A method, device, storage medium and equipment for predicting brain tumor infiltration degree

Abstract

The application discloses a brain tumor infiltration degree prediction method, device, storage medium and equipment, and belongs to the technical field of medical image processing. The application innovatively takes tumor segmentation probability and fiber density features of a tumor edge region as infiltration analysis features, constructs a brain tumor infiltration degree prediction model, device, storage medium and equipment, and is used for predicting the infiltration degree of a brain tumor. Experiments show that the brain tumor infiltration degree predicted by the model has a relationship (P<0.05) with survival, and the survival time of the high-infiltration degree is obviously lower than that of the low-infiltration degree, which has important clinical significance.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method, apparatus, storage medium, and device for predicting the degree of brain tumor infiltration. Background Technology

[0002] In neurosurgery, the extent of resection plays a crucial role in surgical outcomes, especially in the treatment of malignant tumors. Gliomas are the most common primary brain tumors in adults, exhibiting infiltrative growth with indistinct boundaries between the lesion and surrounding normal brain tissue. High-grade gliomas, in particular, are often accompanied by varying degrees of peritumoral edema. Previous pathological studies have pathologically confirmed that tumor cells also infiltrate within the peritumoral edema zone of gliomas. If the surgical resection does not reach the tumor infiltration boundary, the remaining tumor cells can become a potential risk of recurrence. Therefore, clearly defining the tumor infiltration boundary and determining the nature of the peritumoral edema area are essential steps in reducing the recurrence rate of glioma patients.

[0003] With the rapid development of artificial intelligence (AI) technology and the continuous accumulation of medical big data, AI is being widely applied in the medical field. AI possesses advantages over humans in processing big data, complex and uncertain data, and in deeply mining the potential information within the data. Building disease classification and prediction models based on AI technology is gradually becoming a hot topic and trend in precision medicine in the big data era. Neurological diseases face challenges such as complex etiologies, high diagnostic difficulty, significant individual differences in treatment, and a lack of standardization and evidence-based medicine. In recent years, the development of AI applications in neurological diseases has shifted from a single-mode approach to a comprehensive disease diagnosis and treatment model. Machine learning has powerful capabilities in analyzing multi-dimensional and complex big data, enabling the integrated analysis of multimodal data on neurological diseases and improving comprehensive diagnosis and treatment capabilities.

[0004] The application of AI in the medical field holds great promise, but numerous challenges exist in clinical practice. AI algorithms possess considerable modeling capabilities on training sets, but their performance and stability need improvement when applied to new, unknown data. AI relies on large amounts of data for training; compared to commercial applications in the technological field, the amount of data in the medical industry is relatively small, making results prone to error. Furthermore, AI algorithms, particularly deep learning, suffer from poor interpretability, requiring an understanding of the factors considered in the model's decision-making process. Summary of the Invention

[0005] To address the above problems, the present invention provides a method, apparatus, storage medium, and device for predicting the degree of brain tumor infiltration.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention provides a method for constructing a model to predict the extent of brain tumor invasion, characterized by the following specific steps:

[0008] The first step is data collection: collecting image data of patients with gliomas, including T1, T2, FL, T1CE, and DTI sequences of MRI images;

[0009] The second step is image preprocessing: standardized procedures are used to process MRI images and align multimodal images.

[0010] The third step is tumor segmentation: a pre-trained NN-Unet brain tumor segmentation model is used to segment the tumor region based on the four input sequences T1, T2, FL, and T1CE, and the tumor region foreground probability value of each pixel is output; a segmentation mask is extracted using a threshold of 0.5.

[0011] The fourth step is feature extraction: the probability distribution of the segmentation boundary, the tumor boundary, and the tissue fiber condition near the edge are used as features;

[0012] Step 5: Construction of the infiltration degree prediction model: Based on the features extracted above, the model is constructed using the degree of infiltration as a label, high infiltration as a positive sample and low infiltration as a negative sample, and an SVM classifier is used to construct the brain tumor infiltration degree prediction model.

