MRI brain tumor grading method for children

Abstract

The invention provides a children MRI brain tumor grading method. The children MRI brain tumor grading method comprises the following steps: preprocessing brain MRI image sequences of T1, T1c, T2, Flair and ADC modals to obtain tensors corresponding to the brain MRI image sequences; inputting the tensor of the T1 mode into an SE feature extraction channel for feature extraction, wherein the SE feature extraction channel performs feature extraction based on an SE attention mechanism; tensors of the T1c, the T2, the Flair and the ADC mode are input into an MSFE multi-scale feature extraction module for feature extraction; features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module are subjected to feature fusion and then are input into a classification module for classification prediction to obtain a classification prediction result; according to the method, undetectable subtle changes in early tumors can be recognized, and the calculation load is effectively reduced while high precision is maintained.

Description

Technical Field

[0001] This invention relates to the fields of medical imaging and deep learning, and in particular to a method and system for grading brain tumors in children using MRI. Background Technology

[0002] In the field of medical imaging, the grading of pediatric brain tumors is an important and challenging task. Traditional radiomics techniques rely on manual diagnosis, which is not only time-consuming and labor-intensive but also easily influenced by subjective factors, making it difficult to guarantee the consistency and accuracy of diagnostic results. In recent years, with the continuous development of artificial intelligence, Transformer has been widely used in the field of 3D medical imaging, providing new ideas for the grading of pediatric brain tumors.

[0003] In existing technologies, 3D medical image modeling often faces a technical bottleneck where accuracy and efficiency are difficult to balance. To achieve high accuracy, the model needs to process massive amounts of voxel data, leading to an exponential increase in computational complexity and placing enormous pressure on hardware, making it difficult to meet the timeliness requirements of real-time clinical diagnosis. On the other hand, downsampling or simplifying the network to improve efficiency inevitably results in the loss of crucial details about minute lesions, directly affecting the accuracy of grading. Meanwhile, pediatric brain tumors present unique challenges in early imaging, with their signs often being extremely subtle, such as only weak enhancement, slight mass effect, or signal characteristics similar to normally developing brain tissue. This makes it easy for traditional methods and human visual identification to miss or misdiagnose them as benign lesions. This diagnostic uncertainty often leads to two extreme consequences: first, treatment delays due to observation and waiting, missing the optimal intervention period; second, overdiagnosis leading to aggressive biopsies or surgeries, causing unnecessary trauma to the child. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for grading pediatric MRI brain tumors. Based on a 3D Transformer model, this invention achieves accurate modeling of complex brain tumor features through the synergy of a multi-scale feature extraction module and a multi-head self-attention mechanism. Firstly, the multi-scale module focuses on the efficient extraction of local details, enhancing the ability to capture local information while effectively controlling computational complexity. Secondly, the self-attention mechanism is responsible for modeling the global context and long-range dependencies between multiple sequences. These two mechanisms work in parallel, ultimately achieving an efficient fusion of local and global features.

[0005] To achieve the above objectives, the first aspect of the present invention provides a method for grading pediatric MRI brain tumors, comprising the following steps: Tensors for the corresponding brain MRI image sequences were obtained by preprocessing the brain MRI image sequences of T1, T1c, T2, Flair, and ADC modes. The tensor of the T1 mode is input into the SE feature extraction channel for feature extraction, and the SE feature extraction channel performs feature extraction based on the SE attention mechanism; The tensors of T1c, T2, Flair, and ADC modes are input into the MSFE multi-scale feature extraction module for feature extraction. After fusing the features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module, the feature is input into the classification module for classification prediction to obtain hierarchical prediction results. The MSFE multi-scale feature extraction module uses the following methods for feature extraction: Input the tensors of T1c, T2, Flair and ADC modes into the gated network to obtain the mode importance weights and initial feature maps corresponding to T1c, T2, Flair and ADC modes; If the modality importance weight of the corresponding modality is not lower than a certain threshold, the first feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with high importance. If the modality importance weight of the corresponding modality is lower than a certain threshold, the second feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with low importance. The feature maps obtained from the first and second feature extractions are then weighted and fused.

[0006] The aforementioned scheme employs different attention mechanisms for different brain image tensors. Specifically, the T1 modality tensor utilizes the SE attention mechanism, enabling the model to adaptively calibrate the importance of channel features, thereby significantly improving its ability to represent key anatomical structures. This mechanism effectively focuses on sequence channel features closely related to tumor diagnosis with low computational cost, while suppressing noise and irrelevant information interference. Furthermore, the SE-based T1 sequence not only achieves stable anchoring of anatomical structures and interference filtering but also promotes complementary fusion among multiple sequences—alleviating the "spatial misalignment" and "artifact interference" problems commonly encountered by Flair and T1c sequences in tumor localization, and providing a reliable anatomical reference benchmark for multimodal feature fusion. In contrast, mechanisms such as spatial attention are susceptible to interference from normal tissue structures in brain images and have high computational complexity. Therefore, the SE mechanism, while ensuring the needs of real-time clinical diagnosis, further improves the model's accuracy and robustness in tumor identification.

