A glioblastoma peritumoral tumor infiltration segmentation model construction method

Abstract

The application discloses a glioblastoma peritumoral tumor infiltration segmentation model, which comprises clinical case collection, image information collection, pathological information collection, image data preprocessing, image feature learning based on CPC for self-learning of a Gb peritumoral edema region, migration learning by using a 3DUX-Net segmentation model, construction of a peritumoral infiltration segmentation model, and segmentation model verification. The application learns the artificial ROI segmentation features, establishes the peritumoral infiltration segmentation model, verifies the effectiveness of the model through double-blind testing, and accurately segments the peritumoral tumor infiltration through the peritumoral infiltration segmentation model, thereby providing an important reference for clinical multidisciplinary diagnosis and treatment.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging technology and relates to a method for constructing a segmentation model of peritumoral tumor infiltration in glioblastoma. Background Technology

[0002] Peritumoral edema in glioblastoma (Gb) is mainly composed of tumor infiltration and vasogenic 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 enhancing margin. Preoperatively, it is difficult to accurately identify tumor infiltration using imaging technology, leading to incomplete surgical resection. This is the main reason for the poor prognosis and high recurrence rate of Gb.

[0003] Tumor invasion in current preoperative fractionation planning for treatment (PTR) can be summarized into two research trends: exploring imaging signs or markers of tumor invasion based on multimodal imaging, and determining the invasion boundaries of gliomas based on artificial intelligence models and technologies. However, in clinical practice, it is not possible to directly obtain a large number of accurate tumor cell invasion markers and pure angioedema markers from Gb, thus making it impossible to accurately segment peritumoral tumor invasion using limited invasion segmentation models. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method for constructing a segmentation model of peritumoral tumor infiltration in glioblastoma, which effectively solves the problems in the prior art.

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

[0006] Methods for constructing a segmentation model of peritumoral tumor invasion in glioblastoma include:

[0007] Clinical case collection involved designing inclusion and exclusion criteria for glioblastoma, non-enhancing glioma, and meningioma, and collecting cases that met the requirements.

[0008] Image information acquisition: The images of the case were acquired using a 3TMR scanner. The acquisition included conventional MRI plain scan, multi-b-value diffusion-weighted imaging scan, and high-resolution 3DT1WI enhanced scan.

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

[0010] Image data preprocessing involves parametric map calculation and registration of the images, manual segmentation of regions of interest, and subsequent standardization of the image data.

[0011] Image feature learning: a contrastive predictive coding method is used to learn the Gb peritumoral edema region, and complete feature extraction is performed on the Gb dataset.

[0012] An infiltration segmentation model was established, and a 3DUX-Net segmentation model was used for transfer learning to construct a peritumoral infiltration segmentation model.

[0013] The segmentation model was validated by performing a double-blind analysis of cases using both manual segmentation and peritumoral infiltration segmentation models. The results were compared to evaluate the effectiveness of the model.

[0014] Optionally, the image data standardization includes:

[0015] Image normalization transforms the pixel values ​​of image data to a specific range so that different images have similar ranges of pixel values.

[0016] Image cropping involves extracting the region of interest from the center of a 3D image and cropping it to a uniform size.

[0017] Resampling involves calculating the size or resolution ratio between the original and target images. In each dimension, the position of a pixel in the target image is calculated in the original image using an interpolation method, and the pixel value is calculated using the corresponding interpolation method.

[0018] Optionally, the image feature learning also includes supervised learning based on attention mechanisms and convolutional networks.

[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 features Z. ,

[0021] in, It is the original 3D data. It is the encoded feature representation;

[0022] Predictive modeling: Predictive model P predicts future features based on historical features. Suppose we want to predict features for the next k steps, then:

[0023] ,

[0024] in, It is a predicted future characteristic. It is a historical feature sequence;

[0025] The contrastive loss is constructed based on the InfoNCE loss and is used to maximize the similarity between positive samples and minimize the similarity between negative samples. The contrastive loss can be defined as follows:

[0026] ,

[0027] in, (x) is the similarity measurement function. Parameters used to adjust the scale of contrast loss;

[0028] Model training uses an optimizer and a contrastive loss function to update the encoder. and prediction models Parameters:

[0029] ,

[0030] in, These are model parameters. It's the learning rate. It is the gradient operator.

