A Method and System for Three-Dimensional Reconstruction of Brain MRI Tumor Regions Based on Convolutional Neural Networks

Abstract

This invention provides a method and system for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks, belonging to the field of three-dimensional reconstruction technology. The method includes: preprocessing multi-sequence MRI raw voxel data of the patient's brain to obtain standard feature tensors; inputting the standard feature tensors into a three-dimensional depth-separable convolutional encoder to output feature maps; and performing directional filtering on the feature maps to obtain a set of feature maps. This invention improves the overall accuracy, computational efficiency, and morphological realism of three-dimensional reconstruction of brain MRI tumor regions by standardizing data preprocessing, lightweight feature extraction, performing three-level feature calibration by region, and correcting the probability distribution with anatomical prior constraints, ultimately optimizing the reconstruction mesh.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction technology, and in particular to a method and system for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks. Background Technology

[0002] When performing 3D reconstruction of tumor regions from brain MRI data using convolutional neural networks, existing methods typically employ a 3D convolutional encoder to extract features from multiple image sequences, and then generate isosurface meshes using probabilistic voxel maps. However, in practical data processing, these methods still present some noteworthy technical issues in certain situations. A common problem relates to the left-right symmetry of brain MRI data. Because brain anatomy naturally exhibits mirror symmetry in images, and certain types of tumors are often unilaterally distributed, existing segmentation models sometimes generate false positive activations at symmetrical locations on the healthy side when certain normal tissues (such as choroid plexuses, vascular structures, or calcifications) show similarities in grayscale distribution and texture features to the target tumor region. These false positive regions and the actual tumor region may each form a connected volume exceeding a certain preset threshold, and the spatial symmetry distance between them may be less than or equal to the usual tolerance range. In this situation, when extracting isosurface meshes from the probabilistic voxel map, the algorithm often struggles to automatically identify and remove false positive regions on the healthy side based on the symmetry distance and volume relationship of the morphological primitives. This results in the final reconstructed 3D surface model containing two symmetrically distributed connected volumes with similar volumes, thus generating unnecessary model redundancy.

[0003] For example, in data reconstruction, in a set of brain MRI voxel data, the actual tumor region is located in the left temporal lobe, showing a high signal intensity on T2-weighted images. A normal choroidal fissure cyst exists in the corresponding location in the right temporal lobe, and its signal intensity on T2-weighted images is similar to that of the left tumor, with an approximately spherical shape. Existing segmentation methods using a three-dimensional U-shaped network (3DU-Net) output connected regions with volumes exceeding 0.5 cm³ on both sides when processing this data. The centroid distance between the two on the axial section is approximately 2.5 mm, within the common symmetry tolerance range. Because existing methods do not incorporate a post-processing mechanism based on symmetry constraints, the subsequent isosurface extraction step incorporates both connected regions into the three-dimensional surface model. The resulting reconstructed model simultaneously presents two similarly shaped, spatially symmetrical surface structures, failing to automatically suppress false positives from the healthy side at the algorithmic level. Summary of the Invention

[0004] This invention provides a method and system for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks, which can realistically restore the spatial morphology, size, boundaries and positional relationships of tumors with complete details and smooth morphology.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks, the method comprising:

[0007] Step 1: Preprocess the raw voxel data from the multi-sequence MRI of the patient's brain to obtain a standard feature tensor; input the standard feature tensor into a 3D depth-separable convolutional encoder to output a feature map, and perform directional filtering on the feature map to obtain a set of feature maps;

[0008] Step 2: Extract the first morphological primitive from the first local region from the feature map set. When the anisotropy of the first morphological primitive is lower than the set value and continues to decay along the depth direction, perform the first recalibration on the feature map set to obtain the first corrected feature map.

[0009] Step 3: Fuse the first corrected feature map with the standard feature tensor to construct a multi-scale feature pyramid. Extract the second morphological primitives in the second local region from the multi-scale feature pyramid. When the feature difference gradient between the second morphological primitives and the adjacent regions is less than the set tolerance, perform a second recalibration on the second local region to obtain the second corrected feature map.

[0010] Step 4: Convert the second correction feature map into a tumor probability voxel map, extract the third morphological primitives in the third local region from the tumor probability voxel map, and when there are two connected bodies in the third morphological primitives whose symmetric distance is within the tolerance and whose volume exceeds the minimum volume, perform a third calibration on the tumor probability voxel map to obtain the corrected probability voxel map.

[0011] Step 5: Extract isosurface mesh from the modified probability voxel map based on the set reference, and smooth and fill the holes in the isosurface mesh to obtain a three-dimensional surface model of the brain MRI tumor region.

[0012] Secondly, a three-dimensional reconstruction system for brain MRI tumor regions based on convolutional neural networks includes:

[0013] The processing module is used to preprocess the raw voxel data of multi-sequence MRI of the patient's brain to obtain a standard feature tensor; the standard feature tensor is input into a three-dimensional depth-separable convolutional encoder to output a feature map, and the feature map is directionally filtered to obtain a set of feature maps;

[0014] The first calibration module is used to extract the first morphological primitives in the first local region from the feature map set. When the anisotropy of the first morphological primitives is lower than a set value and continues to decay along the depth direction, the first calibration is performed on the feature map set to obtain the first corrected feature map.

[0015] The second calibration module is used to fuse the first calibration feature map with the standard feature tensor to construct a multi-scale feature pyramid. It extracts the second morphological primitives in the second local region from the multi-scale feature pyramid. When the feature difference gradient between the second morphological primitives and the adjacent regions is less than the set tolerance, the second calibration is performed on the second local region to obtain the second calibration feature map.

[0016] The third calibration module is used to convert the second correction feature map into a tumor probability voxel map, extract the third morphological primitives in the third local region from the tumor probability voxel map, and when there are two connected bodies in the third morphological primitives whose symmetric distance is within the tolerance and whose volume exceeds the minimum volume, the third calibration is performed on the tumor probability voxel map to obtain the corrected probability voxel map.

[0017] The extraction module is used to extract isosurface meshes from the modified probability voxel map based on a set benchmark, and to smooth and fill holes in the isosurface meshes to obtain a three-dimensional surface model of the brain MRI tumor region.

[0018] Thirdly, a computing device includes:

[0019] One or more processors;

[0020] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0021] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0022] The above-described solution of the present invention has at least the following beneficial effects:

[0023] This method employs a targeted verification and third-level calibration mechanism using third-morphological primitives to detect two connected entities within a third local region whose symmetric distance is within tolerance and whose volume exceeds the minimum volume. Regularization adjustments are performed based on the anatomical symmetry characteristics of the brain, automatically correcting the probability values ​​of symmetrical voxel pairs. This effectively eliminates false-positive regions on the healthy side, ensuring that the reconstructed 3D surface model corresponds only to the actual tumor region, avoiding redundant structural interference, and improving the accuracy and reliability of the reconstruction results. A three-stage progressive feature calibration process is designed, with conditional trigger calibration performed in three key stages: the initial stage of feature extraction, after multi-scale feature fusion, and after probabilistic voxel map generation. This gradually corrects deviations in the feature extraction process and optimizes the accuracy of feature representation. Simultaneously, a 3D depth-separable convolutional encoder combined with directional filtering can capture 3D features from multi-sequence MRI data.

[0024] The construction of a multi-scale feature pyramid enables comprehensive extraction of tumor features at different scales, further improving the accuracy of tumor region segmentation. A dedicated preprocessing workflow is designed for the raw voxel data from multiple sequences of patient brain MRI, adaptable to various clinically commonly used brain MRI sequences, eliminating the need for complex manual intervention on the raw data. Simultaneously, multi-level calibration mechanisms are based on the characteristics of brain anatomy to locate local regions, effectively addressing complex situations such as individual differences among patients, MRI data noise, and image artifacts, reducing reconstruction errors caused by data heterogeneity. Through a coherent workflow of isosurface mesh extraction, smoothing, and hole filling, problems such as mesh coarseness, holes, and morphological distortion that occur during the conversion of probabilistic voxel maps into mesh models are effectively solved. The generated 3D curved surface model can realistically restore the spatial morphology, size, boundaries, and positional relationships of the tumor, with complete details and smooth morphology. A 3D depth-separable convolutional encoder is employed, which, compared to traditional 3D convolutional structures, reduces the number of model parameters and computational load while ensuring feature extraction accuracy, improving the efficiency of feature extraction and overall reconstruction. Attached Figure Description

[0025] Figure 1 This is a schematic flowchart of a three-dimensional reconstruction method for brain MRI tumor regions based on convolutional neural networks provided in an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of a three-dimensional reconstruction system for brain MRI tumor regions based on convolutional neural networks, provided in an embodiment of the present invention. Detailed Implementation

[0027] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0028] like Figure 1 As shown, embodiments of the present invention propose a method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks, the method comprising the following steps:

[0029] Step 1: Preprocess the raw voxel data from the multi-sequence MRI of the patient's brain to obtain a standard feature tensor; input the standard feature tensor into a 3D depth-separable convolutional encoder to output a feature map, and perform directional filtering on the feature map to obtain a set of feature maps;

[0030] Step 2: Extract the first morphological primitive from the first local region from the feature map set. When the anisotropy of the first morphological primitive is lower than the set value and continues to decay along the depth direction, perform the first recalibration on the feature map set to obtain the first corrected feature map.

[0031] Step 3: Fuse the first corrected feature map with the standard feature tensor to construct a multi-scale feature pyramid. Extract the second morphological primitives in the second local region from the multi-scale feature pyramid. When the feature difference gradient between the second morphological primitives and the adjacent regions is less than the set tolerance, perform a second recalibration on the second local region to obtain the second corrected feature map.

[0032] Step 4: Convert the second correction feature map into a tumor probability voxel map, extract the third morphological primitives in the third local region from the tumor probability voxel map, and when there are two connected bodies in the third morphological primitives whose symmetric distance is within the tolerance and whose volume exceeds the minimum volume, perform a third calibration on the tumor probability voxel map to obtain the corrected probability voxel map.

[0033] Step 5: Extract isosurface mesh from the modified probability voxel map based on the set reference, and smooth and fill the holes in the isosurface mesh to obtain a three-dimensional surface model of the brain MRI tumor region.

[0034] In this embodiment of the invention, a targeted verification and third-level calibration mechanism using third morphological primitives detects two connected bodies within the third local region whose symmetric distance is within tolerance and whose volume exceeds the minimum volume. Regularization adjustments are performed based on the anatomical symmetry characteristics of the brain, automatically correcting the probability values ​​of symmetrical voxel pairs. This effectively eliminates false-positive regions on the healthy side, ensuring that the reconstructed 3D surface model corresponds only to the real tumor region, avoiding redundant structural interference, and improving the accuracy and reliability of the reconstruction results. A three-stage progressive feature calibration process is designed, with conditional trigger calibration performed in three key stages: the initial stage of feature extraction, after multi-scale feature fusion, and after probabilistic voxel map generation. This gradually corrects deviations in the feature extraction process and optimizes the accuracy of feature representation. Simultaneously, a 3D depth-separable convolutional encoder combined with directional filtering can capture the 3D features of multi-sequence MRI data; the construction of a multi-scale feature pyramid enables comprehensive extraction of tumor features at different scales, further improving the accuracy of tumor region segmentation. A dedicated preprocessing workflow is designed for the patient's brain multi-sequence MRI raw voxel data, adaptable to various clinically commonly used brain MRI sequences, eliminating the need for complex manual intervention on the raw data. Meanwhile, the multi-level calibration mechanism locates local regions based on the characteristics of brain anatomy, effectively addressing complex situations such as individual differences among patients, MRI data noise, and image artifacts, reducing reconstruction errors caused by data heterogeneity. Through a coherent process of isosurface mesh extraction, smoothing, and hole filling, it effectively solves problems such as mesh coarseness, holes, and morphological distortion that occur during the conversion of probabilistic voxel maps into mesh models. The generated 3D curved surface model can realistically restore the spatial morphology, size, boundaries, and positional relationships of the tumor, with complete details and smooth morphology. Employing a 3D depth-separable convolutional encoder, compared to traditional 3D convolutional structures, it reduces the number of model parameters and computational load while ensuring feature extraction accuracy, thus improving the efficiency of feature extraction and overall reconstruction.