[0007] The tensors for T1c, T2, FLAIR, and ADC modalities utilize the MSFE multi-scale feature extraction module. For brain MRI image sequences with modality importance weights not lower than a certain threshold, a first feature extraction method is used for fine extraction, capturing differential features from minute textures to macroscopic regions, especially subtle changes that are difficult to detect in early-stage tumors. For brain MRI image sequences with modality importance weights lower than a certain threshold, a second feature extraction method is used for rapid extraction to handle low-importance image regions. These two feature extraction methods do not perform complex and time-consuming calculations on all modalities; the model can dynamically guide images with importance weights below the threshold to a less computationally expensive path, significantly improving the model's efficiency and speed.

[0008] By employing a gating mechanism, the model becomes more efficient, and the final model output integrates clinical dimensions while highlighting key evidence. Furthermore, it can still function smoothly even when a certain modality is missing.

[0009] By fusing features extracted by different feature extraction modules at multiple scales, and integrating the enhanced multi-sequence functional information with the clear anatomical structure provided by the T1 modality tensor, a more comprehensive and discriminative feature representation is constructed.

[0010] In one possible embodiment, the method for the gated network to obtain modality importance weights and initial feature maps includes: Initial feature maps are obtained by extracting features from the tensors of the input T1c, T2, Flair, and ADC modes respectively, and the initial feature maps are then transformed into feature vectors. The gated network pre-sets four expert networks. The four expert networks first score the feature vectors of each modality, then the scores of the four expert networks are weighted and summed, and the sum is normalized. The normalized weighted sum score is the modality importance weight of each modality. Each expert network is used to evaluate a specific clinical dimension, which includes: tumor enhancement features, overall tumor morphology and surrounding edema extent, microedema and infiltrative foci, and cell density.

[0011] The above scheme evaluates the tensor of each modality across four expert networks using clinical dimensions, obtaining scores for each dimension. The final decision is based on the combined scores of all modalities across each dimension; it integrates the opinions of all expert networks to arrive at the most comprehensive diagnosis; and it embeds clinical prior knowledge into the expert networks, forcing the model to solve problems in the way human experts think, thus achieving a qualitative leap in performance, interpretability, and robustness.

[0012] In one possible embodiment, the method for extracting the second feature includes: Let the number of input channels be C. in Number of output channels C out And the number of output channels C out The number of input channels C in 2 times; Initial feature maps whose input modality importance weights are below a certain threshold; where each input channel corresponds to an initial feature map of one modality; For the initial feature map of each input channel, a 3×3×3 depthwise convolution is first performed to extract features to obtain a size H. in ×W in ×D in The input feature map is then subjected to a 1×1×1 pointwise grouped convolution. The pointwise grouping convolution includes: Get C in The input feature map after performing a 3×3×3 depthwise convolution; Based on the number of input channels C in With the number of output channels C out The greatest common divisor determines the number of groups G, and the number of input channels C. in and the number of output channels C out They were all divided into Group G; Within each group, perform a 1×1×1 pointwise convolution, using C for each group. out / G 1×1×1 convolution kernels, for the C of this group in Processing / G input channels yields a group size of H. out ×W out ×D out The output feature map; The output feature maps of all G groups are concatenated along the channel dimension, resulting in an output size of C. out ×H out ×W out ×D out The first feature map; where H out W out and D out These represent the height, width, and depth of the output feature map, respectively, and their values ​​are equal to the corresponding dimensions of the input feature map.

[0013] The above scheme performs multi-channel processing on initial brain feature maps with modality importance weights below a certain threshold. Modalities with low importance weights often have low signal-to-noise ratios, unstable quality, or weak relevance to the core of the current task. Processing such data via a single path can easily lead to misinterpretations by noise or artifacts. Multi-channel processing can address these data from different perspectives. By aggregating the outputs of multiple channels, the model can average out the errors of individual channels, effectively suppressing noise interference. Furthermore, through pointwise grouped convolution, the number of parameters and computational cost are further reduced to 1 / G compared to the original pointwise convolution, enabling the construction of a more lightweight model.

[0014] In one possible embodiment, the method for the first feature extraction includes: The initial feature maps with modality importance weights not lower than a certain threshold are concatenated. The concatenated initial feature maps are then subjected to local feature extraction through a 3×3×3 depthwise convolution. A non-linear mapping is introduced through a non-linear activation function to enhance the model's ability to represent complex features. Then, a pooling layer is used to perform downsampling operations to achieve feature dimension compression and key information preservation, and a second feature map is output.