[0013] Furthermore, the specific steps of the second image preprocessing step are as follows:

[0014] 1) For T1CE sequences, the scan orientation is automatically read, corrected to LPS, and N4 field correction and head correction are performed. Head correction here is primarily based on rigid registration from T1CE to the standard template (SRI-24 atlas) achieved using ants.

[0015] 2) Then, the scan direction correction and N4 field correction are performed on the T1, T2, and FL sequences, and then rigid registration is used to register the T1, T2, and FL sequences onto T1CE;

[0016] 3) Use a pre-trained Unet model to debone the T1, T2, FL, and T1CE sequences;

[0017] 4) For DTI sequences, first extract the b0 image, then use FSL software to debone the DTI data, and finally rigidly register the b0 of the deboned DTI onto T1CE.

[0018] Furthermore, the specific steps of the fourth step of feature extraction are as follows:

[0019] 1) Based on the edge characteristics of the tumor mask: The mask is segmented by traversing a 3*3*3 window. If there is a background value in the window and the center of the window is in the foreground, then the center pixel is the tumor edge, and the segmentation probability value of the center point is recorded. The edge probability value is extracted using a first-order statistical method.

[0020] 2) Extract global fiber features from DTI: Extract FA, AD, RD and MD fiber description features from the b0 image registered with DTI, and use deterministic fiber tracking to obtain the whole brain fiber direction;

[0021] 3) Extract DTI features of the edge region: Traverse all tumor edges and areas outside the edges in a 3*3*3 window (the foreground probability value of the tumor region is in the range of 0.1~0.5), record the average FA, AD, RD and MD values ​​of each window, and record the number of fibers passing through each window as fiber density; use the same first-order statistical method to extract features locally; in addition, record the average fiber density distribution of each window in the range of 0.5~1 at intervals of 0.1 as features.

[0022] Furthermore, the first-order statistical method is variance, mean, and kurtosis.

[0023] Furthermore, the specific steps for constructing the fifth step of the infiltration degree prediction model are as follows:

[0024] 1) Dataset partitioning: The dataset is partitioned in an 8:2 ratio to serve as the training set and the test set, with the positive and negative samples remaining consistent in each experiment;

[0025] 2) Use z_score to normalize the training data;

[0026] 3) Select the RBF kernel as the kernel function, and then use cross-validation and grid tuning to obtain the optimal values ​​of parameters C and g, where C iterates between 0.001 and 1000, and g iterates between 0.05 and 10.

[0027] 4) Apply the optimal parameters C and g to train the model and obtain the optimal prediction model;

[0028] 5) Use z_score to normalize the test set data, and then apply the test set to validate the above best model.

[0029] The present invention also provides a method for predicting the degree of brain tumor invasion, characterized in that the method includes: constructing an optimal prediction model using the method described in any one of the above-mentioned brain tumor invasion degree prediction models to predict the degree of brain tumor invasion in new samples.

[0030] The present invention also provides a device for predicting the extent of brain tumor invasion, characterized in that the device comprises:

[0031] Data collection module: Data collection: Collects image data of patients with glioma, including T1, T2, FL, T1CE, and DTI sequences of MRI images;

[0032] Image preprocessing module: Standardizes MRI images using a standardized process and aligns multimodal images;

[0033] Tumor segmentation module: The pre-trained NN-Unet brain tumor segmentation model is used to segment the tumor region based on the four input sequences T1, T2, FL, and T1CE, and outputs the tumor region foreground probability value of each pixel; a segmentation mask is extracted using a threshold of 0.5.

[0034] Feature extraction module: uses the probability distribution of segmentation boundaries, tumor boundaries, and tissue fiber conditions near the edges as features;

[0035] Infiltration degree prediction model construction module: Based on the above extracted features, with high or low infiltration as labels, high infiltration as positive samples and low infiltration as negative samples, an SVM classifier is used to construct a brain tumor infiltration degree prediction model.

[0036] Results output module: Outputs the results of the prediction of the new sample after executing the above best prediction model.

[0037] The present invention also provides a model, characterized in that the model is constructed by performing the brain tumor infiltration degree prediction method described above.

[0038] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, characterized in that the processor executes the program to implement the steps in the brain tumor infiltration degree prediction method described above.

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, performs the steps in the brain tumor infiltration degree prediction method described above.

