Tumor placeholder brain function partition nerve influence image segmentation method based on deep learning model

Abstract

The invention discloses a tumor space occupying brain function partition nerve influence image segmentation method based on a deep learning model, and particularly relates to the technical field of artificial intelligence and medical cross. Comprising the following collaborative operation steps: multi-modal data fusion acquisition, tumor edge feature enhancement, anatomy-functional feature collaborative modeling, occupation deformation compensation, adversarial boundary optimization, dynamic loss regulation and control, cascade segmentation strategy, clinical interaction verification, and construction of a U-Net + + based improved three-dimensional feature fusion network. Utilizing a multi-scale cavity convolution module to extract edge gradient features of the tumor infiltration area in a layered manner; in order to solve the problem of multi-modal fusion and dynamic adaptation mechanism deficiency in the prior art, a three-dimensional cross-modal fusion network of a channel competition mechanism is constructed, multi-modal key features are dynamically screened through spatial pyramid pooling and soft attention weight, and a tumor volume self-adaptive dynamic loss function is combined, so that multi-modal fusion and dynamic adaptation are realized. The problems of feature confusion and poor generalization of a traditional method are solved.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology of artificial intelligence and medicine, and more specifically, to a method for segmenting images of neural impact of tumor-occupying brain functional zonation based on a deep learning model. Background Art

[0002] A deep learning-based neuroimaging image segmentation method for tumor-occupying brain functional zonation uses artificial intelligence technology to accurately identify tumor areas and their relationship to brain functional zonation, significantly improving the efficiency and accuracy of medical image analysis. It can help doctors more clearly and quickly determine tumor boundaries, size, and impact on brain functional areas, providing a reliable basis for developing personalized surgical and radiotherapy plans. It also helps assess disease prognosis, promotes the development of precision medicine for tumors, and is of great significance for improving the diagnosis and treatment of brain tumor patients and their quality of life.

[0003] The existing technology has the following problems:

[0004] 1. Lack of multimodal data fusion and dynamic adaptation mechanisms

[0005] Existing methods for the fusion of structural and functional images mostly use simple feature splicing, and fail to filter key modal information through the attention mechanism, resulting in insufficient extraction of tumor boundary features (such as confusion between T2 edema areas and fMRI activation areas). At the same time, the fixed weight loss function cannot adapt to different tumor volumes and boundary complexities, resulting in the coexistence of missed segmentation of small tumors and blurred boundaries of large tumors, which seriously restricts clinical generalization capabilities.

[0006] 2. Insufficient separation and verification of anatomical and functional features

[0007] Traditional models independently process gray matter anatomical structure and functional connectivity characteristics and lack a cross-modal collaborative optimization mechanism, resulting in a contradiction between the anatomical rationality and functional continuity of the segmentation results. For example, the motor area is morphologically intact but the BOLD signal connection is interrupted. In addition, DTI fiber tracking and biomechanical models are not integrated for result verification, making it impossible to detect the anatomical integrity of key pathways such as the pyramidal tract, affecting the surgical safety assessment.

[0008] 3. Tumor space deformation compensation and real-time defects

[0009] Existing algorithms ignore the brain tissue displacement effect caused by tumor growth and directly apply standard brain maps, resulting in functional area positioning deviations (such as misjudging the deformed precentral gyrus as a tumor). Complex deformation field calculation and multi-scale feature fusion result in single-case processing time exceeding 2 minutes and GPU memory usage exceeding 12GB, making it difficult to meet the real-time navigation requirements during surgery.

[0010] 4. Weak clinical interactivity and quantitative output

[0011] Fully automated segmentation results lack a physician-correction interface and standardized quantitative reporting, forcing physicians to spend an additional 30 minutes or more performing manual measurements. Furthermore, the single mask output is not deeply integrated with the neuronavigation system (due to poor DICOM interface compatibility), resulting in a disconnect between AI-assisted and clinical workflows and an actual adoption rate of less than 42%.

[0012] Therefore, to address the above problems, a neural impact image segmentation method for tumor-occupied brain functional zonation based on a deep learning model is proposed. Summary of the Invention

[0013] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for segmenting a tumor-occupied brain functional zonation neural impact image based on a deep learning model to solve the problems raised in the above-mentioned background technology.