[0015] Using the first feature extraction method, for initial feature maps with modality importance weights not lower than a set threshold (i.e., key modalities), these maps are concatenated along channels to form a feature map that fuses multi-sequence information. This operation aims to aggregate complementary information across modalities. Subsequently, the first feature extraction method is used to perform deep feature extraction on the synthesized image to accurately capture its complex structures and details.

[0016] In one possible embodiment, the multi-scale feature fusion operation includes: The features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module are aligned and embedded with positional encoding to restore their feature space structure. Then, the fused third feature map is output through the Dropout operation.

[0017] The above scheme uses weighted fusion of T1c, T2, Flair, and ADC sequences to highlight the features most relevant to tumor grading in different sequences, achieving information complementarity. Subsequently, it aligns with T1 features, integrating the enhanced multi-sequence functional information with the clear anatomical structure provided by the T1 sequence to construct a more comprehensive and discriminative feature representation.

[0018] To achieve the above objectives, a second aspect of the present invention provides a pediatric MRI brain tumor grading system, comprising: an input module, a feature extraction module, and a result output module; The input module is used to preprocess brain MRI image sequences of T1, T1c, T2, Flair and ADC modes to obtain tensors of the corresponding brain MRI image sequences. The feature extraction module is used to input the tensor of the T1 mode into the SE feature extraction channel for feature extraction. The SE feature extraction channel performs feature extraction based on the SE attention mechanism. The feature extraction module includes an MSFE multi-scale feature extraction module, which is used to input the tensors of T1c, T2, Flair and ADC modes into the MSFE multi-scale feature extraction module for feature extraction. The MSFE multi-scale feature extraction module uses the following methods for feature extraction: Input the tensors of T1c, T2, Flair and ADC modes into the gated network to obtain the mode importance weights and initial feature maps corresponding to T1c, T2, Flair and ADC modes; If the modality importance weight of the corresponding modality is not lower than a certain threshold, the first feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with high importance. If the modality importance weight of the corresponding modality is lower than a certain threshold, the second feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with low importance. The feature maps obtained from the first and second feature extractions are then weighted and fused. The result output module is used to fuse the features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module, and then input the fusion into the classification module to obtain the hierarchical prediction result.

[0019] To achieve the above objectives, a third aspect of the present invention provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. A memory for storing computer programs; a processor for executing the programs stored in the memory to implement the steps of a pediatric MRI brain tumor grading method as described in the first aspect above.

[0020] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the steps of a pediatric MRI brain tumor grading method as described in the first aspect above.

[0021] The beneficial effects of this invention are as follows: 1. Through gating networks, the model can automatically identify and distinguish between critical and secondary modalities. For critical modalities, a refined first feature extraction is used, concentrating computational resources to capture deep details; for secondary modalities, a lightweight second feature extraction or multi-channel processing is used, significantly improving computational efficiency while ensuring no information loss.

[0022] 2. The MSFE module ensures the extraction of the most suitable multi-scale features from images of different modalities and importance levels, achieving comprehensive perception from local details to global context. Simultaneously, it significantly reduces computational resource consumption while maintaining accuracy, making it more suitable for practical clinical applications. Through fine feature extraction within multiple modalities and dynamic weighting between modalities, it solves the technical challenge of balancing accuracy and efficiency in 3D image modeling.

[0023] 3. Different attention mechanisms are used for different sequences, making the image sequences more compatible with the attention mechanisms, improving the utilization of multi-sequence data and reducing resource waste. Attached Figure Description

[0024] Figure 1 This is an overall model structure diagram provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the SE attention mechanism provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the MSFE module provided in an embodiment of the present invention; Figure 4 This is a diagram of the depth-separable grouped convolutional structure provided in an embodiment of the present invention; Figure 5 This is a structural diagram of the multi-scale feature fusion module provided in an embodiment of the present invention; Figure 6 This is a structural diagram of the multi-head self-attention mechanism provided in an embodiment of the present invention; Figure 7 These are the AUC curves of various models provided in the embodiments of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0026] To facilitate understanding of the technical solutions provided in this application, the technical terms involved in the embodiments of this application are explained below.

[0027] AUC: Defined as the area under the ROC curve and the coordinate axis. Since the ROC curve is generally above the line y=x, the value of AUC ranges between 0.5 and 1. The closer the AUC is to 1.0, the higher the realism of the detection method; when it is equal to 0.5, the realism is the lowest.

[0028] Magnetic Resonance Imaging (MRI): MRI is a relatively new medical imaging technique that uses static and radiofrequency magnetic fields to image human tissues. During the imaging process, it obtains high-contrast, clear images without the need for electron ionizing radiation or contrast agents. It can reveal abnormalities and early lesions in human organs from within the molecular structure, and in many ways, it is superior to X-ray CT. MRI images typically include multiple sequences, such as T1, T1c, T2, FLAIR, and ADC sequences. These different sequences can highlight different lesion areas.