[0031] Optionally, the 3DUX-Net segmentation model includes,

[0032] Feature extraction involves inputting the preprocessed data into the model. Let the multimodal data be X, satisfying... , indicating that it has One sample Each channel, depth ,high and width The data, with extracted feature F, similarly satisfies Feature extraction:

[0033] ,

[0034] in, Representative feature extractor, 3D convolutional neural network;

[0035] Feature fusion, let the fused features be... ,satisfy Fusion features:

[0036] ,

[0037] in, The proposed fusion methods include element-by-element addition, element-by-element multiplication, or channel splicing.

[0038] Feature decoding, for the fused features Perform decoding, and let the decoded result be... ,satisfy Feature decoding:

[0039] ,

[0040] in, Represents a decoder, a 3D transposed convolutional neural network;

[0041] Loss calculation and optimization: Let the loss value be L, and the true label be... Calculate the loss:

[0042] ;

[0043] Model training, based on training data and feature extractors Fusion mechanism, decoder Loss function We use an optimizer and a loss function to update the model parameters to minimize the loss during training.

[0044] ,

[0045] in, These are model parameters. It's the learning rate. It is the gradient operator.

[0046] Optionally, the feature extraction includes,

[0047] Spatial features are extracted from multimodal MRI using 3D convolutional layers.

[0048] ,

[0049] in, Indicates the first Location in the layer ( The output feature map of ) Indicates the first The size of the layer is ( 3D convolution kernel, symbol This represents the convolution operation. It is the first Layer bias terms, It is an activation function;

[0050] Through a hierarchical Transformer structure with a self-attention mechanism,

[0051] ,

[0052] in, These represent the query, key, and value, respectively, and are typically obtained from the input feature map through a linear transformation. It is the dimension of the key vector, used to scale the attention score. It is a normalization function.

[0053] Optionally, the methods for judging the accuracy of segmentation in the segmentation model verification include Dice similarity coefficient, Hausdorff distance, and intersection-union ratio.

[0054] Optionally, at least two different people are used to perform 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%.

[0055] Compared with existing technologies, the present invention has the following beneficial effects: by learning the artificial ROI segmentation features, a peritumoral invasion segmentation model is established, and the effectiveness of the model is verified through a double-blind test. The peritumoral invasion segmentation model provides an important reference for the accurate segmentation of peritumoral tumor invasion, enabling multidisciplinary clinical diagnosis and treatment. Detailed Implementation

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The method for constructing a peritumoral tumor invasion segmentation model for glioblastoma disclosed in this invention includes:

[0058] Clinical case collection involved designing inclusion and exclusion criteria for glioblastoma, non-enhancing glioma, and meningioma, and collecting cases that met the requirements.

[0059] Image information acquisition: The images of the case were acquired using a 3TMR scanner, including conventional magnetic resonance imaging (MRI) plain scan, multi-b-value diffusion-weighted imaging (DWI) scan, and high-resolution 3DT1WI enhanced scan.

[0060] Pathological information collection includes stereotactic biopsy information from glioblastoma patients, as well as cell morphology, mitotic count, and Ki-67 labeling index under pathological examination. Pathological information is also used for segmentation model validation.

[0061] Image data preprocessing involves parametric map calculation and registration of the images, manual segmentation of regions of interest (ROIs), and subsequent standardization of the image data.

[0062] Image feature learning uses contrastive predictive coding (CPC) to learn the Gb peritumoral edema region and extract complete features from the Gb dataset.

[0063] An infiltration segmentation model was established, and a 3DUX-Net segmentation model was used for transfer learning to construct a peritumoral infiltration segmentation model.