[0040] The beneficial effects of the present invention compared with the prior art.

[0041] This invention innovatively uses tumor segmentation probability and fibrous density characteristics of the tumor margin region as features for invasiveness analysis to predict the degree of brain tumor invasion. Experiments show that there is a relationship between high and low invasiveness of brain tumors predicted by this model and survival (P<0.05), and the survival time of high invasiveness is significantly lower than that of low invasiveness. Attached Figure Description

[0042] Figure 1 There is a relationship between high and low invasiveness of brain tumors and survival. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings. The following embodiments are merely illustrative of the present invention and should not be construed as limiting the invention.

[0044] Example 1: This invention proposes a method for predicting the degree of brain tumor infiltration. The specific method steps and implementation methods are as follows.

[0045] [Data Collection] Data on the impact of glioma patients were collected, including T1, T2, FL, T1CE, and DTI sequences of MRI images.

[0046] [Image Preprocessing] A standardized workflow was used to normalize MRI images and align multimodal images. The specific workflow is as follows.

[0047] 1) For T1CE sequences, the scanning direction is automatically read, corrected to LPS, and N4 field correction and head correction are performed. Here, head correction is mainly based on the rigid registration from T1CE to the standard template (SRI-24 atlas) achieved by ants.

[0048] 2) Perform scan direction correction and N4 field correction on the T1, T2, and FL sequences, and then use rigid registration to register the T1, T2, and FL sequences onto T1CE.

[0049] 3) Use a pre-trained Unet model to debone the T1, T2, FL, and T1CE sequences.

[0050] 4) For DTI sequences, first extract the b0 image, then use FSL software to debone the DTI data, and finally rigidly register the b0 of the deboned DTI onto T1CE.

[0051] [Tumor Segmentation] A pre-trained NN-Unet brain tumor segmentation model is used to segment the tumor region based on the four input sequences T1, T2, FL, and T1CE, and outputs the foreground (tumor region) probability value for each pixel. A segmentation mask is typically extracted using a threshold of 0.5.

[0052] [Feature Extraction] Tumor invasiveness refers to the phenomenon of diseased tissue spreading to surrounding tissues. Typically, the segmentation boundaries are blurred and spiky, causing some damage to surrounding tissues (such as fibrosis). Based on these characteristics, this invention innovatively uses the probability distribution of segmentation boundaries, the tumor boundary, and the fibrous condition of tissues near the edge as features for tumor invasiveness analysis. The specific steps are as follows.

[0053] 1) Based on the edge characteristics of the tumor mask: The mask is segmented by traversing a 3*3*3 window. If there is a background value in the window and the center of the window is in the foreground, then the center pixel is the tumor edge, and the segmentation probability value of the center point is recorded. First-order statistical methods are used to extract features (such as variance, mean, and kurtosis) from the edge probability values.

[0054] 2) Extract global fiber features from DTI: Extract FA, AD, RD and MD fiber description features from the b0 image registered with DTI, and use deterministic fiber tracking to obtain the direction of whole brain fibers.

[0055] 3) Extracting DTI features from the edge region: Using a 3x3x3 window, traverse all tumor edges and the regions outside the edges (foreground (tumor region) probability values ​​are in the range of 0.1~0.5). Record the average FA, AD, RD, and MD values ​​for each window, and simultaneously record the number of fibers passing through each window as fiber density. Similarly, first-order statistical methods are used to extract local features (such as variance, mean, and kurtosis); additionally, the average fiber density distribution of each window within the 0.5~1 interval (with 0.1 as an interval) is recorded as a feature.

[0056] [Construction of Invasiveness Discrimination Model] Based on the extracted features above, and using the level of invasiveness as labeled by doctors as the tag, high invasiveness is the positive sample and low invasiveness is the negative sample. An SVM classifier is used to construct a tumor invasiveness prediction model. The training process is as follows:

[0057] 1) Dataset partitioning: A total of 224 samples were collected and partitioned into training and testing sets in an 8:2 ratio. The positive and negative samples were kept consistent within each experimental group.