[0029] refer to Figure 1 The steps for grading brain tumors in children using MRI are as follows: MRI sequence images were acquired and preprocessed to obtain tensors of the MRI image sequences. These tensors were then input into five channels in the order of T1, T1c, T2, FLAIR, and ADC. The T1 sequence was used to extract features based on the SE attention mechanism, effectively focusing on the sequence channel features most relevant to tumor diagnosis while suppressing noise and irrelevant information. The remaining T1c, T2, FLAIR, and ADC brain MRI image tensors were extracted using the MSFE multi-scale feature extraction module. refer to Figure 5 The T1c, T2, Flair and ADC are weighted and fused, and the weighted and fused feature map is aligned with the T1 sequence feature map and embedded with position encoding to restore its feature space structure. Then, the fused feature map is output through the Dropout operation.

[0030] The output after fusing multi-scale features is concatenated and input into a multi-head self-attention mechanism, passing through the first scaling layer, normalization layer, multilayer perceptron, and second scaling layer in sequence, and finally the prediction result is output by the classification head.

[0031] Example 1 To achieve the above objectives, the first aspect of the present invention provides a method for grading pediatric MRI brain tumors, comprising the following steps: Step S1: After preprocessing the brain MRI image sequences of T1, T1c, T2, Flair and ADC modalities, tensors of the corresponding brain MRI image sequences are obtained; It should be noted that after acquiring five-sequence brain MRI image data, N4 field correction is first performed to eliminate the influence of magnetic field inhomogeneity. Subsequently, the images are resampled using a 0.5×0.5×5 mm³ voxel size, and the T1, T2, FLAIR, and ADC sequences are uniformly registered to the T1c space to ensure strict spatial alignment of multimodal data. The pediatric MRI brain tumor grading model receives registered five-sequence MRI data from the same patient. By fusing multi-sequence information, it improves the accuracy of lesion detection, addresses the problem of unclear early tumor identification in children leading to treatment delays or overtreatment, and strives for the best intervention window for children.

[0032] Step S2: Input the tensor of mode T1 into the SE feature extraction channel for feature extraction. The SE feature extraction channel performs feature extraction based on the SE attention mechanism. The reason for using the SE feature fusion mechanism for T1 sequences is that, compared with other sequences, T1 sequences can achieve "anatomical anchoring + interference filtering + multi-sequence complementarity" through SE, which not only solves the problems of "spatial misalignment" and "artifact interference" in tumor localization of sequences such as Flair and T1c, but also provides a stable anatomical reference for multi-sequence feature fusion.

[0033] For specific details on the structure of SE, please refer to [link / reference]. Figure 2 The SE attention mechanism's processing flow begins with an input feature block of shape (H', W', C'). First, it is transformed into an intermediate feature block U (H×W×C) through feature transformation operations. Next, a compression operation is performed: global average pooling is applied to U, compressing the spatial information of each channel into a single scalar, generating a channel descriptor containing global statistical information. Then, an activation operation is performed: a gating mechanism including fully connected layers and activation functions is used to learn the non-linear dependencies between channels, generating channel weights between 0 and 1. Finally, a recalibration operation is performed: the learned channel weights are multiplied channel-by-channel with the original feature block U, enhancing important features and suppressing minor features, outputting a calibrated feature block of the same size as U, completing the "compression-activation-recalibration" process.

[0034] Step S3: Input the tensors of T1c, T2, Flair and ADC modes into the MSFE multi-scale feature extraction module for feature extraction; While extracting features from T1 brain MRI image sequences based on the SE feature fusion mechanism, feature extraction was performed on T1c, T2, FLAIR, and ADC brain MRI image sequences using the MSFE multi-scale feature extraction module. Specifically, the feature extraction methods of the MSFE multi-scale feature extraction module include: Input the tensors of T1c, T2, Flair and ADC modes into the gated network to obtain the mode importance weights and initial feature maps corresponding to T1c, T2, Flair and ADC modes; If the modality importance weight of the corresponding modality is not lower than a certain threshold, the first feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with high importance. If the modality importance weight of the corresponding modality is lower than a certain threshold, the second feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with low importance. The feature maps obtained from the first and second feature extractions are then weighted and fused.

[0035] refer to Figure 3 The diagram illustrates the MSFE module mechanism. The MSFE module processes tensors from brain MRI image sequences via a gating network, generating importance weights and initial feature maps for four modalities: T1c, T2, FLAIR, and ADC. These weights are used to assess the importance of different MRI brain tumor sequences and dynamically select between first and second feature extraction methods. Finally, the feature images output from the two paths are weighted and fused, then input into the classification module for classification prediction to obtain a hierarchical prediction result.