[0064] The segmentation model was validated by performing a double-blind analysis of cases using both manual segmentation and peritumoral infiltration segmentation models. The results were compared to evaluate the effectiveness of the model.

[0065] Specifically, the process involves collecting clinical cases, selecting suitable cases for imaging and pathological information collection, manually segmenting regions of interest (ROIs), extracting features using relevant software, training a peritumoral infiltration segmentation model, and finally comparing the results of the peritumoral infiltration segmentation model analysis and prediction with those of manual analysis and prediction under double-blind conditions to evaluate the effectiveness of the model.

[0066] In this way, by learning the segmentation features of artificial ROIs, a peritumoral invasion segmentation model is established. The effectiveness of the model is verified through double-blind testing. The accurate segmentation of peritumoral tumor invasion using the peritumoral invasion segmentation model provides an important reference for multidisciplinary clinical diagnosis and treatment.

[0067] In some feasible approaches, clinical case collection involves inclusion and screening of cases. Inclusion criteria include age group, no prior radiotherapy or chemotherapy, clear imaging, and postoperative pathological confirmation of the disease. Exclusion criteria include missing necessary MRI sequences, use of unconventional imaging parameters, and low-quality images.

[0068] For image acquisition, a 64-channel head and neck coil 3TMR scanner can be used. Conventional MRI plain scan sequences include T1WI, T2WI, and FLAIR. During high-resolution 3DT1WI contrast-enhanced scanning, gadopentetate dimeglumine is injected via the antecubital vein. For ease of implementation, this embodiment provides the following scanning parameters for image acquisition:

[0069] The scanning parameters for axial T1WI are: 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 scan time 37 seconds.

[0070] The scanning parameters for axial T2WI are: TR 4090 ms, TE 99 ms, matrix 733×733, field of view 220 mm×220 mm, slice thickness 5 mm, number of slices 20, and scanning time 34 seconds.

[0071] The axial FLAIR scanning parameters were: TR 8000 ms, TE 81 ms, matrix 314×314, field of view 220 mm×220 mm, slice thickness 5 mm, 20 slices, and scanning time 1 minute 38 seconds. The 3D-T1-MPRAGE scanning parameters were: TR 2300 ms, TE 2.32 ms, matrix 266×266, field of view 240 mm×240 mm, slice thickness 0.9 mm, 176 slices, and scanning time 5 minutes 21 seconds.

[0072] Magnetic resonance diffusion-weighted imaging (DWI) uses a single-excitation planar echo sequence and six b values ​​(0, 500, 1000, 1500, 2000 and 2500 s / mm2) for axial scanning. Each non-zero b value is performed in 30 coding directions. The TR is 3500 ms, TE is 81 ms, the matrix is ​​383×383, the field of view is 230 mm×230 mm, the slice thickness is 5 mm, the number of slices is 20, and the scan time is 54 seconds.

[0073] In some feasible approaches, image data preprocessing includes parametric map calculation, image registration and segmentation, and image data standardization. In this embodiment, the sequence and parametric map are provided as shown in Table 1, wherein the sagittal 3D-T1-MPRAGE image is reconstructed into 20 layers of axial, sagittal and coronal images.

[0074] Table 1 Sequence Parameter Diagram

[0075] Sequence (number) Parameter diagram Conventional MRI (4) T1WI, T2WI, FLAIR, 3D-T1-MPRAGE DMRI (23) DWI:ADCDTI:AD,FA,MD,RDDKI:AD,AK,FA,MD,MK,RD,RKNODDI:ICVF,ISOVF,ODIMAP:MSD,NG,NGAx,NGRad,QIV,RTAP,RTOP,RTPP

[0076] DMRI parametric map calculation: Import DWI images into the post-processing software NeuDilab (based on DIPY) to calculate the parametric maps of DKI, DTI, NODDI, and MAP from the DWI images. The meaning of each parameter is based on existing technology and will not be elaborated here.

[0077] During image registration, the Automatic registration mode of the open-source software ITK-SNAP was used, with FLAIR as the template, to register the various parameter maps of other regular sequences and diffusion sequences. After registration, the images of each sequence have the same voxel size, initial image position, and resolution.