[0058] 2) Use z_score to normalize the training data and apply the normalization method to the test set;

[0059] 3) Select the RBF kernel as the kernel function, and then use cross-validation and grid tuning to obtain the optimal values ​​of parameters C and g (where C iterates between 0.001 and 1000, and g iterates between 0.05 and 10).

[0060] 4) Applying the optimal parameters C and g, train the SVM model on the training set, and then validate the model performance on 44 test cases. The classification metric AUC reaches 85%, and the survival curves for the classification data are shown below. Figure 1 As shown.

[0061] Example 2.

[0062] A computer-aided diagnosis and treatment system based on a brain tumor infiltration degree prediction model, wherein the system performs the following steps.

[0063] Brain tumor MRI images (T1, T2, T1CE, FL, DTI) are input into a trained infiltration degree prediction model, and the model outputs a judgment result, which is either high infiltration or low infiltration.

[0064] The brain tumor infiltration degree prediction model includes an image-based preprocessing module, a segmentation module, a feature extraction module, and a brain tumor classification prediction module.

[0065] Brain tumor MRI images are input into the preprocessing module to obtain preprocessed standardized images. The preprocessing results are then input into the segmentation module to obtain segmented images. The segmentation results and the preprocessed images are input into the feature extraction module to obtain data features. The features are then input into the brain tumor classification and prediction module to output brain tumor invasiveness judgment results.

[0066] The preprocessing module includes orientation correction, deviation correction, and bone removal; the nnU-net network is used for tumor segmentation of the preprocessed T1, T2, T1CE, and FL images; the image-based tumor invasiveness classification model uses the SVM model; and the core innovation of this invention is the feature extraction module, which uses the probability distribution of the segmentation boundary, the tumor boundary, and the tissue fiber condition near the edge as features.

[0067] As a preferred approach, multimodal brain MRI images are input into the preprocessing model to obtain standardized image results, specifically including the following steps.

[0068] 1) For T1CE sequences, the scanning direction is automatically read, corrected to LPS, and N4 field correction and head correction are performed. Here, head correction is mainly based on the rigid registration from T1CE to the standard template (SRI-24 atlas) achieved by ants.

[0069] 2) Perform scan direction correction and N4 field correction on the T1, T2, and FL sequences, and then use rigid registration to register the T1, T2, and FL sequences onto T1CE.

[0070] 3) Use a pre-trained Unet model to debone the T1, T2, FL, and T1CE sequences.

[0071] 4) For DTI sequences, first extract the b0 image, then use FSL software to process and debone the DTI data, and finally rigidly register the deboned DTI b0 onto T1CE.

[0072] As a preferred embodiment, the T1, T2, FL, and T1CE images are input into the nnU-net network to obtain brain tumor segmentation results, specifically including the following steps.

[0073] The brain tumor image was subjected to channel augmentation using a convolutional module with a kernel size of 1×1×1 and a kernel count of 64, with ReLU activation. After channel augmentation, downsampling was performed using a convolution with a stride of 2 and a kernel size of 3×3×3, with LeakyReLU activation, reducing the feature map by a factor of 1. The resulting feature map was then subjected to another convolution operation with a kernel size of 3×3×3, a stride of 1, and a kernel count of 96, using LeakyReLU activation. This process was repeated, shrinking the feature map five times. Each shrinkage involved convolution with a stride of 2 and a stride of 1 to achieve downsampling and feature extraction. The number of convolutional kernels was set to 128, 192, 256, 384, and 512, respectively. The resulting feature maps were expanded 6 times, with each expansion performing a deconvolution operation with a stride of 2 to achieve upsampling. Skip connections were used to concatenate the feature maps of the same layer's contraction path with those of the expansion path. Then, two convolution operations were applied with a kernel size of 3×3 and the number of kernels were set to 384, 256, 192, 128, 96, and 64, respectively. The LeakyReLU activation function was used to expand the feature maps. The feature maps obtained in the above steps were then applied to a single convolution operation with a kernel size of 1×1×1. The 64-channel feature maps were mapped to the corresponding pixel classifications to obtain the 3D tumor segmentation results.