[0036] Specifically, methods for obtaining modality importance weights and initial feature maps in gating networks include: Initial feature maps are obtained by extracting features from the tensors of the input T1c, T2, Flair, and ADC modes respectively, and the initial feature maps are then transformed into feature vectors. The gated network pre-sets four expert networks. The four expert networks first score the feature vectors of each modality, then the scores of the four expert networks are weighted and summed, and the sum is normalized. The normalized weighted sum score is the modality importance weight of each modality. Each expert network is used to evaluate a specific clinical dimension, which includes: tumor enhancement features, overall tumor morphology and surrounding edema extent, microedema and infiltrative foci, and cell density.

[0037] It should be noted that the model employs a multi-expert system architecture, with each expert network built upon a pre-trained model and focusing on a different specific clinical dimension. Each expert performs in-depth evaluation of its specialized dimension, outputting structured semantic features. The final decision layer uses a weighted fusion mechanism to aggregate the outputs of all expert modules. The weight of each expert is determined by its pre-defined level of expertise; a higher expert weight indicates stronger discriminative authority in that clinical dimension. All expert weights and modality importance weights undergo normalization.

[0038] After obtaining the modality importance weights based on the gating network, the system dynamically selects between first and second feature extraction. If the modality importance weight of a corresponding modality is lower than a certain threshold, feature extraction is performed on the initial feature map of that modality. The second feature extraction method includes: Let the number of input channels be C. in Number of output channels C out And the number of output channels C out The number of input channels C in 2 times; Initial feature maps whose input modality importance weights are below a certain threshold; where each input channel corresponds to an initial feature map of one modality; For the initial feature map of each input channel, a 3×3×3 depthwise convolution is first performed to extract features to obtain a size H. in ×W in ×D in The input feature map is then subjected to a 1×1×1 pointwise grouped convolution. The pointwise grouping convolution includes: Get C in The input feature map after performing a 3×3×3 depthwise convolution; Based on the number of input channels C in With the number of output channels C out The greatest common divisor determines the number of groups G, and the number of input channels C. in and the number of output channels C out They were all divided into Group G; Within each group, perform a 1×1×1 pointwise convolution, using C for each group. out / G 1×1×1 convolution kernels, for the C of this group in Processing / G input channels yields a group size of H. out ×W out ×D out The output feature map; The output feature maps of all G groups are concatenated along the channel dimension, resulting in an output size of C. out ×H out ×W out ×D out The first feature map; where H out W out and D out These represent the height, width, and depth of the output feature map, respectively, and their values ​​are equal to the corresponding dimensions of the input feature map.

[0039] refer to Figure 4 B in the figure gives the number of input channels C. in For example, if the input channels change, the above point-by-point grouping convolution method is used for processing.

[0040] This can be understood as follows: if the number of input image sequences using the second feature extraction is reduced, the number of input channels is also reduced by the same amount. However, the setting of grouped convolution still follows the rule of point-by-point grouped convolution mentioned above, that is, the number of groups is consistent with the number of input channels.

[0041] It should be noted that the system can dynamically control the number of initial feature maps entering the gating network by adjusting the modality importance weights. The second feature extraction is a fast extraction path designed for efficient preprocessing of the initial feature maps, rather than replacing the first feature extraction for deep feature extraction. The data obtained from this fast extraction path will provide supplementary information to the first feature map in the subsequent fusion stage, enhancing its feature representation.

[0042] If the modality importance weight of the corresponding brain MRI image sequence is not lower than a certain threshold, then the first feature extraction method is used to extract features from the brain MRI image sequence. The first feature extraction method includes: The initial feature maps with modality importance weights not lower than a certain threshold are concatenated. The concatenated initial feature maps are then subjected to local feature extraction through a 3×3×3 depthwise convolution. A non-linear mapping is introduced through a non-linear activation function to enhance the model's ability to represent complex features. Then, a pooling layer is used to perform downsampling operations to achieve feature dimension compression and key information preservation, and a second feature map is output.

[0043] Step S4: The output after multi-scale feature fusion is concatenated and input into the multi-head self-attention mechanism, and then input into the classification module for classification prediction to obtain the prediction result; refer to Figure 6 Specifically, the multi-scale feature fusion operation includes: The features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module are aligned and embedded with positional encoding to restore their feature space structure. Then, the fused third feature map is output through the Dropout operation.

[0044] Alignment unifies all features into a comparable and consistent space, making subsequent fusion and computation more efficient and significantly improving the clarity and consistency of feature representation.

