Brain glioma boundary recognition system based on nuclear magnetic images

Abstract

The invention relates to the technical field of nuclear magnetic image auxiliary identification, and discloses a brain glioma boundary identification system based on nuclear magnetic images, which comprises a data input module, the data input module establishes a database based on a plurality of multi-modal MRI data, and normalization processing is performed on the multi-modal MRI data in the database through a data standardization processing module. According to the method, features can be extracted in x, y and z axial directions at the same time by using a 3D convolution kernel, a complete three-dimensional structure of the brain glioma is obtained, a continuous space structure of the tumor is completely captured, one or more 3D bounding boxes can be rapidly output for any input standardized multi-mode MRI through a positioning module, all tumor areas are wrapped with high confidence, and the accuracy of the three-dimensional structure of the brain glioma is improved. Therefore, the identification time of a single sample is shortened from dozens of minutes to a second level, and the boundary of the infiltration area can be clearly and efficiently distinguished through the cascade structure design of positioning and the core area and the infiltration area.

Description

Technical Field

[0001] The present invention relates to the field of nuclear magnetic resonance imaging-assisted recognition technology, and in particular to a brain glioma boundary recognition system based on nuclear magnetic resonance imaging. Background Art

[0002] Gliomas are primary tumors of the central nervous system that originate from glial cells, accounting for 27% of all brain tumors and over 80% of malignant brain tumors. Gliomas are characterized by infiltrative growth, with tumor cells spreading along nerve fiber bundles and lacking a clear boundary with normal brain tissue. 90% of postoperative recurrences occur within 2 cm of the original lesion. Currently, magnetic resonance imaging is the preferred diagnostic method in clinical practice. T1-weighted images show low signal intensity, while T2 / FLAIR images show high signal intensity. High-grade gliomas exhibit irregular enhancement due to disruption of the blood-brain barrier, with significant surrounding edema. Traditional manual detection is insensitive to microinfiltration lesions, time-consuming, and has inconsistent clinical standards. The volume of tumors manually outlined by different doctors can vary by up to 4.1 mL. Artificial intelligence has demonstrated excellent performance in the field of medical imaging, providing a powerful tool for boundary identification of gliomas. In particular, convolutional neural networks, represented by the U-Net architecture, have made great progress in the automatic segmentation of glioma MRI images. Advanced AI models are now able to achieve high-precision tumor subregion segmentation on preoperative MRI, with performance indicators even exceeding 0.8 or 0.9, approaching or in some aspects surpassing the level of human experts.

[0003] Most existing models are based on two-dimensional slice image analysis and are unable to reconstruct the three-dimensional infiltration trajectory of the tumor. Most traditional CNN models process 3D MRI data slice by slice in a 2D manner, completely ignoring the spatial contextual information between slices. Glioma is a complete three-dimensional structure, and its invasive boundary changes continuously in three-dimensional space. 2D models cannot capture this cross-layer continuity, resulting in segmentation results that may appear discontinuous and incoherent after three-dimensional reconstruction. This is especially true in areas where the tumor boundary morphology is complex and changes dramatically, where the segmentation accuracy will be significantly reduced. Summary of the Invention

[0004] The present invention provides a brain glioma boundary recognition system based on nuclear magnetic resonance imaging, which has the beneficial effect of clearly and efficiently distinguishing the boundaries of infiltration areas.

[0005] The present invention provides the following technical solution: a brain glioma boundary recognition system based on nuclear magnetic resonance imaging, comprising: A data input module, which establishes a database based on several multimodal MRI data; A data normalization processing module is used to normalize the multimodal MRI data in the database and perform spatial registration on the normalized multimodal MRI data to obtain multimodal MRI data in the same format and coordinate system; Multimodal data alignment module: The multimodal data alignment module aligns and resamples the different modal MRI images of the same patient to be identified, thereby obtaining a multimodal dataset of the same patient; A positioning module, which identifies and obtains a three-dimensional bounding box based on a multimodal dataset of the same patient. The three-dimensional bounding box includes several sub-image blocks of standardized sizes; A multi-task learning framework, comprising a core region boundary recognition module and an infiltration region boundary recognition module, is used to output range parameters of the enhanced tumor region, the necrotic / non-enhancing core region, and the infiltration region; The result verification module summarizes the range parameters of the obtained sub-image blocks, enhanced tumor areas, necrotic / non-enhanced core areas, and infiltration areas and combines them with the multimodal MRI data matching evaluation in the same format and coordinate system to output the recognition results.

