A glioma boundary recognition system based on MRI images

Abstract

This invention relates to the field of MRI image-assisted recognition technology and discloses a glioma boundary recognition system based on MRI images. The system includes a data input module that establishes a database based on several multimodal MRI datasets. A data standardization module normalizes the multimodal MRI data within the database. By using 3D convolutional kernels, this invention can simultaneously extract features along the x, y, and z axes to obtain the complete three-dimensional structure of the glioma, fully capturing the continuous spatial structure of the tumor. Through a localization module, for any input standardized multimodal MRI, it can quickly output one or more 3D bounding boxes to encompass all tumor regions with high confidence. This reduces the recognition time for a single sample from tens of minutes to seconds. Through a cascaded structure design of localization, core area, and infiltrative area, the boundaries of the infiltrative area can be clearly and efficiently distinguished.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging-assisted recognition technology, specifically to a glioma boundary recognition system based on magnetic resonance imaging. Background Technology

[0002] Gliomas are primary tumors of the central nervous system originating 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 clear boundaries with normal brain tissue. Postoperative recurrence occurs within 2 cm of the original lesion in 90% of cases. Currently, MRI 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 blood-brain barrier disruption, and significant surrounding edema is present. Traditional manual detection methods lack sensitivity for microinvasive lesions, are time-consuming, and have inconsistent clinical standards; the difference in tumor volume delineated manually by different doctors can reach 4.1 mL. Artificial intelligence has demonstrated superior performance in medical imaging, providing a powerful tool for glioma boundary identification. In particular, convolutional neural networks, represented by the U-Net architecture, have made great strides in the automatic segmentation of glioma MRI images. Advanced AI models can now achieve high-precision tumor subregion segmentation on preoperative MRI, with performance indicators even exceeding 0.8 or 0.9, approaching or surpassing the level of human experts in some aspects.

[0003] Most existing models are based on two-dimensional slice image analysis and cannot reconstruct the three-dimensional invasion trajectory of tumors. Most traditional CNN models process 3D MRI data slice by slice in a 2D manner, completely ignoring the spatial context 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, which may lead to discontinuities and inconsistencies in the segmentation results after three-dimensional reconstruction. Especially in areas where the tumor boundary morphology is complex and changes drastically, the segmentation accuracy will decrease significantly. Summary of the Invention

[0004] This invention provides a glioma boundary recognition system based on MRI imaging, which has the beneficial effect of clearly and efficiently distinguishing the boundaries of infiltrated areas.

[0005] This invention provides the following technical solution: a glioma boundary recognition system based on MRI imaging, comprising:

[0006] The data input module establishes a database based on several multimodal MRI data sets.

[0007] The data standardization processing module normalizes the multimodal MRI data in the database and performs cospatial registration on the normalized multimodal MRI data to obtain multimodal MRI data in the same format and coordinate system.

[0008] The multimodal data alignment module registers and resamples different modal MRI images of the same patient that need to be identified, thereby obtaining a multimodal dataset of the same patient.

[0009] The localization module identifies and obtains a three-dimensional bounding box based on the multimodal dataset of the same patient. The three-dimensional bounding box includes several sub-image blocks of standardized size.

[0010] 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-enhanced core region, and the infiltration region.

[0011] The results verification module summarizes the range parameters of the obtained sub-image blocks, enhanced tumor areas, necrotic / non-enhanced core areas, and infiltrative areas, and combines them with multimodal MRI data of the same format and coordinate system for matching and evaluation, thereby outputting the recognition results.

[0012] As an optional solution of the glioma boundary recognition system based on MRI images described in this invention, the core area 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-enhanced core areas.

[0013] The infiltrative region boundary recognition module outputs a complete full tumor region segmentation mask within a sub-image block, and subtracts the enhanced tumor region and necrotic / non-enhanced core region masks from the full tumor region to obtain the range parameters of the infiltrative region.