[0014] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based neuroimaging segmentation method for tumor-occupying brain functional zonation, comprising the following collaborative steps:

[0015] S1. Multimodal data fusion acquisition: Structural MRI (T1 / T2 / FLAIR), functional MRI (BOLD signal), and diffusion tensor imaging (DTI) data are acquired through simultaneous registration to establish a three-dimensional voxel-based spatial map that includes tumor mass effect and functional connectivity.

[0016] S2. Tumor edge feature enhancement: We constructed an improved 3D feature fusion network based on U-Net++, hierarchically extracted edge gradient features of tumor infiltration regions using a multi-scale dilated convolution module (with dilation rates of 2 / 4 / 6), and aligned the heterogeneous textures of T1 / T2 / FLAIR sequences using a cross-modal fusion layer.

[0017] S3. Collaborative modeling of anatomical and functional features: A dual-path attention mechanism is designed. The anatomical path uses 3D ResNet-50 to extract gray matter / white matter microstructural features, while the functional path uses a graph convolutional network (GCN) to model the functional connectivity topology of the BOLD signal. The two types of features are dynamically fused using a cross-attention matrix.

[0018] S4. Occupancy deformation compensation: Based on the biomechanical model, an elastic registration sub-network is constructed. Combined with the spatial distribution prior of the functional activation heat map, the vector field generator is used to predict the tumor growth direction and output an adaptive deformation field to correct the spatial displacement of the functional area.

[0019] S5. Adversarial boundary optimization: A conditional generative adversarial network (cGAN) is integrated, and the discriminator is used to extract the diffusion tensor anisotropy (FA) features of the tumor-functional area transition zone. The consistency between the segmentation boundary and the direction of the white matter fiber bundles is optimized through adversarial training.

[0020] S6. Dynamic loss control: Construct a multi-objective loss function W_total = αL_dice + βL_edge + γ*L_topology, where α, β, and γ are dynamically adjusted according to the tumor volume ratio, and a boundary complexity index (curvature / fractal dimension) is introduced to adaptively improve the β weight;

[0021] S7. Cascade segmentation strategy: A two-stage cascade training strategy is used. In the first stage, 3D Mask R-CNN is used to locate the tumor core area and generate a placeholder mask. In the second stage, the mask is used as a spatial constraint to segment the motor / language functional area.

[0022] S8. Clinical interactive verification: Deploy a real-time correction module based on the B / S architecture to support doctors to manually adjust the segmentation results through the 3D volume rendering interface, and synchronize the surgical planning path with the intraoperative navigation system through the DICOM interface.

[0023] Preferably, the cross-modal fusion layer in step S2 adopts a feature channel competition mechanism, specifically including: channel normalization of tumor edema area features extracted from T1 / T2 weighted images; generating a channel attention mask based on the necrotic core features of the FLAIR sequence; and dynamically fusing multimodal features through soft attention weights.

[0024] Preferably, the deformation field estimation module of step S4 includes a dual verification mechanism: the first verification layer verifies the anatomical continuity of the pyramidal tract and the precentral gyrus after deformation through DTI fiber tracking; the second verification layer compares the functional connectivity strength of the BOLD signal before and after deformation (Pearson correlation coefficient ≥ 0.85).

[0025] Preferably, the adversarial training strategy of step S5 includes a multi-scale discriminator design: a macro-scale discriminator analyzes the spatial relationship between the overall tumor occupancy and the ventricular system; a meso-scale discriminator detects the distance threshold (≥5 mm) between the hand knob of the motor area and the tumor boundary; and a micro-scale discriminator verifies the gradient continuity of the FA value at the segmentation boundary.

[0026] Preferably, the cascade training of step S7 adopts a knowledge distillation strategy: in the first stage, the teacher network generates a tumor probability heat map; in the second stage, the student network fuses the heat map with the original image to perform functional area segmentation, and then constrains the consistency of the feature distribution of the two stages through the KL divergence loss.

[0027] Preferably, step S1 includes a quantification preprocessing of the space-occupying effect, which simulates the brain tissue displacement field caused by tumor space occupation based on finite element analysis, quantifies the degree of compression and deformation of white matter fiber bundles using tractography, and then establishes a nonlinear regression model (R 2 ≥0.91)

[0028] Preferably, the interactive correction module of step S8 implements incremental learning, records the doctor's correction trajectory and generates a differential mask, updates the local feature extractor weights through online learning, and establishes a user habit model to predict high-frequency correction areas.