[0045] It should be noted that after multi-scale feature fusion, the data is input into the classification module for classification prediction to obtain the prediction result. The classification module further includes: a first scaling layer, a normalization layer, a multilayer perceptron, and a second scaling layer. Finally, the classification head outputs the prediction result, specifically including: Brain tumor MRI feature maps first enter the first scaling layer for adaptive scaling to match the optimal numerical range for subsequent processing. Then, the data flows through a normalization layer, where its distribution is readjusted and stabilized, accelerating training convergence and improving model robustness. The stabilized features are then fed into a multilayer perceptron, where nonlinear interactions and information fusion in high-dimensional space significantly enhance their feature representation capabilities. The fused high-level features then pass through a second scaling layer for a final scale calibration before output. Finally, the deeply refined feature vectors are sent to the classification head to be converted into the final class prediction probabilities.

[0046] After performing multi-scale feature fusion, the outputs of the multi-scale feature fusion are concatenated and input into the multi-head self-attention mechanism, referencing... Figure 6 The processing flow of the multi-head self-attention mechanism is as follows: The input features first pass through a normalization layer to normalize the channel dimensions of the features, stabilizing the numerical distribution for subsequent calculations. The normalized features are then mapped in parallel into three components: a query vector Q, a key vector K, and a value vector V. These three components have the same shape and correspond to the core roles of "query, matching, and information aggregation" in multi-head self-attention. K and V are first added together, and then Softmax is applied to the result to generate attention weights. The attention weights are multiplied by Q and then input into a fully connected layer for feature transformation. Finally, the output of the fully connected layer is added to the initial Layer Norm output, and then passed through a Concat layer to obtain the final self-attention features.

[0047] Multi-head self-attention mechanisms offer three advantages in brain tumor identification: First, their multi-head design allows the model to focus on features at different levels in parallel; for example, some heads focus on local details while others capture the global context, resulting in a more comprehensive understanding. Second, by calculating the attention weights between each pixel and all pixels in the image, this mechanism can accurately model long-range dependencies, which helps to clearly define the complex boundaries between tumor infiltration areas and normal tissue. Finally, when processing multi-sequence MRI data, multi-head self-attention can efficiently achieve cross-sequence feature fusion, such as automatically associating enhancement regions in T1c sequences with edema signals in FLAIR sequences, thereby constructing richer and more robust tumor representations and significantly improving the utilization efficiency of multimodal data.

[0048] Example 2 This embodiment provides a verification experiment of a pediatric MRI brain tumor grading method compared to Example 1.

[0049] Simulation Experiment In this embodiment, the model reaches a performance plateau after approximately 100 training epochs, therefore the total number of training epochs is set to 100. During training, each weight update is based on the gradient calculated from 16 samples (i.e., the batch size is set to 16). The loss function chosen is BCEWithLogitsLoss, which is suitable for binary classification tasks. To improve training stability and convergence, a cosine annealing strategy is used to dynamically adjust the learning rate, with the initial learning rate set to 1×10⁻⁻⁶. 4 .

[0050] To verify the effectiveness of the method proposed in this embodiment, a clinical model, a radiomics comparison model, and a deep learning comparison model were constructed, as detailed below: (1) Clinical model: Four features were selected: age, gender, type and location. A univariate analysis was performed using the chi-square test. A clinical model was constructed using RandomForest based on the selected features.

[0051] (2) Radiomics Comparison Model: Radiomics features of regions of interest were extracted from five sequences, including first-order features, shape-based features, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and wavelet features. First, a chi-square test was used for univariate analysis, followed by binary minimum absolute shrinkage and selection operator (LASSO) regression for multivariate analysis. CLS token features were extracted from the optimal model. After feature standardization, five-fold cross-validation was used for training, and important features were selected. Finally, radiomics models based on Random Forest, Support Vector Machine (SVM), and Decision Tree were constructed to distinguish between low-grade and high-grade pediatric brain tumors.

[0052] (3) Deep learning comparison models: Using the same dataset partitioning method and hyperparameter settings as the 3D C-Vit model, improved VGG16, DenseNet121, ResNet50 and Transformer-based deep learning models were constructed to distinguish pediatric brain tumors of the same category.

[0053] Experimental results are as follows Figure 7 As shown in the figure, the AUC curves of each model are compared; Table 1 lists the detailed comparison results of each model on different evaluation indicators.

[0054] Table 1 Comparison of pediatric brain tumor grading models To verify the effects of the multi-head self-attention mechanism MHSA, the multi-scale feature extraction module MSFE, and the SE module on the grading of pediatric brain tumors, ablation experiments were conducted, and the results are shown in Table 2.