[0006] As an optional solution of the MRI-based glioma boundary recognition system of the present invention, the core region boundary recognition module accurately segments sub-image blocks using a 3D U-Net network architecture and a hybrid loss function. The sub-image blocks consist of enhanced tumor areas and necrotic / non-enhancing core areas. The invasive area boundary recognition module outputs a complete segmentation mask of the entire tumor area in the sub-image block, and subtracts the enhanced tumor area and necrotic / non-enhanced core area masks from the entire tumor area to obtain the range parameters of the invasive area.

[0007] As an optional solution of the brain glioma boundary recognition system based on nuclear magnetic resonance imaging of the present invention, MRI image registration includes: Set the reference modality and align all MRI images of the patient using the registration tool based on the parameters of the reference modality; Resampling involves resampling the registered MRI images of all modalities of the same patient to the voxel grid and spatial resolution of the reference modality to obtain a multimodal dataset of the same patient.

[0008] As an optional solution of the brain glioma boundary recognition system based on nuclear magnetic resonance imaging of the present invention, wherein: the registration tool uses a rigid transformation or an affine transformation with at least six degrees of freedom to register all modal MRI images of the same patient; Any voxel coordinate in a multimodal dataset corresponds precisely to the same physical location in the brain across all modality channels.

[0009] As an optional solution of the brain glioma boundary recognition system based on nuclear magnetic resonance imaging of the present invention, the positioning module includes a 3DCNN model, which recognizes a multimodal data set in the axial orthogonal plane, the coronal orthogonal plane, and the sagittal orthogonal plane to obtain a three-dimensional bounding box; The sub-image blocks are adjusted to a uniform size by padding or interpolation.

[0010] As an optional solution of the brain glioma boundary recognition system based on nuclear magnetic resonance imaging of the present invention, wherein: the three-dimensional bounding box includes the coordinates of the front point of the upper left corner and the coordinates of the back point of the lower right corner; A safety margin value is set. When the 3DCNN model recognizes a multimodal dataset, the safety margin value is expanded outward on the obtained 3D bounding box to optimize the 3D bounding box. The optimized 3D bounding box ensures that the edge infiltration area is included.

[0011] As an optional solution of the brain glioma boundary recognition system based on nuclear magnetic resonance imaging described in the present invention, it also includes setting an IoU value, and the IoU value of the overlap degree between the three-dimensional bounding box obtained by the positioning module and the real labeled bounding box is greater than the set IoU value.

[0012] As an optional solution of the brain glioma boundary recognition system based on nuclear magnetic resonance imaging of the present invention, the core area boundary recognition module also includes a residual Transformer module; The multi-scale features of multimodal datasets are integrated through the encoder structure, decoder structure and skip connection of 3D U-Net; The Transformer module is used to capture long-range dependencies on the multi-scale features of the multimodal dataset, so as to obtain the overall morphology of the tumor and identify the infiltration boundaries away from the core area, thereby assisting the hybrid loss function to optimize the three-dimensional bounding box and output sub-image blocks.

[0013] As an optional solution to the MRI-based glioma boundary recognition system described in the present invention, the core area boundary recognition module and the infiltration area boundary recognition module in the multi-task learning framework share features, and the position and morphological information of the necrotic / non-enhanced core area are used to assist in inferring the range parameters of the infiltration area.

[0014] As an optional solution to the brain glioma boundary recognition system based on nuclear magnetic resonance imaging described in the present invention, the infiltration area boundary recognition module quantitatively evaluates the range parameters of the infiltration area through the Dice similarity coefficient and 95% Hausdorff distance to verify whether the range parameters of the infiltration area are accurate.

[0015] The Dice similarity coefficient was used to evaluate the degree of regional overlap; The 95% Hausdorff distance was used to evaluate the boundary matching.

[0016] The present invention has the following beneficial effects: 1. This MRI-based glioma boundary recognition system, by using a 3D convolution kernel, can simultaneously extract features in the x, y, and z axes to obtain the complete three-dimensional structure of the glioma and fully capture the continuous spatial structure of the tumor.

[0017] 2. This MRI-based glioma boundary recognition system, through the positioning module, can quickly output one or more 3D bounding boxes for any input standardized multimodal MRI, enclosing all tumor areas with high confidence, thereby reducing the recognition time of a single sample from tens of minutes to seconds.