[0029] Preferably, the final output includes a multi-dimensional clinical report, including a three-dimensional heat map of the volume ratio of the functional area invaded by the tumor, a white matter fiber bundle integrity score, a risk probability distribution of the optimal craniotomy approach, and a machine learning prediction value of the risk of postoperative functional impairment.

[0030] The technical effects and advantages of the present invention are as follows:

[0031] 1. To address the problem of lack of multimodal fusion and dynamic adaptation mechanisms in existing technologies, the present invention constructs a three-dimensional cross-modal fusion network with a channel competition mechanism. It dynamically screens multimodal key features through spatial pyramid pooling and soft attention weights, and combines it with a dynamic loss function for tumor volume adaptation to solve the problems of feature confusion and poor generalization in traditional methods.

[0032] 2. In response to the problems of insufficient anatomical-functional feature separation and verification in existing technologies, the present invention designs a dual-path attention mechanism. The anatomical path and the functional path fuse the gray matter structure and BOLD signal characteristics through a cross-attention matrix, and integrate DTI fiber tracking verification and BOLD functional connection strength threshold, breaking through the limitations of feature separation of traditional methods.

[0033] 3. In response to the problems of tumor space-occupying deformation compensation and real-time defects in existing technologies, the present invention develops an adaptive deformation field module based on a biomechanical model, combines finite element analysis with tumor growth direction prediction, and adopts a cascade segmentation strategy to solve the problems of large positioning deviation and long calculation time in traditional methods.

[0034] 4. In response to the problems of weak clinical interactivity and quantitative output in existing technologies, the present invention deploys a real-time correction module with a B / S architecture, which supports doctors to complete three-dimensional interactive adjustments within 10 seconds and generate multi-dimensional surgical reports. It directly connects to the neuronavigation system through the DICOM interface, solving the clinical disconnection problem of traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the workflow of the present invention. DETAILED DESCRIPTION

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

[0037] like Figure 1 As shown, a deep learning-based neuroimaging segmentation method for brain functional zonation of tumor lesions is disclosed, comprising the following collaborative steps:

[0038] S1. Multimodal data fusion acquisition: Structural MRI (T1 / T2 / FLAIR), functional MRI (BOLD signal), and diffusion tensor imaging (DTI) data are acquired through simultaneous registration to establish a three-dimensional voxel-based spatial map that includes tumor mass effect and functional connectivity.

[0039] S2. Tumor edge feature enhancement: We constructed an improved 3D feature fusion network based on U-Net++, hierarchically extracted edge gradient features of tumor infiltration regions using a multi-scale dilated convolution module (with dilation rates of 2 / 4 / 6), and aligned the heterogeneous textures of T1 / T2 / FLAIR sequences using a cross-modal fusion layer.

[0040] S3. Collaborative modeling of anatomical and functional features: A dual-path attention mechanism is designed. The anatomical path uses 3D ResNet-50 to extract gray matter / white matter microstructural features, while the functional path uses a graph convolutional network (GCN) to model the functional connectivity topology of the BOLD signal. The two types of features are dynamically fused using a cross-attention matrix.

[0041] S4. Occupancy deformation compensation: Based on the biomechanical model, an elastic registration sub-network is constructed. Combined with the spatial distribution prior of the functional activation heat map, the vector field generator is used to predict the tumor growth direction and output an adaptive deformation field to correct the spatial displacement of the functional area.

[0042] S5. Adversarial boundary optimization: A conditional generative adversarial network (cGAN) is integrated, and the discriminator is used to extract the diffusion tensor anisotropy (FA) features of the tumor-functional area transition zone. The consistency between the segmentation boundary and the direction of the white matter fiber bundles is optimized through adversarial training.

[0043] S6. Dynamic loss control: Construct a multi-objective loss function W_total = αL_dice + βL_edge + γ*L_topology, where α, β, and γ are dynamically adjusted according to the tumor volume ratio, and a boundary complexity index (curvature / fractal dimension) is introduced to adaptively improve the β weight;

[0044] S7. Cascade segmentation strategy: A two-stage cascade training strategy is used. In the first stage, 3D Mask R-CNN is used to locate the tumor core area and generate a placeholder mask. In the second stage, the mask is used as a spatial constraint to segment the motor / language functional area.