[0014] As an optional embodiment of the glioma boundary recognition system based on MRI imaging described in this invention, the MRI image registration includes:

[0015] A reference modality is set, and all modal MRI images of the patient are registered using a registration tool based on the parameters of the reference modality;

[0016] Resampling involves resampling all registered MRI images of the same patient using a voxel grid and spatial resolution of the reference modality to obtain a multimodal dataset of the same patient.

[0017] As an optional solution to the glioma boundary recognition system based on MRI images described in this invention, the registration tool uses a rigid transformation or affine transformation with at least six degrees of freedom to register all modal MRI images of the same patient.

[0018] Any voxel coordinate in the multimodal dataset corresponds precisely to the same physical location in the brain across all modal channels.

[0019] As an optional solution of the glioma boundary recognition system based on MRI images described in this invention, the localization module includes a 3DCNN model, which identifies multimodal datasets in the axial orthogonal plane, the coronal orthogonal plane, and the sagittal orthogonal plane to obtain a three-dimensional bounding box.

[0020] Adjust sub-image blocks to a uniform size by padding or interpolation.

[0021] As an optional solution to the glioma boundary recognition system based on MRI images described in this invention, the three-dimensional bounding box includes the coordinates of the upper left front point and the lower right back point.

[0022] Set a safety margin value. When the 3DCNN model recognizes a multimodal dataset, expand the safety margin value outward on the acquired 3D bounding box to optimize the 3D bounding box. The optimized 3D bounding box ensures that the edge immersion area is included.

[0023] As an optional solution to the glioma boundary recognition system based on MRI images described in this invention, it further includes setting an IoU value, wherein the IoU value of the overlap between the three-dimensional bounding box obtained by the localization module and the actual labeled bounding box is greater than the set IoU value.

[0024] As an optional solution to the glioma boundary recognition system based on MRI images described in this invention, the core area boundary recognition module further includes a residual Transformer module;

[0025] Multi-scale features of a multimodal dataset are fused using the encoder structure, decoder structure, and skip connections of a 3D U-Net.

[0026] The Transformer module is used to capture long-distance dependencies of multi-scale features in a multimodal dataset, thereby obtaining the overall morphology of the tumor and identifying the infiltrative boundary far from the core area. This helps to optimize the 3D bounding box using a hybrid loss function and output sub-image patches.

[0027] As an optional solution to the glioma boundary recognition system based on MRI images described in this invention, the core area boundary recognition module and the infiltration area boundary recognition module in the multi-task learning framework share features, and the range parameters of the infiltration area are inferred by utilizing the location and morphological information of the necrotic / non-enhanced core area.

[0028] As an optional solution of the glioma boundary recognition system based on MRI images described in this invention, 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, thereby verifying whether the range parameters of the infiltration area are accurate.

[0029] The Dice similarity coefficient is used to assess the degree of regional overlap.

[0030] 95% Hausdorff distance is used to evaluate boundary matching.

[0031] The present invention has the following beneficial effects:

[0032] 1. This glioma boundary recognition system based on MRI imaging can extract features simultaneously in the x, y, and z axes by using 3D convolution kernels to obtain the complete three-dimensional structure of the glioma and capture the continuous spatial structure of the tumor.

[0033] 2. This glioma boundary recognition system based on MRI images can quickly output one or more 3D bounding boxes for any input standardized multimodal MRI through the localization module, which can wrap all tumor areas with high confidence, thereby reducing the recognition time of a single sample from tens of minutes to seconds.

[0034] 3. This glioma boundary recognition system based on MRI imaging can clearly and efficiently distinguish the boundaries of the infiltrative area through a cascaded structure design of localization, core area and infiltrative area. Attached Figure Description

[0035] Figure 1 This is a diagram of the overall system framework of the present invention.

[0036] Figure 2 This is a schematic diagram of the modalities of the present invention.

[0037] Figure 3 This is a schematic diagram of the labeled area of ​​the present invention.