[0035] In a preferred embodiment of the present invention, step 1 includes:

[0036] Step 100: Preprocess the raw voxel data from the multi-sequence MRI of the patient's brain to obtain standard feature tensors. Specifically, this includes: collecting raw voxel data from the multi-sequence MRI of the patient's brain; selecting clinically diagnosed patients requiring brain MRI examination; and performing multi-sequence scans of the patient's brain using a 1.5T or 3.0T MRI scanner, covering the entire brain from the top to the base of the skull; the scan sequences must include at least T1-weighted images, T2-weighted images, and FLAIR sequences (liquid attenuation inversion recovery sequences). The T1-weighted image scan parameters are set as follows: repetition time (TR) 1500 to 2000 ms, echo time (TE) 1... The scan parameters for T2-weighted images were: repetition time (TR) 3000-4000 ms, echo time (TE) 80-100 ms, slice thickness 3-5 mm, slice spacing 0.5-1 mm, and matrix 256×256. For FLAIR sequences, the parameters were: repetition time (TR) 8000-10000 ms, echo time (TE) 80-100 ms, inversion time (TI) 2000-2500 ms, slice thickness 3-5 mm, slice spacing 0.5-1 mm, and matrix 256×256. After scanning, the raw scan data of the patient's brain was exported using the image storage and transmission system of the MRI equipment in DICOM format. Each sequence corresponds to a three-dimensional voxel matrix, and the value of each element in the matrix represents the MRI signal intensity at the corresponding voxel location. All raw voxel data in DICOM format were converted to NIfTI format. Simultaneously, metadata such as voxel spacing and imaging orientation were extracted for each sequence. The voxel spacing of all sequences was uniformly adjusted to 1×1×1, and the imaging orientation was uniformly corrected to a coordinate system consistent with standard axial, coronal, and sagittal planes to ensure accurate voxel position correspondence across different sequences. A Gaussian filtering algorithm was used to filter the standardized voxel matrix of each sequence, with the Gaussian filter kernel size set to 3×3×3 and the standard deviation of the Gaussian function set to 0.5. For each voxel, a 3×3×3 three-dimensional neighborhood of voxels was selected centered on that voxel. The sum of the products of the signal intensities of all voxels in the neighborhood and the corresponding weights of the Gaussian kernel function was calculated. This summation result was used as the denoised signal intensity of that voxel, as shown in the following formula: ;

[0037] in The original voxel signal intensity, The intensity of the voxel signal after denoising. A 3×3×3 Gaussian kernel function was used. Z-score normalization was employed to normalize the voxel signal intensity of each sequence, eliminating signal intensity differences caused by different patients and scanning devices. First, the mean signal intensity of all voxels in a single sequence was calculated. and standard deviation Then, for each voxel in the sequence, its signal intensity is... Convert to normalized value ,Right now This ensures that the mean signal intensity of all voxels is 0 and the standard deviation is 1 after normalization.

[0038] Based on standard brain anatomy atlases, the smallest three-dimensional cube region encompassing the entire brain region is extracted. Background voxels (voxels with signal intensity close to 0) outside this three-dimensional cube region are removed to obtain the brain region of interest voxel matrix for each sequence. All the brain region of interest voxel matrices processed in the above way are concatenated according to the channel dimension to construct a standard feature tensor. The dimensions of the standard feature tensor are H×W×D×C, where H, W, and D represent the height, width, and depth of the voxel matrix, respectively, and C represents the number of channels of the multiple sequences (e.g., C=3 for 3 sequences). Each channel corresponds to the normalized voxel data of one sequence, thus obtaining the standard feature tensor.

[0039] Step 101: The standard feature tensor is used as the input data of the 3D depth-separable convolutional encoder. The 3D depth-separable convolutional encoder includes multiple cascaded 3D depth-separable convolutional modules. Each module sequentially performs 3D depth-separable convolution, batch normalization, and linear rectified activation to obtain an initial feature map at the corresponding scale. Specifically, this includes:

[0040] A 3D depthwise separable convolutional encoder is constructed, which contains four cascaded 3D depthwise separable convolutional modules. The number of output channels of the four modules are 32, 64, 128, and 256, respectively. Each 3D depthwise separable convolutional module consists of a 3D depthwise separable convolutional layer, a batch normalization layer, and a linear rectified activation layer (ReLU) cascaded together. Each module is followed by a 3D max pooling layer with a stride of 2 (pooling kernel size 3×3×3) to reduce the feature map dimension and retain key features. The standard feature tensor (H×W×D×C) obtained in step 100 is used as the input data of the 3D depth-separable convolutional encoder, and is sequentially input into four cascaded 3D depth-separable convolutional modules. The specific execution process of each module is as follows: the input feature map is divided into single channels in the depth direction, and a 3×3×3 3D depth convolution kernel is used to perform convolution operation on each channel to obtain a single-channel feature map; then, a 1×1×1 3D point convolution kernel is used to fuse all single-channel feature maps to obtain the convolutional output feature map of the module. That is, for the input feature map X (dimension H×W×D×C), the output feature map is obtained. ),in Represents the height of the input feature map. Represents the width of the input feature map. Represents the depth of the input feature map. The number of channels in the input feature map is represented by the number of channels. First, an intermediate feature map Y is obtained through depthwise convolution. Each channel of Y... The c-th channel of X and its corresponding 3×3×3 depth convolution kernel The result is obtained through convolution, and the formula is as follows: (Convolution operation); then, point convolution is used to fuse all channels of Y to obtain the output feature map Z. ,in It has a 1×1×1 point convolution kernel with dimension . , This represents the number of output channels for this module.

[0041] Batch normalization is performed on the feature map Z output by the 3D depthwise separable convolution to eliminate the gradient vanishing problem. For each channel of the feature map Z, the mean value of all voxel values ​​in that channel is calculated. and variance Then, the voxel value z of that channel is normalized to obtain... ,in To prevent the use of tiny constants with a denominator of 0, a value of 1e-5 is chosen; then a scaling factor is applied. and offset factor Perform linear adjustment to obtain the batch-normalized feature map. , and These are trainable parameters, initialized to 1 and 0 respectively. The batch-normalized feature maps... The input is fed into the ReLU activation layer, where a nonlinear mapping is performed. For Each voxel value in Output This involves retaining all voxel values ​​greater than 0 and setting voxel values ​​less than or equal to 0 to 0, resulting in the final output feature map of the module. After the activation output of each module, a 3D max pooling operation with a stride of 2 and a kernel size of 3×3×3 is performed. Using a 3×3×3 pooling window, the feature map is traversed with a stride of 2, and the maximum value of all voxel values ​​within each pooling window is taken as the value of the corresponding output voxel, reducing the spatial dimension of the feature map (H, W, and D are all halved) while retaining key features. After sequential processing by four cascaded 3D depthwise separable convolutional modules, four feature maps at different scales are obtained, each corresponding to a scale of 3D brain MRI features. These four feature maps at different scales are collectively referred to as the initial feature maps.

[0042] Step 102: For each initial feature map, calculate the response intensity of each voxel position along the sagittal, coronal, and vertical axes, and perform per-directional spatial non-maximum suppression. That is, for each direction, only retain the voxel activation values ​​whose response intensity is greater than the response intensity of all voxels other than the corresponding voxel itself in the three-dimensional neighborhood of the corresponding direction, and set the activation values ​​of all voxels that do not meet the condition to zero, thus obtaining the sub-feature maps after filtering in three directions; concatenate the sub-feature maps after filtering in three directions corresponding to the same initial feature map according to the channel dimension to obtain the feature map set, specifically including:

[0043] The three screening directions are defined as follows: the sagittal axis is the axis parallel to the left-right symmetry plane of the brain; the coronal axis is the axis parallel to the front-back direction of the brain; and the vertical axis is the axis parallel to the top-bottom direction of the brain. These three directions are perpendicular to each other and cover all directions in the three-dimensional space of the brain. For each initial feature map obtained in step 101 (a total of 4 different scales), the following operations are performed along the three directions to obtain the sub-feature map after screening in each direction. Along the target directions (sagittal axis, coronal axis, and vertical axis), the response intensity at each voxel location is calculated. The calculation of the response intensity is based on the three-dimensional neighborhood features of the voxel in the target direction. Specifically, a 5×5×5 three-dimensional neighborhood (covering adjacent voxels in the target direction) is selected centered on the current voxel. The absolute values ​​of the feature values ​​(i.e., the voxel values ​​of the initial feature map) of all voxels within the neighborhood are calculated, and this summation is taken as the response intensity of the current voxel in that direction. The formula is ,in voxels in the neighborhood of the initial feature map The eigenvalues ​​and neighborhood range are adjusted according to the target direction to ensure that the neighborhood extends along the target direction.

[0044] For each target direction, spatial nonmaximum suppression is performed on the calculated response intensity map in that direction. The aim is to retain the voxel activation values ​​with the strongest response intensity in each direction and suppress redundant weak response voxels. For each voxel... A 3×3×3 three-dimensional neighborhood is selected in the target direction (containing only the voxels adjacent to the voxel in the target direction, excluding the voxel itself), and the response intensity of the current voxel is compared. The activation value of a voxel is compared with the response intensity of all other voxels in its neighborhood. If the response intensity of the current voxel is greater than that of all other voxels in its neighborhood, the activation value of that voxel (i.e., the eigenvalue of that voxel in the initial feature map) is retained. If the response intensity of the current voxel is less than or equal to the response intensity of any other voxel in its neighborhood, the activation value of that voxel is set to zero. Through the above operations, each initial feature map can obtain sub-feature maps filtered in three directions (corresponding to the sagittal, coronal, and vertical axes, respectively). The dimension of each sub-feature map is consistent with the dimension of the corresponding initial feature map, with only the voxel activation values ​​being filtered and optimized. The sub-feature maps filtered in the three directions corresponding to the same initial feature map are concatenated according to the channel dimension. The concatenation process is as follows: keep the H, W, and D dimensions of the sub-feature maps unchanged, and superimpose the channels of the three sub-feature maps. For example, if the dimension of an initial feature map is H×W×D×C, the dimension of its three directional sub-feature maps is also H×W×D×C. The dimension of the concatenated feature map is H×W×D×(3C). The above concatenation operation is performed on the four initial feature maps of different scales to obtain four concatenated feature maps of different scales. These four concatenated feature maps of different scales are collectively referred to as the feature map set.