[0055] Table 2 Comparison of Ablation Experiments Base MHSA MSFE SE AUC Accuracy Precision F1 Recall √ √ √ √ 91.36 87.50 95.45 89.29 92.59 √ √ × √ 80.22 75.83 76.89 79.34 81.97 √ √ √ × 84.57 80.58 85.07 95.45 98.12 √ √ × × 81.89 73.33 76.64 75.00 95.74 √ × × × 80.23 71.69 76.56 78.77 85.49 Adding the MHSA module separately to the base model improved the AUC to 81.89% (+1.66%) and accuracy to 73.33% (+1.64%). Notably, the recall rate increased significantly from 85.49% to 95.74% (+10.25%) after adding MHSA, indicating that the multi-head self-attention mechanism has advantages in capturing long-distance dependencies between tumor and surrounding tissues and global contextual information, which helps reduce the missed diagnosis of high-grade lesions.

[0056] The SE module effectively improves feature fusion. After adding the SE module, the AUC slightly decreased to 80.22%, but the accuracy improved to 75.83% (+4.14%). This module enhances T1 channel features through SE attention and dynamically processes other channels through a gating network, effectively fusing complementary information from multiple MRI sequences and improving the representation of discriminative features.

[0057] The MSFE module made the largest contribution, enhancing global modeling. After adding the MSFE module, the AUC improved from 80.23% to 84.57% (+4.34%), accuracy from 71.69% to 80.58% (+8.89%), and the F1 score and recall were significantly improved (F1 +16.68%, recall +12.63%). This fully demonstrates that the MSFE module, through its innovative gated network (GN) and deep separable grouped convolutional design, plays a crucial role in efficiently extracting multi-scale local features (such as micro-necrosis and enhanced edges), greatly enhancing the model's ability to perceive tumor heterogeneity.

[0058] When all modules (MHSA + MSFE + SE) are added to the base model simultaneously, the performance reaches its optimum (AUC 91.36%, accuracy 87.50%), far exceeding the contribution of any single module (the largest AUC improvement comes from MSFE +4.34%, and the largest accuracy improvement comes from MSFE +8.89%). This strongly demonstrates the rationality and synergistic effect of the SE, MSFE, and MHSA module design. SE provides optimized multi-sequence feature input, MSFE efficiently extracts key local multi-scale features, and MHSA captures global contextual dependencies. The combination of the three achieves effective fusion of local details and global information, significantly improving the overall hierarchical performance of the model.

[0059] Example 3 To achieve the above objectives, based on the same inventive concept, this application also provides a pediatric MRI brain tumor grading system. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of the pediatric MRI brain tumor grading system described below can be found in the above-described limitations of the pediatric MRI brain tumor grading system, and will not be repeated here.

[0060] Specifically, it includes: the input module, which is used to preprocess brain MRI image sequences of T1, T1c, T2, Flair and ADC modes to obtain tensors of the corresponding brain MRI image sequences; The feature extraction module is used to input the tensor of the T1 mode into the SE feature extraction channel for feature extraction. The SE feature extraction channel performs feature extraction based on the SE attention mechanism. The feature extraction module includes an MSFE multi-scale feature extraction module, which is used to input the tensors of T1c, T2, Flair and ADC modes into the MSFE multi-scale feature extraction module for feature extraction. The MSFE multi-scale feature extraction module uses the following methods for feature extraction: Input the tensors of T1c, T2, Flair and ADC modes into the gated network to obtain the mode importance weights and initial feature maps corresponding to T1c, T2, Flair and ADC modes; If the modality importance weight of the corresponding modality is not lower than a certain threshold, the first feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with high importance. If the modality importance weight of the corresponding modality is lower than a certain threshold, the second feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with low importance. The feature maps obtained from the first and second feature extractions are then weighted and fused. The result output module is used to fuse the features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module, and then input the fusion into the classification module to obtain the hierarchical prediction result.

[0061] Example 4 This application discloses a computer device. The computer device includes a processor and a memory. The processor, memory, and input / output interface are connected via a system bus, and a communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the computer program is executed by the processor, it implements the steps of the pediatric MRI brain tumor grading method described in Embodiment 1.

[0062] Example 5 Based on the above embodiments, this embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the pediatric MRI brain tumor grading method described in Embodiment 1.

[0063] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for grading pediatric brain tumors using MRI, characterized in that, The method includes the following steps: Tensors for the corresponding brain MRI image sequences were obtained by preprocessing the brain MRI image sequences of T1, T1c, T2, Flair, and ADC modes. The tensor of the T1 mode is input into the SE feature extraction channel for feature extraction, and the SE feature extraction channel performs feature extraction based on the SE attention mechanism; The tensors of T1c, T2, Flair, and ADC modes are input into the MSFE multi-scale feature extraction module for feature extraction. After fusing the features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module, the feature is input into the classification module for classification prediction to obtain hierarchical prediction results. The MSFE multi-scale feature extraction module uses the following methods for feature extraction: Input the tensors of T1c, T2, Flair, and ADC modes into the gated network to obtain the mode importance weights and initial feature maps corresponding to T1c, T2, Flair, and ADC modes; If the modality importance weight of the corresponding modality is not lower than a certain threshold, then the first feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with high importance. If the modality importance weight of the corresponding modality is lower than a certain threshold, the second feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with low importance. The feature maps obtained from the first and second feature extractions are then weighted and fused.