[0045] S8. Clinical interactive verification: Deploy a real-time correction module based on a B / S architecture to support doctors in manually adjusting segmentation results through a 3D volume rendering interface and synchronously updating the surgical planning path with the intraoperative navigation system through a DICOM interface.

[0046] A modular pipeline architecture is used, enabling multi-task collaborative processing on the NVIDIA A100 GPU through CUDA stream parallelism. Data preprocessing stage S1 uses the ANTs toolkit for multimodal registration (6-DOF rigid transformation + affine optimization), taking ≤8 seconds per case. The network inference stage (S2-S7) deploys the TensorRT acceleration engine, fusing the Conv-BN-ReLU layers into a single operator, increasing the inference speed to 400 cases per hour. Clinical interaction module S8 develops a 3D editor based on WebGL, supporting the DICOM-RT format and real-time data synchronization with the Brainlab navigation system.

[0047] As a preferred embodiment, the cross-modal fusion layer in step S2 employs a feature channel competition mechanism. Specifically, this includes: performing channel normalization on tumor edema features extracted from T1-weighted and T2-weighted images; generating a channel attention mask based on necrotic core features from FLAIR sequences; and dynamically fusing multimodal features using soft attention weights. Furthermore, a four-level spatial pyramid structure is constructed, consisting of 3×3×3 convolution kernels (with dilation rates of 1, 2, 4, and 6) and a max pooling layer (with stride 2), outputting 64 / 128 / 256 / 512-channel feature maps. The cross-modal fusion layer incorporates a channel competition mechanism: L2 normalization is performed on tumor edema features from T1-weighted images; sigmoid-based soft masks are generated from necrotic core features from FLAIR sequences; and the final fusion weights are dynamically calculated using bilinear interpolation. PyTorch AMP mixed-precision training is used, with feature map storage downgraded to FP16 (occupying 7.2GB of video memory), while key gradients are retained in FP32 (ensuring <0.1% accuracy loss).

[0048] As a preferred embodiment, the deformation field estimation module of step S4 includes a dual verification mechanism: the first verification layer verifies the anatomical continuity of the pyramidal tract and the precentral gyrus after deformation through DTI fiber tracking; the second verification layer compares the functional connectivity strength of the BOLD signal before and after deformation (Pearson correlation coefficient ≥ 0.85). Furthermore, a brain tissue biomechanical model is established based on COMSOL Multiphysics, and 256,000 tetrahedral units (unit size 0.5 mm) are divided into 3), set the Young's modulus of gray matter to 3kPa (Poisson's ratio 0.45), white matter to 7kPa (Poisson's ratio 0.35), and tumor tissue to 15kPa (Poisson's ratio 0.25), and then perform deformation field calculation to achieve functional area displacement compensation with an accuracy of 1.5mm, and the calculation time is compressed from 50s to 8s.

[0049] As a preferred embodiment, the adversarial training strategy of step S5 includes a multi-scale discriminator design: a macro-scale discriminator analyzes the spatial relationship between the overall tumor occupancy and the ventricular system; a meso-scale discriminator detects the distance threshold (≥5mm) between the handknob in the motor area and the tumor boundary; a micro-scale discriminator verifies the gradient continuity of the FA value at the segmentation boundary. Furthermore, a three-layer fully connected neural network is constructed (input layer 3 nodes: tumor volume, fractal dimension, FA gradient; hidden layer 64 nodes; output layer 3 nodes α / β / γ), using Leaky ReLU activation (negative slope 0.01). The fractal dimension is calculated by the box counting method: using 8 levels of grid coverage of 5-256 pixels on a 512×512 slice, linear regression fitting of the log(N)-log(1 / ε) curve, the slope is the fractal dimension (accuracy ±0.02). The dynamic weight is updated once per epoch, and when the tumor volume is greater than 5cm 3 β weight improvement mechanism (0.2→0.5) is automatically triggered.