[0045] This embodiment effectively eliminates problems such as inconsistent voxel spacing, signal intensity differences, data noise, and background interference caused by different scanning devices and patients through format standardization, Gaussian denoising, intensity normalization, and region of interest extraction. It converts raw voxel data from multi-sequence MRI into standardized feature tensors, solving the feature extraction bias problem caused by non-standard input data in existing methods. The employed 3D depth-separable convolutional encoder, compared to traditional 3D convolutional encoders, reduces the number of model parameters and computational load while ensuring the capture of 3D spatial features of brain MRI, thus improving feature extraction efficiency through the separation of depth convolution and point convolution. Simultaneously, the concatenated batch normalization and ReLU activation layers in each convolutional module effectively avoid gradient vanishing, enhancing the model's feature representation capability and enabling the extraction of brain MRI features at different scales, thus solving the problems of low feature extraction efficiency and inaccurate feature representation in existing methods. Response intensity calculation and spatial nonmaximum suppression are performed along three key directions: sagittal axis, coronal axis, and vertical axis. Key features in each direction are selected and then a feature map set is formed by channel splicing. This enhances the directionality and specificity of the features, captures the feature differences of brain anatomy in different directions, and effectively suppresses redundant weak features.

[0046] In a preferred embodiment of the present invention, step 2 includes:

[0047] Step 200: From the feature map set, locate the voxel coordinate range of the intersection of the anterior horn of the lateral ventricle and the septum pellucidum based on the brain anatomy atlas. Take a subset of the feature maps within this range as the first local region. Specifically, this includes: retrieving a standard digital brain anatomy atlas. This atlas is a clinically used, three-dimensional standardized brain anatomy template containing precise spatial coordinate information of various brain structures (including the lateral ventricle and septum pellucidum), and maintaining consistency with the coordinate system of the data processed in steps 100-102 to ensure accurate coordinate localization. In the standard digital brain anatomy atlas, locate the intersection of the anterior horn of the lateral ventricle and the septum pellucidum. Using the atlas's built-in coordinate annotation function, extract the three-dimensional spatial voxel coordinate range of this intersection. Specifically, the coordinate range is defined by extending in all directions of three-dimensional space from the geometric center of the intersection until the entire area of ​​the intersection of the anterior horn of the lateral ventricle and the septum pellucidum is completely covered. Determine the minimum and maximum voxel coordinate values ​​of this region to form a complete voxel coordinate interval, i.e., the x-axis coordinate range [x...]. min ,x max ], y-axis coordinate range [y min ,y max ], z-axis coordinate range [z min ,z max The coordinate system is defined as follows: x-axis corresponds to the sagittal axis, y-axis corresponds to the coronal axis, and z-axis corresponds to the vertical axis, maintaining complete consistency with the coordinate system of the previous feature map set. In the feature map set obtained in step 102, spatial coordinates (x, y, z) that simultaneously satisfy x ∈ [x...] are selected for each voxel. min ,x max ]、y∈[y min ,y max ]、z∈[z min ,z max All voxels are analyzed, and their corresponding feature data are integrated to form a feature map subset. This feature map subset is the first local region, and its dimension is consistent with the dimension of the feature map at the corresponding scale in the feature map set. Only the feature data of the intersection region is retained, thus completing the localization and extraction of the first local region. The feature map set contains four feature maps at different scales (corresponding to the four scales of the stitched feature map in step 102), each with a different spatial resolution. The absolute voxel coordinate range [x] obtained from the brain anatomical atlas is used. min ,x max ]、[y min ,y max ]、[z min ,z max The resolution is based on the original whole-brain resolution. When extracting feature maps at each scale, coordinate scaling is required according to the downsampling factor for that scale. Let s be the downsampling factor for the feature map at the i-th scale. i(The original resolution corresponds to s=1, after one pooling s=2, after two pooling s=4, and after three pooling s=8), then the corresponding voxel coordinate range at this scale is [x min / s i ,x max / s i ]、[y min / s i ,y max / s i ]、[z min / s i ,z max / s i After rounding and filtering voxels, the first local region subset at that scale is obtained. The subsets at the four scales are collectively referred to as the feature map subsets of the first local region.

[0048] Step 201: For each voxel in the first local region, take the set of feature vectors in the three-dimensional neighborhood centered on the corresponding voxel position, calculate the gradient component of each feature vector, and perform a weighted summation of the outer product of each feature vector with itself to construct a three-dimensional structure tensor matrix. The three-dimensional structure tensor matrix is ​​a third-order symmetric positive semi-definite matrix. Perform eigenvalue decomposition on the three-dimensional structure tensor matrix to obtain three non-negative eigenvalues, which are denoted as the first eigenvalue, second eigenvalue, and third eigenvalue in descending order of value. Calculate the anisotropy score based on the first eigenvalue, second eigenvalue, and third eigenvalue. The anisotropy score is used to quantify the orientational consistency of the voxel structure in the first local region. The closer the value is to 1, the stronger the anisotropy. Specifically, this includes:

[0049] Traverse each voxel within the first local region, starting from the position of the currently traversed voxel. Using the geometric center as the reference point, a 3×3×3 three-dimensional neighborhood is selected. Feature vectors corresponding to all voxels within this neighborhood are collected to form a feature vector set. The dimension of each feature vector is consistent with the number of channels in the feature map set, i.e., each feature vector is a C-dimensional vector (C is the number of channels in multi-sequence MRI). The gradient components of each feature vector in this feature vector set are calculated, with the gradient component calculation performed separately for each channel of the feature vector. For any feature vector v in the feature vector set, its gradient components in the x-axis (sagittal axis), y-axis (coronal axis), and z-axis (vertical axis) are respectively... Finally, the gradient components of each feature vector in the three directions are obtained, forming the gradient vector. .

[0050] Calculate the outer product of each gradient vector with itself. The outer product of gradient vectors is a third-order matrix. (in ), its outer product The calculation results are as follows: ;

[0051] The outer product of all gradient vectors is summed with weights to construct a 3D structure tensor matrix. The weighting coefficients are Gaussian weighted, and the weight value is determined by the Euclidean distance between the voxel corresponding to the gradient vector and the central voxel; the closer the voxel is to the central voxel, the larger the weight value. The specific weight calculation is as follows: ,in Let be the Euclidean distance between the current voxel and the central voxel. The standard deviation is Gaussian, with a value of 1.0; 3D structure tensor matrix. The calculation formula is: ;

[0052] in The total number of voxels in a 3×3×3 three-dimensional neighborhood ( =27), For the first The weighting coefficients of individual elements, For the first The gradient vector of the individual element, the three-dimensional structure tensor matrix It is a third-order symmetric positive semi-definite matrix, which conforms to the spatial distribution characteristics of brain voxel structures.

[0053] For three-dimensional structure tensor matrix Eigenvalue decomposition is performed using the Jacobi iteration method. First, the orthogonal matrix U is initialized as the identity matrix, and then the matrix is ​​repeatedly searched... Find the off-diagonal element with the largest absolute value. For the row and column containing this off-diagonal element, construct the corresponding rotation matrix R. Then, use matrix multiplication U = U × R and S = ... Using the method of ×S×R, the largest off-diagonal element is eliminated. The iterative process of finding, constructing the rotation matrix, and updating the matrix is ​​repeated until the absolute value of all off-diagonal elements in matrix S is less than the preset convergence threshold (set to 1e-6). At this point, matrix S becomes a diagonal matrix, and the three elements on the diagonal are the three non-negative eigenvalues ​​obtained from the decomposition. Since the matrix is ​​a 3x3 symmetric positive semi-definite matrix, all three eigenvalues ​​are greater than or equal to 0. The three eigenvalues ​​are then sequentially designated as the first eigenvalue in descending order of their numerical values. Second eigenvalue Third eigenvalue ,Right now .

[0054] Anisotropy fractions are calculated based on three eigenvalues, and fractional anisotropy is used. The calculation formula is as follows: ;

[0055] The anisotropy fraction The value ranges from 0 to 1, and is used to quantify the orientational consistency of the voxel structure in the first local region. The closer the value is to 1, the stronger the orientation of the voxel structure and the more uniform the spatial orientation; the closer the value is to 0, the more homogeneous the voxel structure and the weaker the orientation, thus completing the quantitative representation of the first morphological primitive.

[0056] Step 202: Calculate the difference in anisotropy fractions of corresponding voxels on adjacent slices along the depth direction to form a difference sequence. When the arithmetic mean of the anisotropy fractions of all voxels in the first local region is lower than a preset homogeneity threshold, and the difference sequence is negative on multiple consecutive slices, trigger the first calibration. That is, perform three-dimensional global average pooling on all channels corresponding to the first local region in the feature map set to obtain a channel description vector, specifically including:

[0057] The depth direction is determined as the z-axis direction (consistent with the depth direction in steps 100 to 102). The first local region is divided into slices along the z-axis direction. Each slice is a two-dimensional plane (x, y) with a fixed z-axis coordinate. All slices are numbered sequentially as slice 1, slice 2, ..., slice M according to the z-axis coordinate from largest to smallest (M is the total number of slices in the first local region along the z-axis direction).

[0058] The anisotropy fractions of corresponding voxels on adjacent slices are calculated slice by slice along the depth direction to construct a difference sequence: For any two adjacent slices k and k+1 (k ranges from 1 to M-1), the anisotropy fraction of each voxel on slice k and the anisotropy fraction of the voxel at the corresponding coordinate (x, y) on slice k+1 are extracted, and the difference between the two is calculated; the differences of the corresponding voxels of all adjacent slices are integrated in slice order to form a difference sequence. The dimension of this difference sequence is consistent with the number of slices and the number of voxels in each slice of the first local region. It can intuitively reflect the increase and decrease of the anisotropy fraction along the depth direction (z-axis direction) through the difference of the anisotropy fractions of corresponding voxels of adjacent slices. A positive difference indicates that the anisotropy fraction of the voxel at that position is increasing along the depth direction, and a negative difference indicates that the anisotropy fraction of the voxel at that position is decreasing along the depth direction. Through the distribution of the differences of all corresponding voxels, the overall change trend of the anisotropy fraction along the depth direction of the entire first local region can be presented.

[0059] The arithmetic mean of the anisotropy scores of all voxels within the first local region is calculated. Simultaneously, a preset homogeneity threshold is set. This involves collecting a large amount of clinically normal brain multi-sequence MRI data and, following the method in step 201, calculating the anisotropy scores of all normal voxels at the junction of the anterior horn of the lateral ventricle and the septum pellucidum (located consistent with the first local region). The statistical distribution of these scores is then obtained. Based on this distribution, a threshold range of 0.8 to 1.2 times the mean anisotropy score of normal voxels is selected. Considering the accuracy requirements for determining homogeneous tissue in clinical practice, the threshold range is determined to be 0.15 to 0.25. Taking into account the median of normal voxel scores, a default value of 0.2 is set to ensure accurate differentiation of anisotropy differences between normal homogeneous and abnormal heterogeneous tissues. Furthermore, a threshold for the number of consecutive slices is set. This threshold is determined based on the total number of slices in the first local region, with a range of 3 to 5 and a default value of 4. This means that one of the judgment conditions is met when the difference sequence has negative values ​​on four consecutive slices.

[0060] The calculated arithmetic mean is compared with a preset homogeneity threshold, and the difference sequence is detected to determine whether there are multiple consecutive slices (reaching a set threshold) where the differences between their corresponding voxels are all negative. When the arithmetic mean is less than the homogeneity threshold and the difference sequence is negative across a set number of consecutive slices, the first calibration is triggered. The specific operation of the first calibration is as follows: For all channels corresponding to the first local region in the feature map set obtained in step 102, a three-dimensional global average pooling operation is uniformly performed. The window size of the three-dimensional global average pooling is consistent with the three-dimensional dimensions of the first local region, i.e., the window dimension is [H1, W1, D1], where H1, W1, and D1 are the height, width, and depth of the first local region, respectively. The arithmetic mean of the feature values ​​of all voxels in each channel is calculated to obtain the feature mean of that channel. The feature mean of all channels is arranged in channel order to form a one-dimensional vector, which is the channel description vector. The dimension of the channel description vector is consistent with the number of channels in the feature map set, i.e., the dimension is C (C is the number of channels in multi-sequence MRI).