2. The method for grading pediatric MRI brain tumors according to claim 1, characterized in that, Methods for obtaining modal importance weights and initial feature maps in gating networks include: Initial feature maps are obtained by extracting features from the tensors of the input T1c, T2, Flair, and ADC modes respectively, and the initial feature maps are then transformed into feature vectors. The gated network pre-sets four expert networks. The four expert networks first score the feature vectors of each modality, then the scores of the four expert networks are weighted and summed, and the sum is normalized. The normalized weighted sum score is the modality importance weight of each modality. Each expert network is used to evaluate a specific clinical dimension, which includes: tumor enhancement features, overall tumor morphology and surrounding edema extent, microedema and infiltrative foci, and cell density.

3. The method for grading pediatric MRI brain tumors according to claim 1, characterized in that, The second feature extraction method includes: Let the number of input channels be C. in Number of output channels C out And the number of output channels C out The number of input channels C in 2 times; Initial feature maps whose input modality importance weights are below a certain threshold; where each input channel corresponds to an initial feature map of one modality; For the initial feature map of each input channel, a 3×3×3 depthwise convolution is first performed to extract features to obtain a size H. in ×W in ×D in The input feature map is then subjected to a 1×1×1 pointwise grouped convolution. The pointwise grouped convolution includes: Get C in The input feature map after performing a 3×3×3 depthwise convolution; Based on the number of input channels C in With the number of output channels C out The greatest common divisor determines the number of groups G, and the number of input channels C. in and the number of output channels C out They were all divided into Group G; Within each group, perform a 1×1×1 pointwise convolution, using C for each group. out / G 1×1×1 convolution kernels, for the C of this group in Processing / G input channels yields a group size of H. out ×W out ×D out The output feature map; The output feature maps of all G groups are concatenated along the channel dimension, resulting in an output size of C. out ×H out ×W out ×D out The first feature map; where H out W out and D out These represent the height, width, and depth of the output feature map, respectively, and their values ​​are equal to the corresponding dimensions of the input feature map.

4. The method for grading pediatric MRI brain tumors according to claim 1, characterized in that, The method for the first feature extraction includes: The initial feature maps with modality importance weights not lower than a certain threshold are concatenated. The concatenated initial feature maps are then subjected to local feature extraction through a 3×3×3 depthwise convolution. A non-linear mapping is introduced through a non-linear activation function to enhance the model's ability to represent complex features. Then, a pooling layer is used to perform downsampling operations to achieve feature dimension compression and key information preservation, and a second feature map is output.

5. A method for grading pediatric MRI brain tumors according to claim 1, characterized in that, The multi-scale feature fusion operation includes: The features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module are aligned and embedded with positional encoding to restore their feature space structure. Then, the fused third feature map is output through the Dropout operation.

6. A pediatric MRI brain tumor grading system, characterized in that, include: The module consists of an input module, a feature extraction module, and a result output module. The input module is used to preprocess brain MRI image sequences of T1, T1c, T2, Flair and ADC modes to obtain tensors of the corresponding brain MRI image sequences. The feature extraction module is used to input the tensor of the T1 mode into the SE feature extraction channel for feature extraction. The SE feature extraction channel performs feature extraction based on the SE attention mechanism. The feature extraction module includes an MSFE multi-scale feature extraction module, which is used to input the tensors of T1c, T2, Flair and ADC modes into the MSFE multi-scale feature extraction module for feature extraction. The MSFE multi-scale feature extraction module uses the following methods for feature extraction: Input the tensors of T1c, T2, Flair, and ADC modes into the gated network to obtain the mode importance weights and initial feature maps corresponding to T1c, T2, Flair, and ADC modes; If the modality importance weight of the corresponding modality is not lower than a certain threshold, then the first feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with high importance. If the modality importance weight of the corresponding modality is lower than a certain threshold, the second feature extraction method is used to extract features from the initial feature map of the modality in order to process the features with low importance. The feature maps obtained from the first and second feature extractions are then weighted and fused. The result output module is used to fuse the features extracted by the SE feature extraction channel and the MSFE multi-scale feature extraction module, and then input the fusion into the classification module to obtain the hierarchical prediction result.

7. A computer device, characterized in that: It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; A memory for storing computer programs; a processor for executing the programs stored in the memory to implement a pediatric MRI brain tumor grading method as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements a pediatric MRI brain tumor grading method according to any one of claims 1 to 5.

Images