[0050] As a preferred embodiment, the cascade training of step S7 adopts a knowledge distillation strategy: in the first stage, the teacher network generates a tumor probability heat map; in the second stage, the student network fuses the heat map with the original image to perform functional area segmentation, and then constrains the consistency of the feature distribution of the two stages through the KL divergence loss. Furthermore, the anatomical path uses a pre-trained 3D ResNet-50 (input size 256×256×64), and the resolution of the last layer feature map is 0.5mm 3 , extracting fine structures such as the hippocampus and precentral gyrus. A graph convolutional network (GCN) was constructed based on the functional pathways, with nodes representing the AAL90 brain regions and edge weights defined by the Pearson correlation coefficient of the BOLD signal (threshold > 0.7). An eight-head cross-attention mechanism was used, with queries derived from the gray matter thickness map of the anatomical pathways and keys and values ​​derived from the functional connectivity matrix output by the GCN. Attention scores were normalized using Softmax to generate fused features.

[0051] As a preferred embodiment, step S1 includes a quantification preprocessing of the mass effect, which simulates the brain tissue displacement field caused by tumor mass based on finite element analysis, quantifies the degree of compression and deformation of white matter fiber bundles using tractography, and then establishes a nonlinear regression model (R) of tumor volume-functional area displacement. 2≥0.91), further, the Unity3D physics engine was integrated to simulate tumor growth: brain tissue viscoelastic parameters were set (damping coefficient 0.4, stiffness coefficient 0.8), and the tumor expansion rate was 0.5-2mm 3 / day (obeying log-normal distribution). The GAN network adopts CycleGAN architecture, the generator contains 9 residual blocks, and the discriminator uses PatchGAN to generate 12 types of tumor morphology data (resolution 0.5mm 3 DTI fiber bundle constraints are embedded in the synthetic data (using DSI Studio tracking to retain the directions of the pyramidal fasciculus and arcuate fasciculus) to ensure the anatomical rationality of the enhanced data.

[0052] As a preferred implementation, the interactive correction module in step S8 implements incremental learning, records the doctor's correction trajectory and generates a differential mask, updates the local feature extractor weights through online learning, and establishes a user habit model to predict high-frequency correction areas. Furthermore, Monte Carlo Dropout sampling is performed 50 times to generate a KL divergence heat map, and the interactive design supports click-to-jump to multi-plane views for verification.

[0053] As a preferred embodiment, the final output includes a multi-dimensional clinical report, including a three-dimensional heat map of the volume ratio of the functional area invaded by the tumor, a white matter fiber bundle integrity score, a risk probability distribution of the optimal craniotomy approach, and a machine learning prediction value for the risk of postoperative functional impairment. Furthermore, the three-dimensional rendering uses VTK's GPU ray casting algorithm to set a transparency gradient mapping for the risk area (risk value 0-100% corresponds to a transparency of 0.8-0.2). The quantitative report automatically generates 17 indicators, including: Wernicke zone invasion volume ratio (error <3%, based on Monte Carlo integration calculation); pyramidal tract integrity score (0-5 levels, based on the rate of decrease in DTI anisotropy fraction FA value); surgical approach risk probability (logistic regression model, AUC = 0.91); the data interface is based on FastAPI to develop RESTful services, supporting DICOM-RT format and seamless integration with the StealthStation navigation system.

[0054] Example 1: Multimodal Fusion Segmentation and Deformation Compensation

[0055] Scenario: Left temporal lobe glioma (volume 4.2 cm 3 ) oppression of the Wernicke language area

[0056] Implementation steps:

[0057] 1. Data collection:

[0058] Siemens Prisma 3T MRI was used to acquire T1-MPRAGE images (resolution 0.8 × 0.8 × 1 mm3 ), fMRI language task (TR = 2000ms, voxel 3 × 3 × 3mm 3 ); DTI (b = 1000s / mm 2 , 32 directions)

[0059] 2. Feature Fusion:

[0060] Input 3D feature fusion network (U-Net++ improvement), then perform spatial pyramid pooling, 4-level dilated convolution (expansion rate 1 / 2 / 4 / 6), cross-modal fusion, T1 edema area features (L2 normalized) and fMRI activation map (Z-score>3) through soft attention weight fusion.