[0038] Figure 4 This is a schematic diagram of the AUC curve representing the recognition performance of this invention.

[0039] Figure 5 This is a schematic diagram of the Dice coefficient comparison in this invention.

[0040] Figure 6 This is a schematic diagram of the Hausdorff distance comparison of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1: Please refer to Figures 1-6 One such glioma boundary recognition system based on MRI imaging includes:

[0043] The data input module establishes a database based on several multimodal MRI data sets.

[0044] Several multimodal MRI datasets include labeled multimodal MRI data from patients with gliomas;

[0045] Each patient's data includes the following four core sequences, meticulously annotated at the pixel level by a senior neuroradiologist. (See attached image.) Figure 2 ;

[0046] T1-weighted imaging (T1w): provides excellent detail of anatomical structures;

[0047] T2-weighted imaging (T2w): sensitive to edematous and non-contrast tumor areas;

[0048] Fluid attenuated inversion recovery sequence (FLAIR): can suppress cerebrospinal fluid signals, making the peritumoral edema area more clearly visualized;

[0049] T1 contrast enhancement imaging (T1ce): By injecting contrast agent, it can highlight the area where the blood-brain barrier is disrupted, which usually corresponds to the active core area of ​​the tumor;

[0050] The annotation of the four core sequence data followed the internationally accepted BraTS (Brain Tumor Segmentation Challenge) standard, dividing the tumor region into three subcategories:

[0051] Necrosis and non-enhancing tumor core area;

[0052] Enhance the tumor area;

[0053] Peritumoral edema area;

[0054] The markings of the necrotic and non-enhanced tumor core area, the enhanced tumor area, and the peritumoral edema area will collectively constitute the label of "whole tumor area".

[0055] Thus, this study has established a structured database that efficiently stores and indexes the original images, preprocessed images, labeled data, and model outputs of thousands of patients, ensuring the reproducibility of the study;

[0056] It should be noted that the database uses the BIDS-BrainImagingDataStructure format;

[0057] Upon completion of this module, a large-scale, quality-controlled, and structured MRI dataset of gliomas will be obtained. All data will be formatted into a unified format (NIfTI) and accompanied by clear metadata, providing a high-quality and reliable data source for subsequent standardization processing and model training.

[0058] The data standardization processing module normalizes the multimodal MRI data in the database and performs cospatial registration on the normalized multimodal MRI data to obtain multimodal MRI data in the same format and coordinate system.

[0059] Among them, the data standardization module is key to eliminating data heterogeneity and ensuring the model's generalization ability. Specifically:

[0060] Because different scanning devices and parameters can lead to inconsistent image intensity (grayscale value) ranges, normalization is necessary. Therefore, normalization includes N4ITK bias field correction and Z-Score normalization.

[0061] Among them, N4ITK bias field correction: MRI images often exhibit low-frequency intensity drift (bias field) caused by magnetic field inhomogeneity, which severely 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. Based on the successful experience of the BraTS Challenge and related research, key parameters are configured, and the optimal parameter combination is systematically tested and selected. The initial configuration will refer to: iteration count set to [50,50,50,50], shrinkage factor of 2, and B-spline fitting order of 3. For the B-spline grid resolution, instead of setting a fixed value directly, the control point distance is set to adapt to the image size. The initial value can be set to 200mm, and gradually decreased at multiple resolution levels.

[0062] Among them, Z-Score normalization: After completing the bias field correction, for each modal image of each patient, this study only focuses on the brain parenchyma region and performs Z-Score normalization through a preliminary brain mask extraction.

[0063] To ensure that all patients' brain images are in a common coordinate system, facilitating model learning and comparison, the image data is registered to a standard space, and the T1w images of all patients are linearly (affine transformation) registered to a standard anatomical template, such as the MNI152 template.

[0064] It is worth noting that the antsRegistration tool in the industry-leading ANTs (Advanced Normalization Tools) software package is used because its SyN algorithm performs exceptionally well in nonlinear registration.