[0061] Step 203: Based on the dimension of the channel description vector and the preset compression ratio, construct a first linear transformation matrix and a second linear transformation matrix; multiply the channel description vector by the first linear transformation matrix, then sequentially perform ReLU nonlinear mapping, restore the dimension using the second linear transformation matrix, and finally output a channel weight vector between 0 and 1 using a sigmoid saturation activation function; multiply the channel weight vector element-wise with each channel feature map corresponding to the first local region in the feature map set to obtain the first corrected feature map, specifically including:

[0062] The dimension of the channel description vector is determined to be C (where C is the number of channels in the feature map set). A preset compression ratio is set, with a value ranging from 4 to 8, and a default value of 4. The compression ratio controls the dimension of the first linear transformation matrix, thereby compressing the dimension of the channel description vector. Based on the dimension C of the channel description vector and the preset compression ratio r... a Construct a first linear transformation matrix and a second linear transformation matrix. The dimension of the first linear transformation matrix W1 is [C / r]. a [,C] is used to compress the channel description vector from C dimensions to C / r. a Dimension, where C / r a Take the integer (if C cannot be taken by r) a (Divisible by rounding down); the dimensions of the second linear transformation matrix W2 are [C, C / r] a The vector is used to restore the compressed vector to C dimensions, consistent with the initial dimension of the channel description vector; the elements of the two linear transformation matrices are trainable parameters, and the initial values ​​are generated using the Xavier initialization method to ensure that the mean of the matrix elements is 0 and the variance is 2 / (C+C / r). a This is to avoid initialization deviations affecting the calculation results.

[0063] Perform matrix multiplication on the channel description vector V (of dimension C×1) and the first linear transformation matrix W1 to obtain the compressed vector V1, where the dimension of V1 is (C / r). aThe channel description vector V1 is compressed by first performing a multiplication of 1 (C×1). The compressed vector V1 is then input into the ReLU nonlinear mapping function for nonlinear mapping. This mapping sets all elements less than or equal to 0 to 0, retaining elements greater than 0, eliminating invalid negative features and enhancing feature representation. The ReLU-mapped vector is then multiplied with the second linear transformation matrix W2 to restore the dimension, resulting in vector V2 with a dimension of C×1, consistent with the initial dimension of the channel description vector V. The dimension-restored vector V2 is then input into the sigmoid activation function for numerical mapping. This mapping maps all elements of vector V2 to the interval between 0 and 1, generating the channel weight vector W with a dimension of C×1. Each element corresponds to the weight value of a channel in the feature map set. The closer the weight value is to 1, the higher the proportion of effective features in the corresponding channel; the closer it is to 0, the higher the proportion of interference features in the corresponding channel. The channel weight vector W is multiplied element-wise with each channel feature map in the first local region of the feature map set. Specifically, for the c-th channel (c from 1 to C) in the first local region of the feature map set, the feature values ​​of all voxels in the channel are multiplied position-wise with the c-th element of the channel weight vector W to obtain the calibrated feature value of the channel. After performing the above element-wise multiplication operation on all channels, all calibrated channel feature data are integrated to obtain the first corrected feature map. The dimension of the first corrected feature map is consistent with the dimension of the first local region, and its spatial resolution is the same as that of the standard feature tensor (both are the original voxel resolution). Only the feature values ​​have been optimized by weight calibration.

[0064] This embodiment, relying on brain anatomical atlases to delineate the first local region, can pinpoint the intersection of fixed tissues in the brain that are prone to feature mixing and structural homogeneity abnormalities. It abandons the indiscriminate operation mode across the entire domain, narrows the scope of feature analysis, and reduces interference from irrelevant brain tissue feature data on the calculation results. Employing a three-dimensional structural tensor combined with eigenvalue decomposition to solve for anisotropy fractions, it can accurately characterize the spatial structural orientation and arrangement of local voxel clusters in three-dimensional space. It can distinguish the feature differences between normal homogeneous brain tissue and tissues with abnormal texture, and the calculation results closely match the actual distribution of the brain's soft tissue spatial structure. By utilizing a dual judgment mechanism combining depth-direction slice difference sequences and overall mean, it can identify abnormal distribution states where local region features gradually attenuate along the depth direction, capturing the problem of undesirable feature distributions in image data where shallow features are obvious and deep features gradually weaken, effectively avoiding judgment errors caused by single-parameter judgment. The first feature calibration is completed by using a channel attention mechanism. The overall feature information of the channel is extracted by global pooling, and then the weights of different feature channels are adaptively allocated by linear transformation and activation function. This can automatically weaken the invalid feature channels with strong interference and low effective information ratio, and strengthen the effective feature channels with high target correlation, thus correcting the feature offset problem generated in the early convolution screening process from the source of feature extraction.

[0065] In a preferred embodiment of the present invention, step 3 includes:

[0066] Step 300: The first corrected feature map is element-wise added to the sub-tensor corresponding to the first local region in the standard feature tensor, while the non-first local region portion of the standard feature tensor remains unchanged, resulting in a fused feature map. The fused feature map is then subjected to 3D max pooling with downsampling ratios of 2, 4, and 8 to obtain a first-scale feature copy, a second-scale feature copy, and a third-scale feature copy. The third-scale feature copy is upsampled to the resolution of the second-scale feature copy and added element-wise to the second-scale feature copy to obtain a first fusion result. The first fusion result is upsampled to the resolution of the first-scale feature copy and added element-wise to the first-scale feature copy to obtain a second fusion result. The second fusion result is upsampled to the original resolution of the fused feature map to obtain a multi-scale feature pyramid, specifically including:

[0067] The first corrected feature map obtained in step 203 and the standard feature tensor obtained in step 100 are called to ensure that their spatial dimensions (height H, width W, depth D) are completely consistent with the channel dimension (C) to avoid operational anomalies caused by dimension mismatch. Element-wise addition is performed on the two feature data. Specifically, for any voxel coordinate and any channel, the feature value of the first corrected feature map at that coordinate and channel is directly added to the feature value of the standard feature tensor at that coordinate and channel to obtain the fused feature value. The fused feature values ​​of all voxels and all channels are integrated to obtain the fused feature map. This fused feature map combines the completeness of the original global features with the accuracy of the first corrected feature map, making up for the information shortcomings of a single feature.

[0068] Three-dimensional max pooling operations with downsampling ratios of 2, 4, and 8 were performed on the fused feature maps respectively. The three sets of operation parameters were set uniformly: the pooling kernel size was 3×3×3, and the same filling method was used, that is, the filling value = 1. Specifically, it means that one layer of voxels is filled around the spatial boundary of the fused feature map, and the feature value of the filled voxels is set to 0 to ensure a smooth transition of the spatial dimension of the feature map after the pooling operation and to avoid loss of boundary voxel features; the pooling step size corresponds one-to-one with the downsampling ratio (ratio 2 corresponds to step size 2, ratio 4 corresponds to step size 4, and ratio 8 corresponds to step size 8). During computation, a 3×3×3 pooling window is used. The fused feature map is traversed according to the corresponding step size. The maximum value of all voxel feature values ​​in each window is taken as the output voxel feature value. Three sets of pooling are completed in sequence: a downsampling factor of 2 yields a first-scale feature copy (the spatial dimension is 1 / 2 of the fused feature map, corresponding to local detail features); a downsampling factor of 4 yields a second-scale feature copy (the spatial dimension is 1 / 4 of the fused feature map, corresponding to mesoscale structural features); and a downsampling factor of 8 yields a third-scale feature copy (the spatial dimension is 1 / 8 of the fused feature map, corresponding to global features).

[0069] A trilinear interpolation method is used to perform cross-scale upsampling and fusion, ensuring accurate matching of feature resolution and smooth transition of feature information after upsampling. The specific process of the trilinear interpolation method is as follows: For any newly added voxel coordinates after upsampling, firstly, find the corresponding 4×4×4 neighboring voxels in the original low-resolution feature copy, and record the coordinates and corresponding feature values ​​of each neighboring voxel; then, perform linear interpolation calculations on the feature values ​​of the neighboring voxels in the three spatial directions of x-axis, y-axis, and z-axis respectively. First, complete the interpolation in the x-axis direction, then complete the interpolation in the y-axis direction based on the x-axis interpolation result, and finally complete the interpolation in the z-axis direction based on the y-axis interpolation result, finally obtaining the feature values ​​of the newly added voxels, ensuring that the feature information is continuous and without obvious discontinuities after interpolation. First, the third-scale feature copy is upsampled to the resolution of the second-scale feature copy, and the newly added voxel feature values ​​are calculated according to the above trilinear interpolation method; then, the upsampled third-scale feature copy and the second-scale feature copy are added element by element (corresponding coordinates and channel feature values ​​are directly added) to obtain the first fusion result, realizing the complementary fusion of low-resolution global features and medium-resolution structural features. The first fusion result is then upsampled to the resolution of the first-scale feature copy using the same trilinear interpolation method, and added element-wise to the first-scale feature copy to obtain the second fusion result, thus strengthening the correlation between multi-scale features. Finally, the second fusion result is upsampled to the original resolution of the fused feature map using trilinear interpolation, restoring it to the dimensions of H×W×D×C, resulting in the multi-scale feature pyramid.

[0070] Step 301: Based on the white matter fiber bundle map, locate the boundary layer voxel set outside the splenium of the corpus callosum from the multi-scale feature pyramid as the second local region; within the second local region, calculate the cosine similarity between the feature vector of each voxel and the feature vector of voxels in adjacent white matter fiber bundle regions, and calculate the root mean square value of the gradient magnitude of the cosine similarity in the sagittal, coronal, and vertical axes as the feature difference gradient. Use the feature difference gradient as the representation quantity of the second morphological primitive, specifically including:

[0071] A clinically standardized white matter fiber tract map is retrieved. This map is a three-dimensional digital anatomical template containing precise spatial coordinates and orientation information of the white matter fiber tracts, and maintains complete consistency with the three-dimensional coordinate system of all previously processed data to ensure accurate localization. In the multi-scale feature pyramid obtained in step 300, based on the anatomical annotations of the white matter fiber tract map, the boundary layer region on the outer side of the splenium of the corpus callosum is located. This region is the transition area between the white matter fiber tracts of the corpus callosum and the surrounding gray matter and cerebrospinal fluid tissue, which is prone to problems such as feature blurring and boundary offset, and is a key area for feature calibration. All voxels corresponding to this boundary layer region are extracted and integrated to form a voxel set. This voxel set is the second local region, and its spatial dimension is consistent with the feature map dimension of the corresponding scale in the multi-scale feature pyramid. Only the feature data of the boundary layer on the outer side of the splenium of the corpus callosum are retained.