[0018] 3. This MRI-based brain glioma boundary recognition system can clearly and efficiently distinguish the boundaries of the infiltration area through the cascade structure design of positioning, core area and infiltration area. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the overall system framework diagram of the present invention.

[0020] Figure 2 Schematic diagram of the present invention.

[0021] Figure 3 This is a schematic diagram of the marked area of ​​the present invention.

[0022] Figure 4 Schematic diagram of the AUC curve of the recognition performance of the present invention.

[0023] Figure 5 Schematic diagram for comparing Dice coefficients of the present invention.

[0024] Figure 6 Schematic diagram of Hausdorff distance comparison of the present invention. DETAILED DESCRIPTION

[0025] 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.

[0026] Example 1: Please refer to Figures 1-6 , one of the brain glioma boundary recognition systems based on nuclear magnetic resonance imaging includes: A data input module, which establishes a database based on several multimodal MRI data; Several multimodal MRI data sets including annotated multimodal MRI data of patients with glioma; Each patient's data contains the following four core sequences, which are annotated at the pixel level by senior neuroradiologists. Figure 2 ; T1-weighted imaging (T1w): provides excellent anatomical details; T2-weighted imaging (T2w): sensitive to edema and non-enhancing tumor areas; Fluid-attenuated inversion recovery (FLAIR) sequence: can suppress cerebrospinal fluid signals and make the peritumoral edema area more clearly visible; T1 contrast-enhanced imaging (T1ce): By injecting contrast agent, it can highlight areas where the blood-brain barrier is disrupted, usually corresponding to the active core of the tumor; The annotation of the four core sequence data follows the internationally accepted BraTS (Brain Tumor Segmentation Challenge) standard, dividing the tumor regions into three subcategories: necrotic and non-enhancing tumor core areas; Enhanced tumor area; Peritumoral edema area; The annotations of necrotic and non-enhancing tumor core areas, enhancing tumor areas, and peritumoral edema areas will together constitute the label of the “whole tumor area”.

[0027] To this end, this study has established a structured database to efficiently store and index thousands of patients’ original images, pre-processed images, annotated data, and model outputs to ensure the reproducibility of the research; It should be noted that the database uses the BIDS-BrainImagingDataStructure format; After completing this module, a large-scale, quality-controlled, and structured multimodal MRI dataset of gliomas is obtained. All data will be organized into a unified format (NIfTI) and accompanied by clear metadata, providing a high-quality, reliable data source for subsequent standardized processing and model training.

[0028] A data normalization processing module is used to normalize the multimodal MRI data in the database and perform spatial registration on the normalized multimodal MRI data to obtain multimodal MRI data in the same format and coordinate system; Among them, the data standardization processing module is the key to eliminating data heterogeneity and ensuring the generalization ability of the model. Specifically: Since different scanning devices and parameters will lead to inconsistent image intensity (grayscale value) ranges, normalization is necessary. Therefore, normalization is required, which includes N4ITK bias field correction and Z-Score normalization: Among them, N4ITK bias field correction: MRI images often have low-frequency intensity drift (bias field) caused by magnetic field inhomogeneity, which seriously affects quantitative analysis. The N4 algorithm is the gold standard for correcting this artifact. This function is implemented based on open source libraries such as SimpleITK. According to the successful experience of the BraTS challenge and the configuration of key parameters in related research, this study systematically tests and selects the optimal parameter combination. The initial configuration will refer to: the number of iterations is set to [50,50,50,50], the shrinkage factor is 2, and the B-spline fitting order is 3. For the B-spline grid resolution, instead of setting a fixed value directly, it is made adaptive to the image size by setting the control point distance. The initial value can be set to 200mm and gradually reduced at multiple resolution levels; Among them, Z-Score normalization: After completing the bias field correction, for each modality image of each patient, this study only focuses on the brain parenchyma area and performs Z-Score normalization through a preliminary brain mask extraction.

[0029] To ensure that all patients' brain images are in a common coordinate system for easy model learning and comparison, the image data are registered to a standard space, and all patients' T1w images are linearly (affine transformed) registered to a standard anatomical template, such as the MNI152 template.