[0078] In ROI segmentation, experienced professionals in the relevant discipline should manually delineate the peritumoral edema area, including infiltrative and vasogenic edema areas, on glioblastomas; delineate the peritumoral infiltrative area on non-enhancing gliomas; and delineate the edema area, generally considered to be the vasogenic edema area, on meningiomas. When delineating ROIs, the corresponding standard sequences should be referenced.

[0079] In some feasible approaches, to compensate for human segmentation bias, if the difference between 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 party makes the judgment or the images are re-segmented for comparison. The 5% figure is a relatively suitable empirical value used in the experiment.

[0080] As a specific implementation of the method for constructing a segmentation model of peritumoral tumor invasion in glioblastoma provided in the application, the standardization of image data includes:

[0081] Image normalization transforms the pixel values ​​of image data to a specific range so that different images have similar ranges of pixel values.

[0082] Image cropping involves extracting the region of interest (ROI) from the 3D image data based on the image center and cropping it to a uniform size.

[0083] Resampling involves calculating the size or resolution ratio between the original and target images. In each dimension, the position of a pixel in the target image is calculated in the original image using an interpolation method, and the pixel value is calculated using the corresponding interpolation method.

[0084] Overall, standardizing image data, ensuring similar pixel value ranges across different images, and regulating image size, while emphasizing the region of interest (ROI) in the image center, can facilitate subsequent processing and improve the accuracy of feature extraction.

[0085] In some feasible approaches, spline interpolation can be selected as the interpolation method in this embodiment. Higher-order spline interpolation is a commonly used interpolation method due to its smoothness and high accuracy, making it suitable for scenarios requiring precise resampling or transformation, such as image registration, reconstruction, and post-segmentation processing. Spline interpolation calculates interpolation values ​​in the neighborhood of the target pixel based on a spline function. In a two-dimensional image, higher-order spline interpolation utilizes the values ​​of multiple pixels surrounding the target pixel location and calculates the target pixel value by fitting a spline function. This embodiment uses cubic spline interpolation. In a two-dimensional image, cubic spline interpolation is a process of performing one-dimensional cubic interpolation on both rows and columns (i.e., separate axis interpolation). Assume the coordinates of the target point are... The calculation steps are as follows:

[0086] (1) Determine the neighborhood of the target point and find the target point. The nearest integer coordinates in the surrounding 4 rows and 4 columns are denoted as: and ,in,

[0087] ,Right now The target point The two most recent integer rows;

[0088] ,Right now The target point The two most recent integer sequences;

[0089] (2) Perform one-dimensional cubic spline interpolation on each column. The intermediate value is calculated using cubic spline interpolation in the row direction:

[0090] ,

[0091] in, These are the coefficients of cubic spline interpolation, which depend on the pixel values. .

[0092] Four intermediate values ​​were obtained. ;

[0093] (3) Perform another cubic spline interpolation on the median value in the row direction, and on the median value in the column direction... Then, cubic spline interpolation is applied again in the column direction to calculate the final target value:

[0094] ,

[0095] in, These are the coefficients of the cubic spline interpolation, which depend on... .

[0096] As a specific implementation of the method for constructing a segmentation model of peritumoral tumor infiltration in glioblastoma provided in the application, image feature learning also includes supervised learning based on attention mechanism and convolutional network (3DUX-Net).

[0097] By combining specific use cases, supervised learning is carried out through attention mechanisms and convolutional networks (3DUX-Net), enabling the model to more clearly distinguish between tumor infiltration and vasogenic edema.

[0098] Among the feasible approaches, supervised learning processes based on attention mechanisms and convolutional networks (3DUX-Net) include:

[0099] (1) Data preparation: three-dimensional magnetic resonance imaging data of non-enhancing glioma and meningioma were obtained. Non-enhancing glioma lesions were marked as tumor infiltration tags and meningioma lesions were marked as vasogenic edema tags.