[0072] Within the second local region, each voxel is traversed. Centered on the currently traversed target voxel, voxels belonging to the white matter fiber bundle region within a 3×3×3 three-dimensional neighborhood are selected as adjacent white matter fiber bundle region voxels. The selection criteria for adjacent voxels are based on the labeling of the white matter fiber bundle map, ensuring that all selected adjacent voxels are normal white matter fiber bundle voxels. The cosine similarity between the target voxel's feature vector and the feature vector of each adjacent white matter fiber bundle region voxel is calculated. Cosine similarity quantifies the spatial distribution consistency of two feature vectors; the closer the similarity is to 1, the more similar the feature distributions of the two voxels; the closer it is to 0, the greater the difference in feature distributions. For each target voxel, after calculating its cosine similarity with all adjacent white matter fiber bundle region voxels, the average of all similarities is taken as the final cosine similarity of the target voxel, ensuring the stability of the calculation results. The gradient magnitude of this cosine similarity in the sagittal axis (x-axis), coronal axis (y-axis), and vertical axis (z-axis) is calculated. The gradient calculation uses the central difference method, first calculating the gradient in the sagittal axis direction. coronal axis gradient Vertical axis gradient Gradient modulus The gradient magnitudes of the cosine similarity of each target voxel in three directions are obtained. The root mean square value of the gradient magnitude of each target voxel is calculated as the feature difference gradient of that voxel. This feature difference gradient is used as a representation of the second morphological primitive. The larger the feature difference gradient, the more obvious the feature difference between the voxel and the adjacent white matter fiber bundle voxels, and the more likely there is a feature ambiguity or offset problem; conversely, the smaller the feature difference, the more uniform the feature distribution.

[0073] Step 302: When the feature difference gradient is less than the set tolerance, a three-dimensional diffusion tensor is constructed based on the local fiber orientation, with each voxel in the second local region as the center. The principal axis of the diffusion tensor is parallel to the fiber orientation, and the principal axis diffusion coefficient is set as a fixed multiple of the tangential diffusion coefficient. Specifically, this includes:

[0074] Tolerance for feature difference gradients was set. A large amount of clinical normal brain multi-sequence MRI data was collected, and the feature difference gradients of all voxels in the white matter fiber tract boundary layer were calculated according to the method in step 301. The mean and median of these gradient values ​​were statistically obtained. Using the statistical mean as the core benchmark, combined with the clinical requirements for feature ambiguity judgment, and taking into account the reasonable fluctuation range of the feature difference gradients of normal voxels, a tolerance range of 0.05 to 0.1 was defined. When the statistical mean was low, the lower limit of the interval (0.05) was used; when the mean was high, the upper limit of the interval (0.1) was used, ensuring that the tolerance could cover the feature difference gradient range of all normal voxels while meeting the clinical judgment accuracy. The median was then considered, and a default value of 0.08 was set. This tolerance was used to determine whether a three-dimensional diffusion tensor needed to be constructed for feature calibration. The feature difference gradient of each voxel calculated in step 301 is compared with the set tolerance one by one. When the root mean square value of the feature difference gradient of a voxel is less than the set tolerance, it indicates that the feature difference between the voxel and the adjacent white matter fiber bundle voxels is too small, resulting in feature ambiguity and unclear boundaries. It is necessary to calibrate by constructing a three-dimensional diffusion tensor. For voxels whose feature difference gradient is greater than or equal to the set tolerance, there is no need to construct a diffusion tensor, and their original feature values ​​remain unchanged.

[0075] For voxels requiring calibration, the local white matter fiber orientation at the voxel's location is determined using a white matter fiber bundle map as the center. The local white matter fiber orientation is determined by the fiber orientation markings in that region of the white matter fiber bundle map, ensuring consistency with the actual clinical white matter fiber distribution. A three-dimensional diffusion tensor is constructed based on the local fiber orientation. This three-dimensional diffusion tensor is a third-order symmetric tensor, denoted as Ds. Its principal axis is parallel to the local white matter fiber orientation, and its two tangential directions are perpendicular to the principal axis and also perpendicular to each other, conforming to the spatial distribution characteristics of white matter fibers. The principal axis diffusion coefficient and the tangential diffusion coefficient of the diffusion tensor are set, with the principal axis diffusion coefficient set as a fixed multiple of the tangential diffusion coefficient. The fixed multiple ranges from 2 to 3, with a default value of 2.5, reflecting the diffusion characteristics of clinical white matter fiber bundles (the diffusion ability in the principal axis direction is stronger than in the tangential direction). The specific setting process is as follows: Let the tangential diffusion coefficient be... The value ranges from 0.1 to 0.2, with a default value of 0.15. Therefore, the main axis diffusion coefficient... Based on the above parameters, a three-dimensional diffusion tensor matrix is ​​constructed, specifically in the form of: ;

[0076] In this matrix, the diagonal elements correspond to the diffusion coefficients along the principal axis and the two tangential directions, respectively, while the off-diagonal elements are all 0, ensuring the symmetry and rationality of the diffusion tensor and completing the construction of the three-dimensional diffusion tensor.

[0077] Step 303: Perform three-dimensional nonlinear diffusion iteration based on the diffusion tensor. In each iteration, the new eigenvalue of each voxel is equal to the sum of the products of the current eigenvalues ​​of all voxels in the neighborhood and the diffusion weights. Repeat the iteration until the sum of the absolute values ​​of the changes in the eigenvalues ​​of all voxels is less than the convergence threshold. Use the feature map after the iteration as the second correction feature map, specifically including:

[0078] Initialize the iteration parameters, setting the iteration step size to 0.1, the initial number of iterations to 0, and the convergence threshold to 1e-4. The convergence threshold is used to determine whether the iteration terminates. The second local region feature map obtained in step 301 is used as the initial feature map, denoted as... Based on the three-dimensional diffusion tensor constructed in step 302, a three-dimensional nonlinear diffusion iterative operation is performed. The core operation process of each iteration is as follows: for each voxel in the second local region ( Using the voxel as the center, select a 3×3×3 three-dimensional neighborhood of voxels, and calculate the diffusion weight between the voxel and each voxel in the neighborhood. The diffusion weight is calculated based on the Euclidean distance between the three-dimensional diffusion tensor and the voxels. The specific calculation formula is as follows: ;

[0079] in For target voxels With neighboring voxels Diffusion weights between them For target voxels With neighboring voxels The Euclidean distance between them For target voxels The corresponding principal axis diffusion coefficient of the three-dimensional diffusion tensor (if the neighboring voxels) A diffusion tensor is also constructed, and the average of the principal axis diffusion coefficients of both is taken. In each iteration, the new eigenvalue of each voxel is equal to the sum of the products of the current eigenvalues ​​of all voxels in its neighborhood and their corresponding diffusion weights. The specific calculation formula is as follows: ;

[0080] in, For the first After the iteration, the coordinates ( ),aisle The new voxel eigenvalues, This represents the current iteration number. coordinates ( The set of 3×3×3 three-dimensional neighborhood voxels of a voxel. coordinates ( Voxels and neighboring voxels diffusion weight, For the first After the next iteration, the neighborhood voxels ,aisle The current eigenvalue. After one iteration, calculate the sum of the absolute values ​​of the changes in all voxel eigenvalues, i.e. ,in To sum the absolute values ​​of the changes in eigenvalues ​​of all voxels, iterate through all voxels and all channels within the second local region, calculate the absolute values ​​of the changes in eigenvalues, and sum them. Then, use the calculated... Compare with the preset convergence threshold, if The convergence threshold determines the number of iterations. Add 1, and repeat the above nonlinear diffusion iteration operation; if If a convergence threshold is reached, the iteration is considered convergent, and the iteration process is terminated. After the iteration terminates, the feature map obtained at the end of the iteration (i.e., the one obtained in the last iteration) is... As a second corrected feature map, this second corrected feature map has the same whole-brain spatial resolution (H×W×D×C) as the standard feature tensor. Only the voxel feature values ​​within the second local region are updated by the aforementioned iterative process; the voxel feature values ​​in other regions directly adopt the values ​​at the corresponding positions in the standard feature tensor (i.e., regions without second-level calibration retain their original features). In this feature map, the originally blurred and offset voxel features are effectively corrected, and the feature distribution more closely matches the actual distribution of clinical white matter fiber tracts, improving feature clarity and accuracy.

[0081] This embodiment first fuses features to fill information gaps, then constructs a multi-scale feature pyramid to fully integrate multi-resolution features. Based on the white matter fiber map, it identifies key regions prone to feature anomalies and performs precise local calibration, reducing unnecessary computations. It quantifies feature differences using cosine similarity and the root mean square value of the gradient modulus, accurately determining calibration trigger conditions based on these quantified indicators, resulting in more objective and reliable judgments. By setting relevant parameters for the three-dimensional diffusion tensor based on the actual orientation of white matter fibers, it completes feature correction through stable and controllable three-dimensional nonlinear diffusion iteration, effectively improving regional feature blurring and boundary offset issues, ultimately obtaining a more accurate calibrated feature map.

[0082] In a preferred embodiment of the present invention, step 4 includes:

[0083] Step 400: Construct a three-dimensional deconvolution layer based on the ratio of the current resolution of the second corrected feature map to the target spatial resolution of the standard feature tensor; input the second corrected feature map into the three-dimensional deconvolution layer to obtain an upsampled feature map, specifically including:

[0084] Extract the current spatial resolution of the second corrected feature map obtained in step 303, denoted as (H1, W1, D1), where H1 is the current height, W1 is the current width, and D1 is the current depth. Simultaneously, extract the target spatial resolution of the standard feature tensor obtained in step 100, denoted as (H, W, D), ensuring that the channel dimensions of both remain consistent. Calculate the resolution ratios along the three dimensions: the ratio r along the height direction. H =H / H1=1, the ratio r along the width direction W =W / W1=1, the ratio r along the depth direction D =D / D1=1. If r H =r W =r D If the value is 1, it indicates that the second corrected feature map has a completely consistent whole-brain spatial resolution with the standard feature tensor. At this time, there is no need to perform an upsampling operation; simply set the upsampled feature map to be equal to the second corrected feature map.

[0085] If any of the three ratios is greater than 1, a 3D deconvolution layer needs to be constructed for upsampling. The core parameters of the 3D deconvolution layer are set as follows: the deconvolution kernel size is set to 4×4×4; the stride of each dimension is set to the corresponding resolution ratio, i.e., the stride along the height direction is r. H The step size along the width direction is r W The step size along the depth direction is r D The same padding method is used, that is, the boundaries of each dimension are padded during the deconvolution operation to ensure that the spatial resolution of the output feature map after deconvolution accurately matches the target resolution (H, W, D). The number of input channels of the deconvolution layer is consistent with the number of channels of the second corrected feature map, and the number of output channels is also consistent with the number of channels of the standard feature tensor to avoid channel dimension mismatch. The second corrected feature map is used as input to the constructed 3D deconvolution layer to perform 3D deconvolution operation. A 4×4×4 deconvolution kernel is used, and the stride (r) is calculated according to the corresponding stride of each dimension. H r W r D Traverse the second corrected feature map, perform convolution operation between the deconvolution kernel and the corresponding region of the feature map, and gradually enlarge the spatial dimension of the feature map through the deconvolution operation, and finally output a feature map with spatial resolution of (H, W, D), which is the upsampled feature map.