[0030] It is important to note that the antsRegistration tool in the industry-leading ANTs (Advanced Normalization Tools) software package was used because its SyN algorithm performs excellently in nonlinear registration; The output of the data normalization module is a fully normalized multimodal MRI dataset. All MRI sequences for each patient have similar intensity distributions and are registered to the same three-dimensional coordinate space. Specifically, the output data is a four-dimensional tensor of [Channels, Depth, Height, Width], where Channels corresponds to different modalities such as T1, T2, FLAIR, and T1ce. This is the ideal format for input to subsequent CNN models.

[0031] Multimodal data alignment module: The multimodal data alignment module aligns and resamples the different modal MRI images of the same patient to be identified, thereby obtaining a multimodal dataset of the same patient; MRI image registration includes: Set the reference modality and align all MRI images of the patient using the registration tool based on the parameters of the reference modality; Resampling involves resampling the registered MRI images of all modalities of the same patient to the voxel grid and spatial resolution of the reference modality to obtain a multimodal dataset of the same patient; The registration tool uses a rigid transformation or an affine transformation with at least six degrees of freedom to register all modality MRI images of the same patient; Any voxel coordinate in a multimodal dataset corresponds precisely to the same physical location in the brain across all modality channels.

[0032] Specifically, the multimodal data alignment module ensures accurate voxel-level alignment between MRI images of different modalities for the same patient. Although in modern clinical scanning, multimodal sequences are often completed in the same scan, with minimal motion artifacts, slight head movements or distortions between sequences may still exist. Therefore, intra-modality rigid registration is required, including the registration process, technology selection, and resampling. Registration process: In this study, the T1ce modality with the clearest anatomical structure was selected as the reference modality. All other modalities of the patient (T1w, T2w, FLAIR, etc.) were then registered to this reference modality.

[0033] Technology Selection: This approach also utilizes registration tools from ANTs or FSL (FMRIB Software Library), such as FLIRT / FNIRT. Because registration is performed between modalities within the same patient, a rigid or affine transformation with 6 or 7 degrees of freedom can typically achieve high accuracy.

[0034] Resampling: After registration, all modalities need to be resampled to the same voxel grid and spatial resolution as the reference modality. In this study, resampling to an isotropic resolution of 1mm*1mm*1mm is crucial for the 3D CNN model to handle all three spatial dimensions fairly.

[0035] After passing through this module, a perfectly aligned and resampled multimodal dataset is output. Now, any voxel coordinate (i, j, k) in the dataset accurately corresponds to the same physical location in the brain across all modal channels. This allows the 3D CNN model to simultaneously analyze feature vectors from T1, T2, FLAIR, and T1ce at each location, achieving true multimodal feature fusion. This allows the use of 3D convolution kernels to simultaneously extract features in the x, y, and z axes, obtaining the complete three-dimensional structure of the brain glioma and fully capturing the continuous spatial structure of the tumor.

[0036] A positioning module, which identifies and obtains a three-dimensional bounding box based on a multimodal dataset of the same patient. The three-dimensional bounding box includes several sub-image blocks of standardized sizes; The localization module includes a 3DCNN model that recognizes a multimodal dataset in the axial, coronal, and sagittal orthogonal planes to obtain a 3D bounding box. Adjust the sub-image blocks to a uniform size by padding or interpolation; The 3D bounding box includes the coordinates of the upper left front point and the lower right back point; A safety margin value is set. When the 3DCNN model recognizes a multimodal dataset, the safety margin value is expanded outward on the obtained 3D bounding box to optimize the 3D bounding box. The optimized 3D bounding box ensures that the edge infiltration area is included.

[0037] Furthermore, to improve computational efficiency and segmentation accuracy, this study adopted a two-stage strategy: first, quickly locate the approximate region of interest (ROI) of the tumor, and then perform fine boundary segmentation within this ROI.

[0038] The goals at this stage are speed and recall, not pixel-level accuracy. Therefore, this study employed a computationally less expensive 3D CNN model (a variant of Faster R-CNN) to identify the tumor's bounding box on three orthogonal planes (axial, coronal, and sagittal). Considering the completeness of the 3D information, a lightweight, simplified 3D U-Net is preferred to directly predict a rough tumor bounding box on the downsampled 3D image.

[0039] The output of the localization module is a 3D bounding box defined by the coordinates of its upper left corner (x_min, y_min, z_min) and its lower right corner (x_max, y_max, z_max). This bounding box should encompass the entire tumor region (WT). To ensure that the marginal invasive area is not missed, a safety margin of 20 voxels is added to the initially predicted bounding box.