[0100] Input image: , where H, W, and D are the height, width, and depth of the image, respectively;

[0101] Labeling: Where 0 represents normal tissue, 1 represents the tumor infiltration area, and 2 represents...

[0102] Areas of vasogenic edema;

[0103] Normalize the image pixels to a standard normal distribution: ,in, It is the image mean. It is the standard deviation;

[0104] (2) Network design: 3DUX-Net is used as the basic network structure, combined with a self-attention mechanism, enabling the model to better learn the features of tumor infiltration and vasogenic edema. The encoder uses 3D convolutional layers to extract multi-scale features, gradually reducing spatial resolution through downsampling. After feature extraction at each layer, an attention mechanism is added to enhance attention to tumor and edema regions. The decoder uses an upsampling module to gradually restore resolution, fusing low-level features from the encoder with high-level features from the decoder through skip connections to preserve 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 implemented through a hierarchical Transformer structure with a self-attention mechanism.

[0105] The loss function is set using Dice loss to measure the overlap between the segmentation result and the true label. The formula is as follows:

[0106] ,

[0107] Cross-entropy (CE) loss is used to measure the difference between the predicted probability and the true label. The formula is as follows:

[0108] ,

[0109] The total loss is expressed as:

[0110] ,

[0111] in, and This represents the weighting coefficient.

[0112] As a specific implementation of the method for constructing a peritumoral tumor invasion segmentation model for glioblastoma provided in the application, CPC self-learning includes:

[0113] Feature encoding: The encoder network E maps the original data to a low-dimensional feature space and outputs the encoded features Z.

[0114] ,

[0115] in, It is the original 3D data. It is the encoded feature representation;

[0116] Predictive modeling: Predictive model P predicts future features based on historical features. Suppose we want to predict features for the next k steps, then:

[0117] ,

[0118] in, It is a predicted future characteristic. It is a historical feature sequence;

[0119] The contrastive loss is constructed based on the InfoNCE loss and is used to maximize the similarity between positive samples and minimize the similarity between negative samples. The contrastive loss can be defined as follows:

[0120] ,

[0121] in, (x) is the similarity measurement function. Parameters used to adjust the scale of contrast loss;

[0122] Model training uses an optimizer and a contrastive loss function to update the encoder. and prediction models Parameters:

[0123] ,

[0124] in, These are model parameters. It's the learning rate. It is the gradient operator.

[0125] It should be understood that the CPC self-learning component can perform complete feature extraction on the Gb dataset from the outset, providing a certain level of discriminative ability for Gb. Furthermore, as a self-supervised learning method, CPC learns data representations 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, helping to better acquire and distinguish feature information. Typically, if the model is initially trained on similar learning tasks and then transferred, its model extraction ability relies more on the similarity of the tasks themselves, and the strength of transfer learning is difficult to control, limiting the model's performance on the final target task. The introduced CPC self-learning component, however, can build a complete feature representation extraction capability for the target dataset from the outset.

[0126] As a specific implementation of the method for constructing a segmentation model of peritumoral tumor infiltration in glioblastoma provided in the application, the 3DUX-Net segmentation model includes:

[0127] Feature extraction involves inputting the preprocessed data into the model. Let the multimodal data be X, satisfying... , indicating that it has One sample Each channel, depth ,high and width The data, with extracted feature F, similarly satisfies Feature extraction:

[0128] ,

[0129] in, Representative feature extractor, 3D convolutional neural network;

[0130] Feature fusion, let the fused features be... ,satisfy Fusion characteristics:

[0131] ,

[0132] in, The proposed fusion methods include element-by-element addition, element-by-element multiplication, or channel splicing.

[0133] Feature decoding, for the fused features Perform decoding, and let the decoded result be... ,satisfy Feature decoding:

[0134] ,

[0135] in, Represents a decoder, a 3D transposed convolutional neural network;

[0136] Loss calculation and optimization: Let the loss value be L, and the true label be... Calculate the loss:

[0137] ,

[0138] The loss function can be the Dice similarity coefficient (DSC), cross-entropy loss, etc.