[0074] As a preferred embodiment, the tumor segmentation results and standardized image data are input into the feature extraction module to obtain the probability value distribution of the segmentation boundary, the tumor boundary and the tissue fiber features near the edge, specifically including the following steps.

[0075] The mask is segmented using a 3x3x3 window traversal. If the window contains background values ​​and the window center is in the foreground, then the center pixel is considered the tumor edge. The segmentation probability value of the center point is recorded, forming a 1D probability distribution matrix. Then, the variance, mean, kurtosis, skewness, median, interquartile range, entropy, maximum, and minimum values ​​of the matrix are calculated. There are a total of 9 feature values.

[0076] Then, FSL software was used to extract FA, AD, RD, and MD fiber description features from the preprocessed DTI and to perform deterministic fiber tracking.

[0077] Finally, using a 3x3x3 window, all tumor margins and areas outside the margins (tumor region probability values ​​in the range of 0.1 to 0.5) were traversed. The average FA, AD, RD, and MD values ​​were recorded for each window, along with the number of fibers passing through each window as fiber density. Similarly, variance, mean, kurtosis, skewness, median, interquartile range, entropy, maximum, and minimum features were extracted, for a total of 5x9 features. Additionally, the average fiber density distribution of each window within the 0.5 to 1 range (with 0.1 intervals) was recorded as a feature, for a total of 5 features.

[0078] As a preferred approach, the features are input into the brain tumor classification and prediction module to obtain classification results based on the degree of invasiveness, specifically including the following steps.

[0079] The above features are normalized using the z-score normalization method in the training set. The data is then fed into the trained SVM classifier. An output of 1 indicates high infiltration, and an output of 0 indicates low infiltration.

Claims

1. A method for constructing a model to predict the degree of brain tumor infiltration, characterized in that, The specific steps of the method are as follows: The first step is data collection: collecting image data of patients with gliomas, including T1, T2, FL, T1CE, and DTI sequences of MRI images; The second step is image preprocessing: standardized procedures are used to process MRI images and align multimodal images. The third step is tumor segmentation: a pre-trained NN-Unet brain tumor segmentation model is used to segment the tumor region based on the four input sequences T1, T2, FL, and T1CE, and the tumor region foreground probability value of each pixel is output; a segmentation mask is extracted using a threshold of 0.

5. The fourth step is feature extraction: the probability distribution of the segmentation boundary, the tumor boundary, and the tissue fiber condition near the edge are used as features; The specific steps for feature extraction are as follows: (1) Based on the edge characteristics of the tumor mask: The mask is segmented by traversing a 3*3*3 window. If there is a background value in the window and the center of the window is in the foreground, then the center pixel is the tumor edge, and the segmentation probability value of the center point is recorded. The edge probability value is extracted using a first-order statistical method. (2) Extract global fiber features of DTI: extract FA, AD, RD and MD fiber description features from the b0 image registered with DTI, and use deterministic fiber tracking to obtain the direction of whole brain fibers; (3) Extracting DTI features of the edge region: Using a 3*3*3 window, all tumor edges and the region outside the edge are traversed. The foreground probability value of the tumor region is in the range of 0.1 to 0.

5. The average FA, average AD, average RD and average MD values ​​of each window are recorded. At the same time, the number of fibers passing through each window is recorded as fiber density. The first-order statistical method is used to extract features locally. In addition, the average fiber density distribution of each window in the range of 0.5 to 1 is recorded as a feature with an interval of 0.

1. Step 5: Construction of the infiltration degree prediction model: Based on the features extracted above, the model is constructed using the degree of infiltration as a label, high infiltration as a positive sample and low infiltration as a negative sample, and an SVM classifier is used to construct the brain tumor infiltration degree prediction model.

2. The construction method according to claim 1, characterized in that, The specific steps of the second step of image preprocessing are as follows: (1) For the T1CE sequence, the scanning direction is automatically read, the scanning direction is corrected to LPS, and N4 field correction and head correction are performed; here, the head correction is based on ants to achieve rigid registration from T1CE to the standard template SRI-24 atlas; (2) Then, the scanning direction correction and N4 field correction are performed on the T1, T2, and FL sequences, and then rigid registration is used to register the T1, T2, and FL sequences onto T1CE; (3) Use the pre-trained Unet model to debone the T1, T2, FL, and T1CE sequences; (4) For the DTI sequence, first extract the b0 image, then use FSL software to debone the DTI data, and finally rigidly register the b0 of the debonded DTI onto T1CE.