[0040] The output of the localization module (bounding box coordinates) directly serves as the input to the subsequent boundary recognition module. The boundary recognition module no longer processes the entire brain image, but only the sub-volume cropped by the bounding box. This "coarse-to-fine" strategy has three major advantages: Reduced computational burden: The fine segmentation model only needs to process a much smaller region containing the tumor, significantly reducing computational effort and memory consumption; Reduced class imbalance: In the cropped image patches, the ratio of tumor voxels to non-tumor voxels is improved, alleviating the class imbalance problem dominated by background voxels, helping the model better learn tumor features; Standardized input size: All cropped sub-image blocks can be adjusted to a uniform size (such as 128x128x128) by padding or interpolation to facilitate batch training; For any input standardized multimodal MRI, the localization module can quickly (in milliseconds) output one or more 3D bounding boxes, enclosing all tumor areas with high confidence, thereby reducing the recognition time of a single sample from tens of minutes to seconds.

[0041] It also includes setting an IoU value, and the IoU value of the overlap between the three-dimensional bounding box obtained by the positioning module and the real labeled bounding box is greater than the set IoU value.

[0042] It should be noted that the accuracy of model recognition is as follows Figure 3 As shown, the instance data is as follows Figure 4 As shown in the figure, the localization model was validated on the validation set using the Intersection over Union (IoU) to assess the overlap between the predicted bounding box and the ground-truth bounding box. For more than 99% of cases, the IoU between the true tumor region and the predicted bounding box should be greater than 0.8, and the true tumor region should be completely contained within the predicted bounding box. Therefore, this study validated all samples, and those that met these requirements were considered accurate.

[0043] Multi-task learning framework, which includes a core region boundary recognition module and an invasive region boundary recognition module. The multi-task learning framework is used to output the range parameters of the enhanced tumor area, necrotic / non-enhancing core area, and invasive area; The core area boundary recognition module and the invasive area boundary recognition module in the multi-task learning framework share features and use the location and morphology information of the necrotic / non-enhancing core area to assist in inferring the range parameters of the invasive area. The core area boundary recognition module uses a 3D U-Net network architecture and a hybrid loss function to accurately segment sub-image blocks, which are composed of enhanced tumor areas and necrotic / non-enhancing core areas. The core area boundary recognition module also includes a residual Transformer module; The multi-scale features of multimodal datasets are integrated through the encoder structure, decoder structure and skip connection of 3D U-Net; The Transformer module is used to capture long-range dependencies on the multi-scale features of the multimodal dataset, so as to obtain the overall morphology of the tumor and identify the infiltration boundaries away from the core area, thereby assisting the hybrid loss function to optimize the three-dimensional bounding box and output sub-image blocks.

[0044] Specifically, the core region boundary recognition module is responsible for accurately segmenting the tumor core region, namely the enhancing tumor region (ET) and the necrotic / non-enhancing core region (NCR / NET), within the ROI provided by the localization module. The input to this module is a multimodal sub-image block that has been localized, cropped, and size-normalized.

[0045] The advanced network architecture 3D U-Net. This is one of the most classic and powerful models in the field of medical image segmentation. Its encoder-decoder structure and skip connections can effectively fuse multi-scale features, capturing global context while preserving local details. The subsequent introduction of residual modules (such as DeepMedic or dResU-Net) can alleviate the gradient vanishing problem in deep networks, allowing the construction of deeper and more expressive models. Finally, the integration of self-attention mechanisms, especially the introduction of Transformer modules (such as TransBTS, BiTr-Unet, and 3D UNet+CoT). Transformers excel at capturing long-range dependencies, which has great potential for understanding the overall morphology of tumors and identifying infiltration boundaries far away from the core area.

[0046] In terms of loss function selection, due to the large size differences among tumor subregions (ET, NCR / NET), class imbalance still exists. This study uses a hybrid loss function: Dice Loss + Focal Loss. Dice Loss directly optimizes the Dice coefficient, an evaluation metric for the segmentation task, and is robust to class imbalance. Focal Loss, on the other hand, focuses the model on difficult-to-separate samples (typically pixels at boundaries), improving the sharpness of boundary segmentation.

[0047] The core region boundary identification module then outputs a pixel-level segmentation mask of the same size as the input sub-image block, in which each voxel is labeled as one of background, enhancing tumor (ET), or necrotic / non-enhancing core region (NCR / NET).