[0139] Model training, based on training data and feature extractors Fusion mechanism, decoder Loss function We use an optimizer and a loss function to update the model parameters to minimize the loss during training.

[0140] ,

[0141] in, These are model parameters. It's the learning rate. It is the gradient operator.

[0142] Furthermore, the feature extraction of the 3DUX-Net segmentation model includes:

[0143] Spatial features are extracted from multimodal magnetic resonance imaging (MRI) using 3D convolutional layers.

[0144] ,

[0145] in, Indicates the first Location in the layer ( The output feature map of ) Indicates the first The size of the layer is ( 3D convolution kernel, symbol This represents the convolution operation. It is the first Layer bias terms, It is an activation function;

[0146] Through a hierarchical Transformer structure with a self-attention mechanism,

[0147] ,

[0148] in, These represent the query, key, and value, respectively, and are typically obtained from the input feature map through a linear transformation. It is the dimension of the key vector, used to scale the attention score. It is a normalization function.

[0149] As a specific implementation of the method for constructing a segmentation model of peritumoral tumor infiltration in glioblastoma provided in the application, the methods for judging the accuracy of segmentation in the segmentation model verification include Dice similarity coefficient, Hausdorff distance, and intersection-union ratio.

[0150] In conjunction with specific application scenarios, the results of the trained segmentation model are evaluated based on pathological validation. Pathological validation is based on stereotactic biopsy specimens from pathological information collection, including biopsy points under neuronavigation. A peritumoral predictive spatial probability map is created based on the segmentation model, and the location of biopsy points is recorded using neuronavigation. The prediction results are estimated using the average probability of the voxels corresponding to the biopsy locations. Pathological diagnosis is a double-blind parallel test, where biopsy tissue is measured without the knowledge of the infiltration segmentation model's prediction results by relevant professionals. A comprehensive evaluation is conducted by observing cell morphology, mitotic count, and Ki-67 labeling index, categorizing the infiltration into normal tissue, suspicious tumor cell infiltration, and confirmed tumor cell infiltration. Normal tissue or suspicious tumor cell infiltration is considered as pure angioedema. The effectiveness of the segmentation model is evaluated by comparing the prediction results with the pathological diagnosis.

[0151] In some feasible ways, using Indicates the true label, This represents the predicted value. The Dice similarity coefficient (DSC) ranges from [0,1] and effectively measures the overlap between the segmentation result and the true label. It is defined as follows:

[0152] ,

[0153] Hausdorff distance is used to measure the distance between any two sets in a space. It is quite sensitive to the boundaries of the segmentation result and is defined as follows:

[0154] ,

[0155] in, Indicates the bidirectional Hausdorff distance. This represents the distance between any two points between X and Y.

[0156] Intersection over Union (IoU) is the most commonly used evaluation metric for tasks such as object detection, semantic segmentation, and tracking. It is defined as follows:

[0157] .

[0158] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a segmentation model of peritumoral tumor infiltration in glioblastoma, characterized in that, include: Clinical case collection involved designing inclusion and exclusion criteria for glioblastoma, non-enhancing glioma, and meningioma, and collecting cases that met the requirements. Image information acquisition: The images of the case were acquired using a 3TMR scanner. The acquisition included conventional MRI plain scan, multi-b-value diffusion-weighted imaging scan, and high-resolution 3DT1WI enhanced scan. Pathological information collection, which includes stereotactic biopsy information of glioblastoma patients and cell morphology, mitotic count, and Ki-67 labeling index under pathological examination; Image data preprocessing involves parametric map calculation and registration of the images, manual segmentation of regions of interest, and subsequent standardization of the image data. Image feature learning: a contrastive predictive coding method is used to learn the Gb peritumoral edema region, and complete feature extraction is performed on the Gb dataset. An infiltration segmentation model was established, and a 3DUX-Net segmentation model was used for transfer learning to construct a peritumoral infiltration segmentation model. Segmentation model validation: A double-blind analysis of cases was conducted using both manual and peritumoral infiltration segmentation models. The analysis results were compared to evaluate the effectiveness of the model. The self-learning method using contrastive predictive coding includes: Feature encoding: The encoder network E maps the original data to a low-dimensional feature space and outputs the encoded features Z. , in, It is the original 3D data. It is the encoded feature representation; Predictive modeling: Predictive model P predicts future features based on historical features. Suppose we want to predict features for the next k steps, then: , in, It is a predicted future characteristic. It is a historical feature sequence; The contrastive loss is constructed based on the InfoNCE loss and is used to maximize the similarity between positive samples and minimize the similarity between negative samples. The contrastive loss can be defined as follows: , in, (x) is the similarity measurement function. Parameters used to adjust the scale of contrast loss; Model training uses an optimizer and a contrastive loss function to update the encoder. and prediction models Parameters: , in, These are model parameters. It's the learning rate. It is the gradient operator.

