbol * represents the convolution operation, and b l is the bias term of the lth layer, and f is the activation function; Through the hierarchical Transformer structure of the self-attention mechanism, Among them, Q, K, and V represent query, key, and value, respectively, which are usually obtained from the input feature map through linear transformation. k is the dimension of the key vector, used to scale the attention scores, and softmax is the normalization function.

7. The glioblastoma peritumoral tumor infiltration segmentation model according to claim 1, characterized in that: The segmentation accuracy judgment method in the segmentation model verification includes Dice similarity coefficient, Hausdorff distance and intersection-over-union ratio.

8. The glioblastoma peritumoral tumor infiltration segmentation model according to claim 1, characterized in that: At least two different manual segmentations are used in the manual segmentation of the region of interest, and the difference in the overlapping area of ​​the two ROIs should not be greater than 5%.