[0061] 3. Deformation compensation:

[0062] Finite element model parameters: Young's modulus of tumor 15kPa, surrounding white matter 7kPa; predicted displacement of the posterior temporal lobe 2.8mm (direction: anterolateral); after deformation field correction, the coordinate error of the center of the Wernicke area was reduced from 3.2mm to 0.9mm

[0063] 4. Verification method:

[0064] Intraoperative direct cortical stimulation (ECS) verification: The overlap between the language interruption point and the predicted area was 92% (68% with traditional methods); DTI tracking of the arcuate fasciculus: The FA value decrease rate after deformation was <8% (21% with traditional methods)

[0065] Example 2: Dynamic Weight Adjustment and Large Tumor Segmentation

[0066] Scenario: Glioblastoma on the right frontal lobe (8.7 cm 3 , boundary fractal dimension 1.72);

[0067] Implementation steps:

[0068] 1. Dynamic parameter settings:

[0069] Tumor volume ratio V_tumor / V_brain = 5.3% → Trigger large tumor mode

[0070] Loss weights are automatically adjusted: α = 0.35, β = 0.5, γ = 0.15

[0071] Fractal dimension calculation: box counting method fitted in R on 256×256 slices 2 =0.992.

[0072] 2. Adversarial training optimization:

[0073] Discriminator design:

[0074] Macroscopic evaluation: assessing the distance between the tumor and the anterior horn of the lateral ventricle

[0075] Microscopic scale: Detecting FA gradients at the tumor-motor zone boundary

[0076] After 50,000 iterations of the generator, the Hausdorff distance dropped from 9.8 mm to 5.3 mm.

[0077] 3. Data Augmentation:

[0078] CycleGAN generates 12 types of tumor morphology data (including cystic changes, necrosis, and calcification)

[0079] Physics engine parameters: tumor expansion speed 1.2mm 3 / day, white matter stiffness coefficient 0.7

[0080] After enhancement, the training set was expanded from 150 cases to 980 cases.

[0081] 4. Effect verification:

[0082] The segmentation Dice coefficient was 0.89, and the positioning error of the hand knot boundary in the motor area was 1.3 mm.

[0083] Example 3: Clinical Interaction and Surgical Planning

[0084] Scenario: Meningioma compressing the precentral motor area (volume 5.6cm 3 )

[0085] Implementation steps:

[0086] 1. Uncertainty Visualization:

[0087] Monte Carlo Dropout sampling was performed 50 times, and the KL divergence heat map was calculated. The high-risk area (KL>0.15) accounted for 12% of the tumor volume and was concentrated at the tumor-motor area interface.

[0088] 2. 3D interactive correction:

[0089] The 3D Slicer plug-in was used for correction, and the volume change rate of the precentral gyrus after adjustment was less than 3%.

[0090] 3. Surgical planning output:

[0091] An automatically generated report showed a pyramidal tract integrity score of 4.1 / 5 (FA value reduction rate of 9.2%), and the optimal approach was a parietal-occipital craniotomy (risk probability of 7.8%). DICOM-RT data were pushed to Medtronic StealthStation via the REST API.

[0092] 4. Intraoperative verification:

[0093] Neuronavigation error: 0.9 mm.

[0094] Finally, a few points should be explained: First, in the description of this application, it should be noted that, unless otherwise specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense, and may refer to mechanical or electrical connections, internal communication between two components, or direct connection. "Up," "down," "left," and "right" are only used to indicate relative positional relationships. When the absolute positions of the objects being described change, the relative positional relationships may also change.