[0086] Step 401: Map the value of each voxel in the upsampled feature map to between 0 and 1 using a sigmoid activation function to obtain a tumor probability voxel map; extract a third local region from the tumor probability voxel map based on the intersection of the falx cerebri anatomical location and the gray matter mask of the cingulate gyrus; within the third local region, label all connected components with probability values ​​higher than a preset confidence threshold, calculate the centroid coordinates of each connected component, and use two connected components that satisfy the condition that the Euclidean distance between their centroids is less than the tolerance and the number of voxels in each connected component is greater than the minimum volume threshold as the representation quantity of the third morphological primitive, specifically including:

[0087] The feature values ​​of each voxel in the upsampled feature map obtained in step 400 are input into the sigmoid activation function for numerical mapping, uniformly mapping all voxel feature values ​​to the range of 0 to 1, thus obtaining a tumor probability voxel map. The closer the tumor probability value of a voxel is to 1, the greater the probability that the voxel is tumor tissue; the closer it is to 0, the greater the probability that the voxel is normal tissue, achieving a preliminary quantitative distinction between tumors and normal tissues. A standard brain anatomy atlas is retrieved to locate the anatomical position of the falx cerebri and extract the voxel set corresponding to the falx cerebri. Simultaneously, a cingulate gyrus gray matter mask is retrieved, i.e., a large amount of clinical normal brain multi-sequence MRI data is collected. First, clinical statistical analysis is performed on the feature value differences between normal cingulate gyrus gray matter and surrounding white matter and cerebrospinal fluid on the highest resolution feature map of the multi-scale feature pyramid obtained in step 300, determining the feature value threshold range of the cingulate gyrus gray matter to be 0.35 to 0.55 (this threshold range can accurately distinguish the cingulate gyrus gray matter from adjacent tissues, avoiding mask deviation caused by feature overlap); a selection of voxels within this threshold range is then made. All voxels within the mask are used to form an initial cingulate gyrus gray matter mask. Morphological dilation and erosion operations are then performed sequentially. The dilation operation uses 3×3×3 spherical structural elements to fill the micro-holes in the initial mask caused by threshold screening. The erosion operation uses spherical structural elements of the same size to eliminate noise voxel interference at the mask edges, simultaneously optimizing the mask's boundary accuracy to ensure a complete match between the mask and the actual clinical cingulate gyrus gray matter anatomical range. This results in a dedicated marker template for the cingulate gyrus gray matter region. All voxels marked as cingulate gyrus gray matter in this mask are extracted to form the voxel set corresponding to the cingulate gyrus gray matter mask. The intersection of two voxel sets is calculated, i.e., voxels that simultaneously belong to the falx cerebri anatomical location and the cingulate gyrus gray matter mask. The voxel region corresponding to this intersection is designated as the third local region. This region is a key area where tumors are prone to occur and probability judgment bias is likely to occur. All voxels belonging to the third local region in the tumor probability voxel map are extracted to obtain the tumor probability submap corresponding to the third local region.

[0088] A pre-set confidence threshold, determined based on the statistical results of the probability distribution of clinical tumor tissue and normal tissue, ranges from 0.5 to 0.7, with a default value of 0.6. This threshold is used to distinguish between tumor candidate voxels and normal voxels, avoiding misclassification of low-probability normal tissue as tumor while reducing missed detections of tumor tissue. In the tumor probability subplot of the third local region, all voxels with probability values ​​higher than the pre-set confidence threshold are marked. These voxels constitute multiple connected components. The criteria for determining a connected component are: two adjacent voxels (six-neighborhood in 3D space, i.e., directly adjacent in the x, y, and z axes) and both having probability values ​​higher than the confidence threshold belong to the same connected component. The centroid coordinates of each connected component are calculated. For a given connected component, the coordinates of all voxels within that component are statistically analyzed, and the average values ​​of the x, y, and z coordinates of all voxels are calculated, which are the centroid coordinates.

[0089] The system presets a tolerance and a minimum volume threshold. The tolerance is used to determine the spatial correlation between two connected components, with a value ranging from 2 to 4 voxels, and a default value of 3. The minimum volume threshold is used to filter valid connected components and avoid false connected components formed by small noise voxels, with a value ranging from 10 to 20 voxels, and a default value of 15. Two connected components that meet the following two conditions are selected as the characterization quantity of the third morphological primitive: first, the Euclidean distance between the centroids of the two connected components is less than the tolerance, indicating that the two candidate regions are spatially close and may be part of the same tumor or a symmetrical tumor; second, the voxel count of both connected components is greater than the minimum volume threshold, ensuring that the candidate region is real tumor tissue and not a false region formed by noise.

[0090] Step 402: For each connected component, calculate the mirror centroid coordinates relative to the sagittal midplane, and calculate the Euclidean distance between the centroid and the mirror centroid. When there are two different connected components such that the Euclidean distance between the centroid of the first component and the mirror centroid of the second component is less than the tolerance, and the voxel count of both components is greater than the minimum volume threshold, trigger the third calibration, i.e., recursively scan all voxel pairs located on both sides of the sagittal midplane and with symmetrical coordinates in the tumor probability voxel map. Each symmetrical voxel pair contains a first voxel and a second voxel, specifically including:

[0091] The spatial coordinates of the midsagittal plane are determined. The midsagittal plane is the plane of symmetry of the brain, and its spatial coordinates are determined based on standard brain anatomy atlases, denoted as the midpoint of the y-axis coordinates, denoted as y0. Therefore, the equation of the midsagittal plane is y = y0. This plane divides the brain into two symmetrical hemispheres, left and right. Using the midsagittal plane as a symmetry reference, the mirror image position of each connected component can be calculated, providing a basis for determining symmetry deviation. For each connected component obtained in step 401, its mirror centroid coordinates relative to the midsagittal plane are calculated. Let the centroid coordinates of a connected component be (Cx, Cy, Cz). Then, the coordinates of its mirror centroid (Cx', Cy', Cz') are calculated as Cx' = Cx, Cy' = 2y0 - Cy, Cz = Cz'. That is, the x-axis and z-axis coordinates remain unchanged, while the y-axis coordinate is symmetrical about the sagittal midplane y = y0. This ensures that the mirror centroid and the original centroid are symmetrical about the sagittal midplane. The mirror centroid corresponds to the theoretical position of the connected component in the symmetrical hemisphere. By comparing the spatial relationship between the original centroid and the corresponding mirror centroid, the symmetry deviation of the tumor probability distribution can be determined. The Euclidean distance between the centroid of each connected component and its own mirror centroid is calculated using the same formula as in step 401. This is used to initially determine the symmetry rationality of the connected component. If the distance is too large, it indicates that the candidate region corresponding to the connected component may have a positional deviation. Two distinct connected components that meet the following two conditions are selected to trigger the third calibration: First, the Euclidean distance between the centroid of the first connected component and the mirror centroid of the second connected component is less than the preset tolerance (consistent with the tolerance value in step 401, default 3 voxels). This indicates that the two connected components are located on opposite sides of the sagittal midplane and are spatially symmetrical, which conforms to the symmetrical anatomical characteristics of the brain. If there is a difference in the probability distribution between the two, calibration is required. Second, the number of voxels in both connected components is greater than the minimum volume threshold (consistent with the minimum volume threshold value in step 401, default 15 voxels). This ensures that both connected components are real tumor candidate regions and avoids invalid calibration of false regions.

[0092] When the above two conditions are met, the third calibration is triggered. The core operation of the calibration is to recursively scan all voxel pairs located on both sides of the sagittal midplane and with symmetrical coordinates in the tumor probability voxel map. Each symmetrical voxel pair contains a first voxel and a second voxel, where the coordinates of the first voxel are (x, y, z) and the coordinates of the second voxel are (x, 2y0-y, z). This ensures that the two voxels are symmetrical about the sagittal midplane y=y0. Traversing all such symmetrical voxel pairs is to correct the probability deviation of the symmetrical position by adjusting the probability values ​​of the symmetrical voxel pairs, so that the tumor probability distribution is more in line with the symmetrical anatomical structure of the brain.

[0093] Step 403: Based on the difference between the probability values ​​of the first voxel and the second voxel, and the spatial weight function, a regularization term is obtained. The value of the spatial weight function is determined by the vertical distance between the location of the symmetrical voxel pair and the falx cerebri. When the vertical distance is zero, the spatial weight function takes the value of 1; as the vertical distance increases, the spatial weight function linearly decreases to 0. The regularization terms of all symmetrical voxel pairs are summed to form the total regularization loss. The probability value of each symmetrical voxel pair in the tumor probability voxel map is updated to reduce the probability difference at symmetrical positions. After adjusting all symmetrical voxel pairs, a corrected probability voxel map is obtained, specifically including:

[0094] Define a spatial weighting function whose value is determined by the vertical distance of the location of the symmetric voxel pair from the falx brain, and use it to adjust the correction weights of symmetric voxel pairs at different locations. The specific definition of the spatial weighting function is as follows: Let the vertical distance of the location of the symmetric voxel pair from the falx brain be... ,when When =0 (i.e., the voxel is located on the falx cerebri), the spatial weighting function takes a value of 1; when When d > 0, the spatial weight function decreases linearly to 0 as d increases. The specific calculation formula is as follows: ,in The maximum vertical distance from the edge of the falx cerebri to the midsagittal plane is determined statistically based on standard brain anatomical atlases, with a default value of 5 voxels to ensure that the weighting distribution closely reflects the actual brain anatomy. For each symmetrical voxel pair (first voxel, second voxel), let the probability value of the first voxel be P1 and the probability value of the second voxel be P2. Calculate the absolute value of the difference between the two voxels and then apply this difference to the spatial weighting function. Multiplying these terms yields the regularization term for the symmetrical voxel pair. A larger regularization term indicates a greater probability difference between the symmetrical voxel pairs, requiring a larger correction. The regularization terms for all symmetrical voxel pairs in the tumor probability voxel map are summed to obtain the total regularization loss. This total regularization loss measures the overall symmetry deviation of the tumor probability voxel map; a larger loss value indicates a more pronounced symmetry deviation.

[0095] Based on the total regularization loss, the probability value of each symmetrical voxel pair in the tumor probabilistic voxel map is updated to reduce the probability difference at symmetrical positions. Specifically, the update rule is as follows: for symmetrical voxel pairs ( , ), calculate the average probability value of the two. Update the probability value of the first voxel to The probability value of the second voxel is updated to ,in This is a correction coefficient, ranging from 0.1 to 0.3, with a default value of 0.2. It controls the correction magnitude to prevent over-correction from distorting the probability values. After performing the above update operation on all symmetrical voxel pairs, the corrected tumor probability voxel map is obtained, i.e., the corrected probability voxel map. The symmetry deviation of this map is effectively corrected, and the tumor probability distribution is more consistent with the symmetrical anatomical structure of the brain.

[0096] This embodiment constructs an adapted 3D deconvolution layer by adjusting the resolution ratio, upsampling the second calibrated feature map to a resolution consistent with the standard feature tensor. This avoids probability determination bias caused by resolution mismatch while preserving accurate feature information after calibration. A third local region is located based on the intersection of the gray matter masks of the falx cerebri and cingulate gyrus, focusing on areas prone to tumor occurrence and bias. A third morphological primitive is constructed through connected component analysis and centroid calculation to accurately screen tumor candidate regions, avoiding interference from irrelevant regions and improving the specificity and accuracy of tumor probability determination. Based on sagittal midplane symmetry, the probability difference of symmetrical voxel pairs is corrected through mirror centroid calculation and regularization loss adjustment. This solves the problem of tumor probability distribution bias caused by neglecting the symmetrical structure of the brain in existing methods, making the corrected probability voxel map more closely match the actual anatomical structure of the brain and improving the accuracy of tumor region localization.