3. The construction method according to claim 1, characterized in that, The first-order statistical methods are variance, mean, and kurtosis.

4. The construction method according to claim 1, characterized in that, The specific steps for constructing the fifth step of the infiltration degree prediction model are as follows: (1) Dataset partitioning: The dataset is divided into training and test sets in an 8:2 ratio, with positive and negative samples remaining consistent in each experiment. (2) Use z_score to normalize the training data; (3) Select the RBF kernel as the kernel function, and then use cross-validation and grid tuning to obtain the optimal values ​​of parameters C and g, where C iterates between 0.001 and 1000, and g iterates between 0.05 and 10; (4) Apply the optimal parameters C and g to train the model and obtain the optimal prediction model; (5) Use z_score to normalize the test set data, and then apply the test set to verify the above best prediction model.

5. A method for predicting the degree of brain tumor infiltration, characterized in that, The method includes: constructing an optimal prediction model using the method described in claim 4 for predicting the degree of brain tumor invasion, and then using this model to predict the degree of brain tumor invasion in new samples.

6. A device for predicting the degree of brain tumor infiltration, characterized in that, The device includes: Data collection module: Data collection: Collects image data of patients with glioma, including T1, T2, FL, T1CE, and DTI sequences of MRI images; Image preprocessing module: Standardizes MRI images using a standardized process and aligns multimodal images; Tumor segmentation module: The pre-trained NN-Unet brain tumor segmentation model is used to segment the tumor region based on the four input sequences T1, T2, FL, and T1CE, and outputs the tumor region foreground probability value of each pixel; a segmentation mask is extracted using a threshold of 0.

5. Feature extraction module: uses the probability distribution of segmentation boundaries, tumor boundaries, and tissue fiber conditions near the edges as features; The specific steps for feature extraction are as follows: (1) Based on the edge characteristics of the tumor mask: The mask is segmented by traversing a 3*3*3 window. If there is a background value in the window and the center of the window is in the foreground, then the center pixel is the tumor edge, and the segmentation probability value of the center point is recorded. The edge probability value is extracted using a first-order statistical method. (2) Extract global fiber features of DTI: extract FA, AD, RD and MD fiber description features from the b0 image registered with DTI, and use deterministic fiber tracking to obtain the direction of whole brain fibers; (3) Extracting DTI features of the edge region: Using a 3*3*3 window, all tumor edges and the region outside the edge are traversed. The foreground probability value of the tumor region is in the range of 0.1 to 0.

5. The average FA, average AD, average RD and average MD values ​​of each window are recorded. At the same time, the number of fibers passing through each window is recorded as fiber density. The first-order statistical method is used to extract features locally. In addition, the average fiber density distribution of each window in the range of 0.5 to 1 is recorded as a feature with an interval of 0.

1. Infiltration degree prediction model construction module: Based on the above extracted features, with high or low infiltration as labels, high infiltration as positive samples and low infiltration as negative samples, an SVM classifier is used to construct a brain tumor infiltration degree prediction model. The specific steps for building a prediction model are as follows: (1) Dataset partitioning: The dataset is divided into training and test sets in an 8:2 ratio, with positive and negative samples remaining consistent in each experiment. (2) Use z_score to normalize the training data; (3) Select the RBF kernel as the kernel function, and then use cross-validation and grid tuning to obtain the optimal values ​​of parameters C and g, where C iterates between 0.001 and 1000, and g iterates between 0.05 and 10; (4) Apply the optimal parameters C and g to train the model and obtain the optimal prediction model; Results output module: Outputs the results of the prediction of the new sample after executing the above best prediction model.

7. A model for predicting the degree of brain tumor infiltration, characterized in that, The model is constructed by performing the brain tumor infiltration degree prediction method described in claim 5.

8. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the brain tumor infiltration degree prediction method of claim 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the steps of the brain tumor infiltration degree prediction method as described in claim 5.

Images