[0095] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0096] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based neuroimaging segmentation method for brain functional zoning of tumor-occupied space, characterized by It includes the following steps that work together: S1. Multimodal data fusion acquisition: Structural MRI (T1 / T2 / FLAIR), functional MRI (BOLD signal), and diffusion tensor imaging (DTI) data are acquired through simultaneous registration to establish a three-dimensional voxel-based spatial map that includes tumor mass effect and functional connectivity. S2. Tumor edge feature enhancement: We constructed an improved 3D feature fusion network based on U-Net++, hierarchically extracted edge gradient features of tumor infiltration regions using a multi-scale dilated convolution module (with dilation rates of 2 / 4 / 6), and aligned the heterogeneous textures of T1 / T2 / FLAIR sequences using a cross-modal fusion layer. S3. Collaborative modeling of anatomical and functional features: A dual-path attention mechanism is designed. The anatomical path uses 3D ResNet-50 to extract gray matter / white matter microstructural features, while the functional path uses a graph convolutional network (GCN) to model the functional connectivity topology of the BOLD signal. The two types of features are dynamically fused using a cross-attention matrix. S4. Occupancy deformation compensation: Based on the biomechanical model, an elastic registration sub-network is constructed. Combined with the spatial distribution prior of the functional activation heat map, the vector field generator is used to predict the tumor growth direction and output an adaptive deformation field to correct the spatial displacement of the functional area. S5. Adversarial boundary optimization: A conditional generative adversarial network (cGAN) is integrated, and the discriminator is used to extract the diffusion tensor anisotropy (FA) features of the tumor-functional area transition zone. The consistency between the segmentation boundary and the direction of the white matter fiber bundles is optimized through adversarial training. S6. Dynamic loss control: Construct a multi-objective loss function W_total = αL_dice + βL_edge + γ*L_topology, where α, β, and γ are dynamically adjusted according to the tumor volume ratio, and a boundary complexity index (curvature / fractal dimension) is introduced to adaptively improve the β weight; S7. Cascade segmentation strategy: A two-stage cascade training strategy is used. In the first stage, 3D Mask R-CNN is used to locate the tumor core area and generate a placeholder mask. In the second stage, the mask is used as a spatial constraint to segment the motor / language functional area. S8. Clinical interactive verification: Deploy a real-time correction module based on the B / S architecture to support doctors to manually adjust the segmentation results through the 3D volume rendering interface, and synchronize the surgical planning path with the intraoperative navigation system through the DICOM interface.

2. The method for neural image segmentation of tumor-occupying brain functional regions based on deep learning according to claim 1, characterized in that: The cross-modal fusion layer in step S2 adopts a feature channel competition mechanism, specifically including: channel normalization of the tumor edema area features extracted from the T1 / T2 weighted image; generation of a channel attention mask based on the necrotic core features of the FLAIR sequence; and dynamic fusion of multimodal features through soft attention weights.

3. The method for neural image segmentation of tumor-occupying brain functional regions based on deep learning according to claim 1, characterized in that: The deformation field estimation module in step S4 includes a dual verification mechanism: the first verification layer verifies the anatomical continuity of the pyramidal tract and the precentral gyrus after deformation through DTI fiber tracking; the second verification layer compares the functional connectivity strength of the BOLD signal before and after deformation (Pearson correlation coefficient ≥ 0.85).

4. The method for neural image segmentation of tumor-occupying brain functional regions based on deep learning according to claim 1, characterized in that: The adversarial training strategy in step S5 includes the design of a multi-scale discriminator: a macro-scale discriminator analyzes the spatial relationship between the overall tumor occupancy and the ventricular system; a meso-scale discriminator detects the distance threshold (≥5 mm) between the hand knob in the motor area and the tumor boundary; and a micro-scale discriminator verifies the gradient continuity of the FA value at the segmentation boundary.

5. The method for neural image segmentation of tumor-occupying brain functional regions based on deep learning according to claim 1, characterized in that: The cascade training in step S7 adopts a knowledge distillation strategy: in the first stage, the teacher network generates a tumor probability heat map; in the second stage, the student network fuses the heat map with the original image to perform functional area segmentation, and then the KL divergence loss is used to constrain the consistency of the feature distribution of the two stages.

6. The method for neural image segmentation of tumor-occupying brain functional regions based on deep learning according to claim 1, characterized in that: Step S1 includes the quantification preprocessing of the mass effect, which simulates the brain tissue displacement field caused by tumor mass based on finite element analysis, quantifies the degree of compression and deformation of white matter fiber bundles using tractography, and then establishes a nonlinear regression model of tumor volume-functional area displacement (R 2 ≥0.91).

7. The method for neural image segmentation of tumor-occupying brain functional regions based on deep learning according to claim 1, characterized in that: The interactive correction module in step S8 implements incremental learning, records the doctor's correction trajectory and generates a differential mask, updates the local feature extractor weights through online learning, and establishes a user habit model to predict high-frequency correction areas.

8. The method for neural image segmentation of tumor-occupying brain functional regions based on deep learning according to claim 1, characterized in that: The final output includes a multi-dimensional clinical report, including a three-dimensional heat map of the volume ratio of the functional area invaded by the tumor, the white matter fiber bundle integrity score, the risk probability distribution of the optimal craniotomy approach, and the machine learning prediction value of the risk of postoperative functional impairment.

Images