[0065] 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 will be a four-dimensional tensor of [Channels, Depth, Height, Width], where Channels correspond to different modalities such as T1, T2, FLAIR, and T1ce. This is the ideal format for inputting into subsequent CNN models.

[0066] The multimodal data alignment module registers and resamples different modal MRI images of the same patient that need to be identified, thereby obtaining a multimodal dataset of the same patient.

[0067] MRI image registration includes:

[0068] A reference modality is set, and all modal MRI images of the patient are registered using a registration tool based on the parameters of the reference modality;

[0069] Resampling involves resampling all registered MRI images of the same patient using a voxel grid and spatial resolution of the reference modality to obtain a multimodal dataset of the same patient.

[0070] The registration tool uses rigid or affine transformations with at least six degrees of freedom to register all modal MRI images of the same patient.

[0071] Any voxel coordinate in the multimodal dataset corresponds precisely to the same physical location in the brain across all modal channels.

[0072] Specifically, the alignment module for multimodal data ensures precise voxel-level alignment between MRI images of the same patient across different modalities. Although multimodal sequences are typically completed in a single scan with minimal motion artifacts in modern clinical scanning, minor head movements or inter-sequence distortions may still exist. Therefore, rigid intramodal registration is required, including registration procedures, technology selection, and resampling.

[0073] Registration process: In this study, the T1ce modality with the clearest anatomical structure was selected as the reference modality. Then, all other modalities of the patient (T1w, T2w, FLAIR, etc.) were registered to this reference modality.

[0074] Technology selection: Registration tools (such as FLIRT / FNIRT) from ANTs or FSL (FMRIB Software Library) are also used. Since it is intermodal registration of the same patient, a rigid or affine transformation with 6 or 7 degrees of freedom is usually sufficient to achieve high accuracy.

[0075] Resampling: After registration, all modalities need to be resampled to the exact same voxel grid and spatial resolution as the reference modal. This study uniformly resamples to an isotropic resolution of 1mm*1mm*1mm, which is crucial for 3D CNN models to fairly handle the three spatial dimensions.

[0076] After this module, a perfectly aligned and resampled multimodal dataset is output. Now, any voxel coordinate (i,j,k) in the dataset precisely 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 enables the extraction of features along the x, y, and z axes simultaneously using 3D convolutional kernels, resulting in the complete three-dimensional structure of gliomas and capturing the continuous spatial structure of the tumor.

[0077] The localization module identifies and obtains a three-dimensional bounding box based on the multimodal dataset of the same patient. The three-dimensional bounding box includes several sub-image blocks of standardized size.

[0078] The localization module includes a 3DCNN model, which identifies multimodal datasets in the axial orthogonal plane, coronal orthogonal plane, and sagittal orthogonal plane to obtain three-dimensional bounding boxes.

[0079] Adjust sub-image blocks to a uniform size using padding or interpolation;

[0080] The 3D bounding box includes the coordinates of the top-left front point and the bottom-right back point;

[0081] Set a safety margin value. When the 3DCNN model recognizes a multimodal dataset, expand the safety margin value outward on the acquired 3D bounding box to optimize the 3D bounding box. The optimized 3D bounding box ensures that the edge immersion area is included.

[0082] Furthermore, to improve computational efficiency and segmentation accuracy, this study adopts a two-stage strategy: first, quickly locate the approximate region of the tumor (ROI), and then perform fine boundary segmentation within this ROI.

[0083] The goal at this stage is speed and recall, rather than pixel-level accuracy. Therefore, this study employs a computationally less computationally intensive 3D CNN model (a variant of Faster R-CNN) on three orthogonal planes (axial, coronal, and sagittal) to identify bounding boxes containing tumors. Considering the integrity of 3D information, a lightweight, simplified 3D U-Net that directly predicts a coarse tumor bounding box on the downsampled 3D image is a better approach.