2. The method for constructing a peritumoral tumor infiltration segmentation model for glioblastoma according to claim 1, characterized in that: The image data standardization includes: Image normalization transforms the pixel values ​​of image data to a specific range so that different images have similar ranges of pixel values. Image cropping involves extracting the region of interest from the center of a 3D image and cropping it to a uniform size. Resampling involves calculating the size or resolution ratio between the original and target images. In each dimension, the position of a pixel in the target image is calculated in the original image using an interpolation method, and the pixel value is calculated using the corresponding interpolation method.

3. The method for constructing a peritumoral tumor infiltration segmentation model for glioblastoma according to claim 1, characterized in that: The image feature learning also includes supervised learning based on attention mechanisms and convolutional networks.

4. The method for constructing a peritumoral tumor infiltration segmentation model for glioblastoma according to claim 1, characterized in that: The 3DUX-Net segmentation model includes, Feature extraction involves inputting the preprocessed data into the model. Let the multimodal data be X, satisfying... , indicating that it has One sample Each channel, depth ,high and width The data, with extracted feature F, similarly satisfies Feature extraction: ， in, Represents a feature extractor, a 3D convolutional neural network; Feature fusion, let the fused features be... ,satisfy Fusion features: ， in, The proposed fusion methods include element-by-element addition, element-by-element multiplication, or channel splicing. Feature decoding, for the fused features Perform decoding, and let the decoded result be... ,satisfy Feature decoding: , in, Represents a decoder, a 3D transposed convolutional neural network; Loss calculation and optimization: Let the loss value be L, and the true label be... Calculate the loss: ; Model training, based on training data and feature extractors Fusion mechanism, decoder Loss function We use an optimizer and a loss function to update the model parameters to minimize the loss during training. , in, These are model parameters. It's the learning rate. It is the gradient operator.

5. The method for constructing a peritumoral tumor infiltration segmentation model for glioblastoma according to claim 4, characterized in that: The feature extraction includes, Spatial features are extracted from multimodal MRI using 3D convolutional layers. ， in, Indicates the first Location in the layer ( The output feature map of ) Indicates the first The size of the layer is ( 3D convolution kernel, symbol This represents the convolution operation. It is the first Layer bias terms, It is an activation function; Through a hierarchical Transformer structure with a self-attention mechanism, ， in, These represent the query, key, and value, respectively, and are typically obtained from the input feature map through a linear transformation. It is the dimension of the key vector, used to scale the attention score. It is a normalization function.

6. The method for constructing a peritumoral tumor infiltration segmentation model for glioblastoma according to claim 1, characterized in that: The methods for judging the accuracy of segmentation in the segmentation model verification include Dice similarity coefficient, Hausdorff distance, and intersection-union ratio.

7. The method for constructing a peritumoral tumor infiltration segmentation model for glioblastoma according to claim 1, characterized in that: In the manual segmentation of the region of interest, at least two different manual segmenters are used, and the difference in the overlapping area of ​​the two ROIs should not exceed 5%.

Images