[0048] The core area boundary identification module will be verified using the following indicators, using the Dice similarity coefficient (Dice Coefficient) and the 95% Hausdorff distance (95% Hausdorff Distance) for quantitative evaluation.

[0049] The invasive area boundary recognition module outputs a complete segmentation mask of the entire tumor area in the sub-image block, and subtracts the enhanced tumor area and necrotic / non-enhanced core area masks from the entire tumor area to obtain the range parameters of the invasive area; The invasive area boundary recognition module quantitatively evaluates the range parameters of the invasive area using the Dice similarity coefficient and 95% Hausdorff distance to verify whether the range parameters of the invasive area are accurate.

[0050] The Dice similarity coefficient was used to evaluate the degree of regional overlap; The 95% Hausdorff distance was used to evaluate the boundary matching.

[0051] The core area boundary recognition module is used to identify areas where tumor cells have infiltrated but have not yet breached the blood-brain barrier, namely, peritumoral edema (ED) areas with high signals on FLAIR sequences. Identification of infiltrated areas relies heavily on detailed analysis of signal abnormalities on FLAIR and T2 sequences, while also combining T1 and T1ce information to exclude core areas. Specifically, the invasive zone boundary identification module and the core zone identification module can be integrated into a multi-task learning framework, with the same 3D CNN model simultaneously outputting segmentation results for the three subregions (ET, NCR / NET, and ED). This design allows for internal feature sharing within the model, leveraging the core zone's position and morphology to aid in inferring the extent of the invasive zone.

[0052] Boundary-aware loss is a technical challenge. The boundaries of invasive regions are extremely fuzzy, making them typical "difficult-to-classify" examples. This study introduces a boundary-specific loss function that directly penalizes misclassified boundary pixels and increases the weight of boundary voxels in the loss calculation. This forces the model to invest more effort in learning and fitting fuzzy boundaries. The invasive region boundary identification module is expected to output a complete "whole tumor region" (WT) segmentation mask within the ROI. The extent of the invasive region (ED) is obtained by subtracting the core region (ET + NCR / NET) mask from the WT mask. The invasive region boundary identification module will be validated using the following metrics: Dice similarity coefficient and 95% Hausdorff distance for quantitative evaluation.

[0053] In summary, through the cascade structure design of positioning, core area and infiltration area, the boundary of the infiltration area can be clearly and efficiently distinguished.

[0054] The result verification module summarizes the range parameters of the obtained sub-image blocks, enhanced tumor areas, necrotic / non-enhanced core areas, and infiltration areas and combines them with the multimodal MRI data matching evaluation in the same format and coordinate system to output the recognition results.

[0055] The results verification module integrates the previous six modules. For any dataset, it performs the following steps: "data input," "data normalization," "multimodal data alignment," "tumor localization," "core region boundary identification," and "infiltration region boundary identification." The results are then compared with the annotated data to verify the recognition effectiveness. For clinical data, the model system directly outputs the identified glioma boundaries for clinicians' reference.

[0056] The results validation module validated the model's performance using an independent dataset of 200 samples. The model's performance was verified based on the annotation results. This module used the following metrics: sensitivity (recall) and specificity (specificity) to assess the model's ability to detect tumor voxels and correctly identify background voxels. The Dice similarity coefficient (which assesses regional overlap) and Hausdorff distance (which assesses boundary matching) are the two primary metrics for evaluating segmentation performance. This study calculated metrics for the three labels: the enhancement region (ET), the core region (ET+NCR / NET), and the total tumor region (WT). This is crucial for assessing clinical risk, such as missed tumors or over-resection.

[0057] This application proposes an end-to-end 3D CNN-based preoperative boundary recognition system for gliomas, border-3D CNN. Through a modular design, it gradually addresses challenges throughout the entire process, from data preparation to model training and result verification. This results in an intelligent segmentation tool with clinical-grade accuracy and robustness, providing strong technical support for precise surgical treatment of gliomas.