[0084] The output of the localization module is a 3D bounding box, defined by the coordinates of its upper-left front corner (x_min, y_min, z_min) and lower-right back corner (x_max, y_max, z_max). This bounding box should encompass the entire tumor region (WT). To ensure no marginal infiltration is missed, a safety margin of 20 voxels is added outward from the initially predicted bounding box.

[0085] The output of the localization module (boundary box coordinates) is directly used as the input to the subsequent boundary recognition module. The boundary recognition module will no longer process the entire brain image, but only the sub-volume cropped from the bounding box. This "coarse-to-fine" strategy has three main advantages:

[0086] Reduced computational burden: The fine segmentation model only needs to process a much smaller region containing the tumor, which significantly reduces computational load and memory consumption;

[0087] Reduced class imbalance: In the cropped image patch, the ratio of tumor voxels to non-tumor voxels was improved, which alleviated the class imbalance problem caused by background voxels dominating and helped the model learn tumor features better.

[0088] Standardized input size: All cropped sub-image blocks can be adjusted to a uniform size (e.g., 128x128x128) through padding or interpolation, which facilitates batch training;

[0089] For any input standardized multimodal MRI, the localization module can quickly (in milliseconds) output one or more 3D bounding boxes to wrap all tumor regions with high confidence, thereby reducing the identification time of a single sample from tens of minutes to seconds.

[0090] It also includes setting the IoU value, where the overlap between the 3D bounding box obtained by the localization module and the actual labeled bounding box is greater than the set IoU value.

[0091] It should be noted that the accuracy of the model recognition is as follows: Figure 3 As shown, the instance data is as follows Figure 4 As shown. The localization model was validated on the validation set using the Intersection over Union (IoU) ratio to evaluate the overlap between the predicted bounding box and the ground truth bounding box. For over 99% of cases, the IoU between the actual tumor region and the predicted bounding box should be greater than 0.8, and the actual tumor region should be completely contained within the predicted box. Therefore, this study validated all samples, and those meeting the above requirements were considered to have accurate identification.

[0092] The multi-task learning framework includes a core area boundary recognition module and an infiltration area boundary recognition module. The multi-task learning framework is used to output the range parameters of the enhanced tumor area, the necrotic / non-enhanced core area, and the infiltration area.

[0093] In the multi-task learning framework, the core area boundary recognition module and the infiltration area boundary recognition module share features. The location and morphological information of the necrotic / non-enhanced core area are used to help infer the range parameters of the infiltration area.

[0094] 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-enhanced core areas.

[0095] The core area boundary identification module also includes a residual Transformer module;

[0096] Multi-scale features of a multimodal dataset are fused using the encoder structure, decoder structure, and skip connections of a 3D U-Net.

[0097] The Transformer module is used to capture long-distance dependencies of multi-scale features in a multimodal dataset, thereby obtaining the overall morphology of the tumor and identifying the infiltrative boundary far from the core area. This helps to optimize the 3D bounding box using a hybrid loss function and output sub-image patches.

[0098] Specifically, the core region boundary recognition module is responsible for accurately segmenting the core region of the tumor within the ROI provided by the localization module, namely the enhanced tumor region (ET) and the necrotic / non-enhanced core region (NCR / NET). The input to this module is a multimodal sub-image patch that has been localized, cropped, and sized normally.

[0099] The advanced network architecture 3D U-Net is one of the most classic and powerful models in the field of medical image segmentation. Its encoder-decoder structure and skip connections effectively fuse multi-scale features, capturing both global context and preserving local details. The introduction of residual modules (such as DeepMedic or dResU-Net) alleviates the vanishing gradient problem in deep networks, allowing for the construction of deeper and more expressive models. Finally, a self-attention mechanism is integrated, particularly by introducing 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 invasive boundaries far from the core region.

[0100] Regarding the choice of loss function, the class imbalance problem still exists due to the large size differences among the various subregions (ET, NCR / NET) within the tumor. This study adopts 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 on making the model pay attention to hard-to-classify samples (usually pixels at the boundaries), which can improve the sharpness of boundary segmentation.