[0097] In a preferred embodiment of the present invention, step 5 includes:

[0098] Step 500: Using a preset isosurface probability value as a benchmark, traverse each voxel in the corrected probability voxel map, compare the probability values ​​at the eight vertices of each voxel with the extraction benchmark, and generate an 8-bit binary index value based on the state combination of each vertex being higher or lower than the extraction benchmark. Specifically, this includes:

[0099] First, a baseline for isosurface extraction was established. A large amount of clinically confirmed brain MRI tumor case data was collected. The pathologically confirmed tumor region and normal tissue region were extracted from each case. The voxel probability values ​​corresponding to these two regions were derived from the corrected probability voxel map obtained in step 403. That is, after regularization loss calculation and probability value update in step 403, the corrected probability voxel map already contains accurate probability values ​​for all voxels. The voxel probability values ​​for the corresponding regions in this map can be directly extracted. Statistical analysis algorithms were used to quantify the extracted voxel probability values, statistically obtaining the minimum and average voxel probability values ​​for the tumor region, and the maximum and average voxel probability values ​​for the normal tissue region. Using 0.8 times the average probability value of the tumor region as a reference baseline, and considering the probability overlap between normal and tumor tissues, an outlier removal algorithm was used to remove invalid data. Finally, the range of isosurface probability values ​​was determined to be 0.5 to 0.6, with a default value of 0.55. This benchmark can accurately distinguish between tumor areas and normal areas, avoiding misclassification of normal tissue as tumor and preventing the omission of tumor boundaries, ensuring that the extracted isosurfaces can accurately correspond to the actual boundaries of tumor tissue.

[0100] Iterate through each voxel in the corrected probability voxel graph obtained in step 403. Each voxel is a cube in three-dimensional space, containing 8 vertices. The 8 vertices are divided according to the positive and negative directions of the three-dimensional space coordinates (x, y, z), with reference to the geometric center of the voxel. min ,y min ,z min ), (x min ,y min ,z max ), (x min ,y max ,z min ), (x min ,y max ,z max ), (x max ,y min ,z min ), (x max ,y min ,z max ), (x max ,y max ,z min ), (x max ,y max ,z max ), where x min x max Let x and y be the minimum and maximum coordinates of the voxel element along the x-axis, respectively. min y max Let x and y be the minimum and maximum coordinates along the y-axis, respectively, and z be the maximum and minimum coordinates along the z-axis. min zmax These are the minimum and maximum coordinates along the z-axis, respectively; the eight vertices are labeled sequentially from vertex 0 to vertex 7 in a fixed order, with the following correspondence: vertex 0 (x... min ,y min ,z min Vertex 1 (x) min ,y min ,z max Vertex 2 (x) min ,y max ,z min Vertex 3 (x) min ,y max ,z ma Vertex 4 (x) max ,y min ,z min Vertex 5 (x) max ,y min ,z max Vertex 6 (x) max ,y max ,z min Vertex 7 (x) max ,y max ,z max Each vertex corresponds to a unique spatial coordinate, and its probability value is calculated by trilinear interpolation using a modified probability voxel map, ensuring the continuity and accuracy of the vertex probability values. The probability values ​​at the eight vertices of each voxel unit are compared with a preset isosurface extraction benchmark. If the vertex probability value is higher than the benchmark, it is marked as 1; if the vertex probability value is lower than or equal to the benchmark, it is marked as 0, completing the probability state quantization and marking of all vertices. Following a fixed order from vertex 0 to vertex 7, the marking results (0 or 1) of the eight vertices of each voxel unit are combined to generate an 8-bit binary number. This binary number is the binary index value of that voxel unit, with the index value ranging from 0 to 255. Each index value corresponds to a vertex probability state combination.

[0101] Step 501: Based on the binary index value, read the vertex connection pattern of the triangular facet within the corresponding voxel element from the predefined isosurface triangle mesh lookup table; for each triangle vertex to be generated, calculate the spatial coordinates of the corresponding vertex; merge the triangular facets generated by all voxel elements to obtain the initial isosurface mesh, specifically including:

[0102] A predefined isosurface triangle subdivision lookup table is used. This lookup table contains vertex connection patterns for triangle patches corresponding to all binary index values ​​from 0 to 255. Each index value corresponds to 1 to 5 triangle patches (determined based on vertex probability state combinations). Each triangle patch consists of 3 vertex indices, clearly defining the vertex connection relationships of the triangles. This lookup table is a commonly used template for clinical isosurface extraction, ensuring that the subdivision pattern closely matches the actual shape of the tumor boundary. For each voxel unit, based on its generated 8-bit binary index value, the corresponding triangle patch vertex connection pattern is read from the predefined triangle subdivision lookup table to determine the number of triangle patches to be generated within that voxel unit and the vertex indices of each patch. For each triangle facet to be generated, the spatial coordinates of the corresponding vertex are calculated based on the vertex probability value corresponding to its vertex index and the extraction datum (the extraction datum is the isosurface probability value determined in step 500, with a value range of 0.5 to 0.6, and a default value of 0.55). The specific calculation process is as follows: For a vertex corresponding to a given vertex index, if its probability value is higher than the extraction datum, the original spatial coordinates of that vertex are directly used; if its probability value is lower than or equal to the extraction datum, linear interpolation is used to interpolate the vertex coordinates located on the isosurface between that vertex and adjacent high-probability vertices. The interpolation formula is... ,in For the interpolated vertex coordinate, , For adjacent vertices coordinate, , These are the probability values ​​of adjacent vertices. To extract the baseline for the isosurface, the interpolation calculation method for the y and z coordinates is the same as that for the x coordinate. After calculating the vertex coordinates of all triangular patches, the triangular patches generated by all voxel elements are merged, and duplicate triangular patches (patterns with completely identical vertex coordinates) are removed to obtain the initial isosurface mesh. This mesh initially outlines the three-dimensional boundary morphology of the tumor region, but it has problems such as mesh coarseness and uneven boundaries.

[0103] Step 502: For each vertex of the initial isosurface mesh, collect the position coordinates of all adjacent vertices in the first-order neighborhood and calculate the arithmetic mean. Move the current vertex towards the arithmetic mean by a certain proportion, update synchronously, and repeat until the maximum distance of a single movement is less than the smooth convergence value to obtain a smooth mesh. Traverse all edges in the smooth mesh, mark the boundary edges that are used only once by a triangle, and form closed boundary loops according to the topological connection relationship of the boundary edges. Each closed boundary loop corresponds to a hole. For each hole, select the three adjacent edges with the smallest interior angle and connect them to form a new triangle. Repeat until the hole is filled to obtain a three-dimensional surface model of the brain MRI tumor region, specifically including:

[0104] The initial isosurface mesh is smoothed using the Laplace smoothing algorithm, which involves traversing each vertex of the initial isosurface mesh, collecting the position coordinates of all neighboring vertices in the first-order neighborhood of that vertex (the first-order neighborhood refers to all vertices directly connected to that vertex), and calculating the arithmetic mean of these neighboring vertex position coordinates, i.e., the neighborhood mean coordinates. The coordinates of the current vertex are shifted towards the average coordinates of its neighborhood by a certain proportion, set to 0.3 (ranging from 0.2 to 0.4 to ensure smoothness while avoiding mesh deformation). The coordinates of the vertex after the shift are ( The calculation formula is: ;

[0105] in The coordinates of the current vertex before it moves are given. The coordinates of all vertices in the mesh are updated synchronously, and the above smoothing iteration process is repeated. After each iteration, the maximum distance of a single movement of all vertices is calculated, which is the maximum value of the Euclidean distance between the coordinates of all vertices before and after the movement. When the maximum distance of a single movement is less than the smoothing convergence value, the iteration stops. The smoothing convergence value ranges from 0.01 to 0.05, with a default value of 0.03. At this point, a smooth mesh is obtained, and the mesh surface becomes flat and continuous, eliminating the roughness of the initial mesh.

[0106] The smooth mesh undergoes hole detection and filling. This involves traversing all edges in the smooth mesh, counting the number of times each edge is used by a triangular facet, and marking edges used only once by a single triangular facet – these are designated as boundary edges. All boundary edges are then sorted according to their topological connections, forming closed boundary loops. Each closed boundary loop corresponds to a hole in the mesh. For each hole, the three adjacent edges with the smallest interior angles on the boundary loop are selected, and their endpoints are connected to form a new triangular facet, filling a portion of the hole. This process is repeated, continuously selecting the three adjacent edges with the smallest interior angles on the remaining boundary loops to connect new triangles, until the hole is completely filled, ensuring the filled mesh is complete and defect-free. After filling all holes, a final three-dimensional surface model of the brain MRI tumor region is obtained. This model accurately recreates the three-dimensional morphology, boundary contours, and spatial location of the tumor, closely matching the actual clinical tumor distribution.

[0107] This embodiment extracts the isosurface of the tumor boundary through a preset isosurface benchmark, binary index value matching, and linear interpolation calculation. Triangular patches are generated based on a predefined subdivision lookup table to ensure the extracted isosurface closely matches the actual tumor boundary, avoiding boundary distortion or omission. The Laplacian smoothing algorithm eliminates mesh roughness, and boundary loop detection and triangle filling repair mesh holes, solving the problems of initial mesh roughness and defects. The final 3D curved surface model has a smooth surface and complete structure, restoring the 3D morphology and spatial location of the tumor.

[0108] like Figure 2 As shown, embodiments of the present invention also provide a three-dimensional reconstruction system for brain MRI tumor regions based on convolutional neural networks, comprising:

[0109] The processing module is used to preprocess the raw voxel data of multi-sequence MRI of the patient's brain to obtain a standard feature tensor; the standard feature tensor is input into a three-dimensional depth-separable convolutional encoder to output a feature map, and the feature map is directionally filtered to obtain a set of feature maps;

[0110] The first calibration module is used to extract the first morphological primitives in the first local region from the feature map set. When the anisotropy of the first morphological primitives is lower than a set value and continues to decay along the depth direction, the first calibration is performed on the feature map set to obtain the first corrected feature map.

[0111] The second calibration module is used to fuse the first calibration feature map with the standard feature tensor to construct a multi-scale feature pyramid. It extracts the second morphological primitives in the second local region from the multi-scale feature pyramid. When the feature difference gradient between the second morphological primitives and the adjacent regions is less than the set tolerance, the second calibration is performed on the second local region to obtain the second calibration feature map.

[0112] The third calibration module is used to convert the second correction feature map into a tumor probability voxel map, extract the third morphological primitives in the third local region from the tumor probability voxel map, and when there are two connected bodies in the third morphological primitives whose symmetric distance is within the tolerance and whose volume exceeds the minimum volume, the third calibration is performed on the tumor probability voxel map to obtain the corrected probability voxel map.

[0113] The extraction module is used to extract isosurface meshes from the modified probability voxel map based on a set benchmark, and to smooth and fill holes in the isosurface meshes to obtain a three-dimensional surface model of the brain MRI tumor region.