[0058] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0059] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A glioma boundary recognition system based on magnetic resonance imaging, characterized in that: include: A data input module, wherein the data input module establishes a database based on a plurality of multimodal MRI data; a data normalization processing module, which performs normalization processing on the multimodal MRI data in the database and performs same-space registration on the normalized multimodal MRI data, thereby obtaining multimodal MRI data in the same format and coordinate system; A multimodal data alignment module, which aligns and resamples the different modal MRI images of the same patient to be identified, thereby obtaining a multimodal dataset of the same patient; A positioning module, wherein the positioning module identifies and obtains a three-dimensional bounding box based on a multimodal dataset of the same patient, wherein the three-dimensional bounding box includes a plurality of sub-image blocks of standardized sizes; A multi-task learning framework, comprising a core region boundary recognition module and an infiltration region boundary recognition module, configured to output range parameters for an enhancing tumor region, a necrotic / non-enhancing core region, and an infiltration region; The result verification module summarizes the range parameters of the obtained sub-image blocks, enhanced tumor areas, necrotic / non-enhanced core areas, and infiltration areas and combines them with the multimodal MRI data matching evaluation in the same format and coordinate system to output the recognition results.

2. The MRI-based brain glioma boundary recognition system according to claim 1, characterized in that: The core area boundary recognition module accurately segments the sub-image block by using a 3D U-Net network architecture and a hybrid loss function, wherein the sub-image block consists of the enhanced tumor area and the necrotic / non-enhanced core area; The infiltration area boundary recognition module outputs a complete segmentation mask of the entire tumor area in the sub-image block, and subtracts the enhanced tumor area and necrotic / non-enhanced core area masks from the entire tumor area to obtain the range parameters of the infiltration area.

3. The MRI-based brain glioma boundary recognition system according to claim 1, characterized in that: The MRI image registration includes: Setting a reference modality, and registering all modality MRI images of the patient using a registration tool based on the parameters of the reference modality; The resampling includes resampling the registered MRI images of all modalities of the same patient to the voxel grid and spatial resolution of the reference modality, thereby obtaining a multimodal dataset of the same patient.

4. The MRI-based brain glioma boundary recognition system according to claim 3, characterized in that: The registration tool registers all modality MRI images of the same patient using a rigid transformation or an affine transformation with at least six degrees of freedom; Any voxel coordinate of the multimodal dataset accurately corresponds to the same physical position in the brain across all modality channels.

5. The MRI-based brain glioma boundary recognition system according to claim 2, characterized in that: The positioning module includes a 3DCNN model, and the 3DCNN model is used to identify the multimodal data set in the axial orthogonal plane, the coronal orthogonal plane, and the sagittal orthogonal plane to obtain the three-dimensional bounding box; The sub-image blocks are adjusted to a uniform size by filling or interpolation.

6. The MRI-based brain glioma boundary recognition system according to claim 5, characterized in that: The three-dimensional bounding box includes the coordinates of the upper left front point and the lower right back point; A safety margin value is set. When the 3DCNN model recognizes a multimodal dataset, the safety margin value is expanded outward on the obtained three-dimensional bounding box to optimize the three-dimensional bounding box. The optimized three-dimensional bounding box ensures that the edge infiltration area is included.

7. The MRI-based brain glioma boundary recognition system according to claim 6, characterized in that: It also includes setting an IoU value, where the IoU value of the overlap between the three-dimensional bounding box obtained by the positioning module and the real labeled bounding box is greater than the set IoU value.

8. The MRI-based brain glioma boundary recognition system according to claim 6, characterized in that: The core area boundary recognition module also includes a Transformer module; The multi-scale features of the multimodal dataset are fused through the encoder structure, decoder structure and skip connection of the 3D U-Net; The Transformer module is used to capture long-range dependencies on the multi-scale features of the multimodal dataset, so as to obtain the overall morphology of the tumor and identify the infiltration boundaries away from the core area, thereby assisting the hybrid loss function to optimize the three-dimensional bounding box and output sub-image blocks.

9. The MRI-based brain glioma boundary recognition system according to claim 1, characterized in that: The core area boundary recognition module and the infiltration area boundary recognition module in the multi-task learning framework share features and assist in inferring the range parameters of the infiltration area by utilizing the position and morphological information of the necrotic / non-enhanced core area.

10. The MRI-based brain glioma boundary recognition system according to claim 6, characterized in that: The infiltration area boundary recognition module quantitatively evaluates the range parameters of the infiltration area using the Dice similarity coefficient and the 95% Hausdorff distance to verify whether the range parameters of the infiltration area are accurate; The Dice similarity coefficient is used to evaluate the region overlap; The 95% Hausdorff distance is used to evaluate the boundary matching.

Images