[0101] Following the core region boundary recognition module, the output will be a pixel-level segmentation mask of the same size as the input sub-image patch, where each voxel is labeled as one of background, enhanced tumor (ET), or necrotic / non-enhanced core region (NCR / NET).

[0102] The core area boundary identification module will be validated using the following metrics: Dice similarity coefficient and 95% Hausdorff distance for quantitative evaluation.

[0103] The infiltration zone boundary recognition module outputs a complete full tumor region segmentation mask within a sub-image block, and subtracts the enhanced tumor region and necrotic / non-enhanced core region mask from the full tumor region to obtain the range parameters of the infiltration zone;

[0104] The immersion zone boundary identification module uses the Dice similarity coefficient and 95% Hausdorff distance to quantitatively evaluate the immersion zone's range parameters, thereby verifying the accuracy of the immersion zone's range parameters.

[0105] The Dice similarity coefficient is used to assess the degree of regional overlap.

[0106] 95% Hausdorff distance is used to evaluate boundary matching.

[0107] The core area boundary identification module is used to identify the area where tumor cells have infiltrated but have not yet broken the blood-brain barrier, namely the peritumoral edema (ED) area with high signal on the FLAIR sequence. The identification of the infiltrated area relies heavily on the detailed analysis of signal abnormalities on the FLAIR and T2 sequences, and also needs to combine T1 and T1ce information to exclude the core area.

[0108] Specifically, the infiltration region boundary recognition module and the core region recognition module can be integrated into a multi-task learning framework, with the same 3D CNN model simultaneously outputting segmentation results for three sub-regions (ET, NCR / NET, ED). This design allows for feature sharing within the model, utilizing the location and morphological information of the core region to assist in inferring the extent of the infiltration region.

[0109] Boundary-aware loss is a technical challenge, as the boundaries of the infiltrative region are extremely blurry, representing typical "hard-to-classify samples." This study introduces a loss function specifically for boundaries, directly penalizing the boundary loss caused by misclassification of 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 the blurry boundaries. The expected result of the infiltrative region boundary recognition module is to output a complete "whole tumor region" (WT) segmentation mask within the ROI. The extent of the infiltrative region (ED) can be obtained by subtracting the core region (ET+NCR / NET) mask from the WT mask. The infiltrative region boundary recognition module will be validated using the following metrics: Dice similarity coefficient and 95% Hausdorff distance for quantitative evaluation.

[0110] In summary, the cascaded structure design of the positioning, core area, and immersion area can clearly and efficiently distinguish the boundaries of the immersion area.

[0111] The results verification module summarizes the range parameters of the obtained sub-image blocks, enhanced tumor areas, necrotic / non-enhanced core areas, and infiltrative areas, and combines them with multimodal MRI data of the same format and coordinate system for matching and evaluation, thereby outputting the recognition results.

[0112] The results validation module integrates the first six modules. For any dataset, it performs the steps of "data input," "data standardization," "multimodal data alignment," "tumor localization," "core area boundary recognition," and "infiltrative area boundary recognition." The results are then compared with the labeled results to verify the recognition effectiveness. For clinical data, the system directly outputs the glioma boundaries identified by the model for clinicians' reference.

[0113] The validation module validated an independent dataset containing 200 samples. Based on the annotation results, the model's performance was validated. This module used the following metrics for validation: sensitivity (Recall) and specificity, to evaluate the model's ability to detect tumor voxels and correctly identify background voxels. Dice similarity coefficient (assessing region overlap) and Hausdorff distance (assessing boundary matching) are two main metrics for evaluating segmentation performance. This study calculated metrics for three labels: enhancement region (ET), core region (ET+NCR / NET), and whole tumor region (WT). This is crucial for assessing clinical risk (such as missed tumors or over-resection).