[0114] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0115] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0116] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0117] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks, characterized in that, The method includes: Step 1: Preprocess the raw voxel data from the multi-sequence MRI of the patient's brain to obtain a standard feature tensor; input the standard feature tensor into a 3D depth-separable convolutional encoder to output a feature map, and perform directional filtering on the feature map to obtain a set of feature maps; Step 2: Extract the first morphological primitive from the first local region from the feature map set. When the anisotropy of the first morphological primitive is lower than the set value and continues to decay along the depth direction, perform the first recalibration on the feature map set to obtain the first corrected feature map. Step 3: Fuse the first corrected feature map with the standard feature tensor to construct a multi-scale feature pyramid. Extract the second morphological primitives in the second local region from the multi-scale feature pyramid. When the feature difference gradient between the second morphological primitives and the adjacent regions is less than the set tolerance, perform a second recalibration on the second local region to obtain the second corrected feature map. Step 4: Convert the second correction feature map into a tumor probability voxel map, extract the third morphological primitives in the third local region from the tumor probability voxel map, and when there are two connected bodies in the third morphological primitives whose symmetric distance is within the tolerance and whose volume exceeds the minimum volume, perform a third calibration on the tumor probability voxel map to obtain the corrected probability voxel map. Step 5: Extract isosurface mesh from the modified probability voxel map based on the set benchmark, and smooth and fill the holes in the isosurface mesh to obtain a three-dimensional surface model of the brain MRI tumor region.

2. The method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks according to claim 1, characterized in that, The standard feature tensor is input into a 3D depth-separable convolutional encoder, which outputs feature maps. The feature maps are then directionally filtered to obtain a set of feature maps, including: The standard feature tensor is used as the input data of the 3D depth separable convolutional encoder, which includes multiple cascaded 3D depth separable convolutional modules. Each module sequentially performs 3D depth separable convolution, batch normalization and linear rectified activation to obtain the initial feature map of the corresponding scale. For each initial feature map, the response intensity of each voxel position is calculated along the sagittal, coronal, and vertical axes, and spatial nonmaximum suppression is performed in each direction. That is, for each direction, only the voxel activation values ​​whose response intensity is greater than the response intensity of all voxels other than the corresponding voxel itself in the three-dimensional neighborhood of the corresponding direction are retained, and the activation values ​​of all voxels that do not meet the condition are set to zero, resulting in sub-feature maps after filtering in three directions. The sub-feature maps after filtering in three directions corresponding to the same initial feature map are concatenated according to the channel dimension to obtain the feature map set.

3. The method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks according to claim 2, characterized in that, Step 2 includes: From the feature map set, the voxel coordinate range of the junction of the anterior horn of the lateral ventricle and the septum pellucidum is located according to the brain anatomy atlas, and the feature map subset within the range is taken as the first local region; Within the first local region, a three-dimensional structure tensor is constructed for the feature vector of each voxel. The three eigenvalues ​​of the three-dimensional structure tensor are calculated to obtain the anisotropy fraction. The anisotropy fraction is used as the representation of the first morphological primitive. The difference in anisotropy fractions of corresponding voxels on adjacent slices is calculated along the depth direction to form a difference sequence. When the arithmetic mean of the anisotropy fractions of all voxels in the first local region is lower than the preset homogeneity threshold, and the difference sequence is negative on multiple consecutive slices, the first calibration is triggered, that is, three-dimensional global average pooling is performed on all channels in the feature map set corresponding to the first local region to obtain a channel description vector. Based on the dimension of the channel description vector and the preset compression ratio, a first linear transformation matrix and a second linear transformation matrix are constructed. After multiplying the channel description vector with the first linear transformation matrix, the dimension is restored by ReLU nonlinear mapping and the second linear transformation matrix. Finally, the channel weight vector between 0 and 1 is output by mapping with the sigmoid saturation activation function. The channel weight vector is multiplied element-wise with each channel feature map corresponding to the first local region in the feature map set to obtain the first corrected feature map.

4. The method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks according to claim 3, characterized in that, Within the first local region, a three-dimensional structure tensor is constructed for the feature vectors of each voxel. The three eigenvalues ​​of the three-dimensional structure tensor are calculated to obtain the anisotropy fraction. This anisotropy fraction is used as a representation of the first morphological primitive, including: For each voxel in the first local region, with the corresponding voxel position as the center, take the set of feature vectors in the three-dimensional neighborhood, calculate the gradient component of each feature vector, and perform a weighted summation of the outer product of each feature vector with itself to construct a three-dimensional structure tensor matrix. The three-dimensional structure tensor matrix is ​​a third-order symmetric positive semi-definite matrix. The eigenvalue decomposition of the three-dimensional structural tensor matrix yields three non-negative eigenvalues, which are denoted as the first eigenvalue, second eigenvalue, and third eigenvalue in descending order of value. Anisotropy scores are calculated based on the first eigenvalue, second eigenvalue, and third eigenvalue. These anisotropy scores are used to quantify the orientational consistency of the voxel structure within the first local region. The closer the score is to 1, the stronger the anisotropy.

5. The method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks according to claim 4, characterized in that, Step 3 includes: The first corrected feature map is added element-wise to the sub-tensor of the corresponding first local region in the standard feature tensor, while the non-first local region parts in the standard feature tensor remain unchanged, resulting in a fused feature map. Three-dimensional max pooling with downsampling ratios of 2, 4, and 8 is applied to the fused feature map to obtain first-scale feature copies, second-scale feature copies, and third-scale feature copies. The third-scale feature copy is upsampled to the resolution of the second-scale feature copy and added element-wise to the second-scale feature copy to obtain the first fusion result. The first fusion result is upsampled to the resolution of the first-scale feature copy and added element-wise to the first-scale feature copy to obtain the second fusion result. The second fusion result is upsampled to the original resolution of the fused feature map to obtain the multi-scale feature pyramid. Based on the white matter fiber bundle map, the boundary layer voxel set outside the splenium of the corpus callosum is located from the multi-scale feature pyramid as the second local region. Within the second local region, the cosine similarity between the feature vector of each voxel and the feature vector of the voxel in the adjacent white matter fiber bundle region is calculated, and the root mean square value of the gradient magnitude of the cosine similarity in the three directions of the sagittal axis, coronal axis and vertical axis is obtained as the feature difference gradient. The feature difference gradient is used as the representation quantity of the second morphological primitive. When the feature difference gradient is less than the set tolerance, a three-dimensional diffusion tensor is constructed based on the local fiber orientation, with each voxel in the second local region as the center. The principal axis of the diffusion tensor is parallel to the fiber orientation, and the principal axis diffusion coefficient is set as a fixed multiple of the tangential direction diffusion coefficient. Three-dimensional nonlinear diffusion iteration is performed based on the diffusion tensor. In each iteration, the new eigenvalue of each voxel is equal to the sum of the products of the current eigenvalues ​​of all voxels in the neighborhood and the diffusion weight. The iteration is repeated until the sum of the absolute values ​​of the changes in the eigenvalues ​​of all voxels is less than the convergence threshold. The feature map after the iteration is completed is used as the second correction feature map.

6. The method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks according to claim 5, characterized in that, Step 4 includes: Based on the ratio of the current resolution of the second corrected feature map to the target spatial resolution of the standard feature tensor, a three-dimensional deconvolution layer is constructed; the second corrected feature map is input into the three-dimensional deconvolution layer to obtain the upsampled feature map; The value of each voxel in the upsampled feature map is mapped to between 0 and 1 using a sigmoid activation function to obtain a tumor probability voxel map. Based on the intersection of the falx cerebri anatomical location and the gray matter mask of the cingulate gyrus, a third local region is extracted from the tumor probability voxel map. Within the third local region, all connected components with probability values ​​higher than a preset confidence threshold are marked. The centroid coordinates of each connected component are calculated, and two connected components that satisfy the condition that the Euclidean distance between the centroids of two connected components is less than the tolerance and the number of voxels in each connected component is greater than the minimum volume threshold are used as the representation of the third morphological primitive. For each connected component, the coordinates of the mirror centroid relative to the sagittal midplane are calculated, and the Euclidean distance between the centroid and the mirror centroid is calculated. When there are two different connected components such that the Euclidean distance between the centroid of the first component and the mirror centroid of the second component is less than the tolerance, and the number of voxels of both components is greater than the minimum volume threshold, the third calibration is triggered, that is, recursively scanning all voxel pairs located on both sides of the sagittal midplane and with symmetrical coordinates in the tumor probability voxel map, each symmetrical voxel pair containing the first voxel and the second voxel. Based on the difference between the probability values ​​of the first voxel and the second voxel, and the spatial weight function, a regularization term is obtained. The value of the spatial weight function is determined by the vertical distance between the location of the symmetrical voxel pair and the falx cerebri. When the vertical distance is zero, the spatial weight function takes the value of 1, and when the vertical distance increases, the spatial weight function linearly decreases to 0. The regularization terms of all symmetrical voxel pairs are accumulated to form the total regularization loss. The probability value of each symmetrical voxel pair in the tumor probability voxel map is updated to reduce the probability difference at symmetrical positions. After adjusting all symmetrical voxel pairs, the corrected probability voxel map is obtained.

7. The method for three-dimensional reconstruction of brain MRI tumor regions based on convolutional neural networks according to claim 6, characterized in that, Step 5 includes: Using the preset isosurface probability value as a benchmark, traverse each voxel in the modified probability voxel map, compare the probability values ​​at the eight vertices of each voxel with the extraction benchmark, and generate an 8-bit binary index value based on the state combination of each vertex being higher or lower than the extraction benchmark. Based on the binary index value, the vertex connection pattern of the triangular facet within the corresponding voxel element is read from the predefined isosurface triangle mesh lookup table; for each triangle vertex to be generated, the spatial coordinates of the corresponding vertex are calculated; the triangular facets generated by all voxel elements are merged to obtain the initial isosurface mesh. For each vertex of the initial isosurface mesh, the position coordinates of all adjacent vertices in the first-order neighborhood are collected and the arithmetic mean is calculated. The current vertex is moved towards the arithmetic mean by a certain proportion, and the process is repeated synchronously until the maximum distance of a single movement is less than the smooth convergence value, thus obtaining a smooth mesh. All edges in the smooth mesh are traversed, and boundary edges that are used only once by a triangle are marked. The boundary edges are connected according to the topological relationship to form closed boundary loops. Each closed boundary loop corresponds to a hole. For each hole, the three adjacent edges with the smallest interior angle are selected and connected to form a new triangle. This process is repeated until the hole is filled, thus obtaining a three-dimensional surface model of the brain MRI tumor region.

8. A three-dimensional reconstruction system for brain MRI tumor regions based on convolutional neural networks, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The processing module is used to preprocess the raw voxel data of multi-sequence MRI of the patient's brain to obtain a standard feature tensor; the standard feature tensor is input into a three-dimensional depth-separable convolutional encoder to output a feature map, and the feature map is directionally filtered to obtain a set of feature maps; The first calibration module is used to extract the first morphological primitives in the first local region from the feature map set. When the anisotropy of the first morphological primitives is lower than a set value and continues to decay along the depth direction, the first calibration is performed on the feature map set to obtain the first corrected feature map. The second calibration module is used to fuse the first calibration feature map with the standard feature tensor to construct a multi-scale feature pyramid. It extracts the second morphological primitives in the second local region from the multi-scale feature pyramid. When the feature difference gradient between the second morphological primitives and the adjacent regions is less than the set tolerance, the second calibration is performed on the second local region to obtain the second calibration feature map. The third calibration module is used to convert the second correction feature map into a tumor probability voxel map, extract the third morphological primitives in the third local region from the tumor probability voxel map, and when there are two connected bodies in the third morphological primitives whose symmetric distance is within the tolerance and whose volume exceeds the minimum volume, the third calibration is performed on the tumor probability voxel map to obtain the corrected probability voxel map. The extraction module is used to extract isosurface meshes from the modified probability voxel map based on a set benchmark, and to smooth and fill holes in the isosurface meshes to obtain a three-dimensional surface model of the brain MRI tumor region.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images