[0114] This application proposes an end-to-end, 3D CNN-based preoperative boundary recognition system for gliomas called border-3D CNN. Through modular design, it progressively addresses the challenges of the entire process, from data preparation to model training and result validation, thereby achieving an intelligent segmentation tool with clinically applicable accuracy and robustness, providing strong technical support for precise surgical treatment of gliomas.

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0116] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical 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 glioma boundary recognition system based on MRI imaging, characterized in that, include: The data input module establishes a database based on several sets of multimodal MRI data. The data standardization processing module normalizes the multimodal MRI data in the database and performs cospatial registration on the normalized multimodal MRI data to obtain multimodal MRI data in the same format and coordinate system. The multimodal data alignment module registers and resamples different modal MRI images of the same patient that need to be identified, thereby obtaining a multimodal dataset of the same patient. The localization module 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 size. 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-enhanced core region, and the infiltration region. The results verification module summarizes the range parameters of the obtained sub-image blocks, enhanced tumor areas, necrotic / non-enhanced core areas, and infiltrative areas, and combines them with multimodal MRI data of the same format and coordinate system for matching and evaluation, thereby outputting the recognition results. 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 the enhanced tumor region and the necrotic / non-enhanced core region. The infiltration zone boundary recognition module outputs a complete full tumor region segmentation mask within a sub-image block, and subtracts the enhanced tumor region and necrotic / non-enhanced core region masks from the full tumor region to obtain the range parameters of the infiltration zone. The localization module includes a 3DCNN model, which identifies multimodal datasets in the axial orthogonal plane, coronal orthogonal plane, and sagittal orthogonal plane to obtain the three-dimensional bounding box. The sub-image blocks are adjusted to a uniform size by padding or interpolation. 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 acquired 3D bounding box to optimize the 3D bounding box. The optimized 3D bounding box ensures that it includes the edge immersion area. The core area boundary identification module also includes a Transformer module; Multi-scale features of a multimodal dataset are fused through the encoder structure, decoder structure, and skip connections of the 3D U-Net. The Transformer module is used to capture long-distance dependencies of multi-scale features in a multimodal dataset, thereby obtaining the overall morphology of the tumor and identifying the infiltrative boundary far from the core area. This helps to optimize the 3D bounding box using a hybrid loss function and output sub-image patches.

2. The glioma boundary recognition system based on MRI imaging according to claim 1, characterized in that, The MRI image registration includes: A reference modality is set, and all modal MRI images of the patient are registered using a registration tool based on the parameters of the reference modality; The resampling involves resampling all registered MRI images of the same patient using a voxel grid and spatial resolution of the reference modality to obtain a multimodal dataset of the same patient.

3. The glioma boundary recognition system based on MRI imaging according to claim 2, characterized in that: The registration tool uses a rigid or affine transformation with at least six degrees of freedom to register all modal MRI images of the same patient. Any voxel coordinate in the multimodal dataset corresponds precisely to the same physical location in the brain across all modal channels.

4. The glioma boundary recognition system based on MRI imaging according to claim 1, characterized in that: It also includes setting an IoU value, wherein the IoU value of the overlap between the 3D bounding box obtained by the positioning module and the actual labeled bounding box is greater than the set IoU value.

5. The glioma boundary recognition system based on MRI imaging according to claim 1, characterized in that: The core region boundary recognition module and the infiltration region boundary recognition module in the multi-task learning framework share features, and use the location and morphological information of the necrotic / non-enhanced core region to help infer the range parameters of the infiltration region.

6. The glioma boundary recognition system based on MRI imaging according to claim 5, characterized in that: The immersion zone boundary identification module quantitatively evaluates the range parameters of the immersion zone using the Dice similarity coefficient and the 95% Hausdorff distance, thereby verifying the accuracy of the range parameters of the immersion zone. The Dice similarity coefficient is used to assess the degree of regional overlap. The 95% Hausdorff distance is used to evaluate the boundary matching degree.

Images