A brain tumor classification method, device, and medium based on magnetic resonance images

Abstract

The application provides a brain tumor classification method and device based on a magnetic resonance image, and a medium, and the method comprises the following steps: acquiring a magnetic resonance image to be processed; pre-processing the magnetic resonance image to obtain a first image after preprocessing; inputting the first image into a trained implicit high-dimensional state space hybrid network to obtain a classification result corresponding to the first image, wherein the implicit high-dimensional state space hybrid network comprises a cascaded convolution embedding layer, a backbone network and a classification module, the backbone network is composed of multiple stages, and each stage comprises a plurality of stacked state space multiplicative interaction blocks. The backbone network adopts a multi-layer state space multiplicative interaction block stacking structure, and through selective long-range aggregation and implicit high-order feature interaction, the perception ability of the tumor region and the boundary thereof is enhanced while the calculation complexity is reduced; under the premise of ensuring the lightweight of the model, the accuracy and the generalization ability of the brain tumor multi-classification task are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, device, and medium for classifying brain tumors based on magnetic resonance imaging. Background Technology

[0002] Brain tumors, due to their high rates of disability and mortality, and their strong reliance on structural imaging in clinical diagnosis, have become a key research focus. As one of the most threatening malignant diseases globally, brain tumors can severely damage the nervous system, leading to blindness, sudden death, paralysis, tumor hemorrhage, and mental disorders. With advancements in medical technology, diagnostic and treatment methods for brain tumors have become increasingly diversified. Among them, magnetic resonance imaging (MRI), with its ability to acquire high-resolution anatomical information based on the arrangement characteristics of hydrogen nuclei in an external magnetic field, has unique advantages in revealing tumor boundaries, morphology, and heterogeneity, making it a core tool for brain tumor image analysis. However, early MRI diagnosis relied primarily on manual image interpretation, which was both time-consuming and susceptible to human error. Therefore, automated brain tumor classification methods have gradually gained attention.

[0003] In early research, traditional machine learning methods such as support vector machines (SVMs) and discrete wavelet transforms were widely used for brain tumor classification. These methods made significant progress in the automatic identification of brain tumor types, providing doctors with auxiliary tools. However, due to their reliance on manual feature design, they have significant shortcomings in modeling complex structures and generalizing across datasets. In recent years, deep learning has shown significant advantages in the field of multi-classification of brain tumors, making up for some of the shortcomings of machine learning. Convolutional neural networks (CNNs) can automatically extract multi-level high-level features and are widely used for brain tumor classification. Among them, fully convolutional neural networks (WCNNs) and deep convolutional neural networks (DCNNs) can extract local features through the sliding window mechanism, but their local receptive field limitations make it difficult to model long-range dependencies. Researchers have proposed various improvements, but existing similar technologies still have several key shortcomings.

[0004] First, adaptive brain region interactions are insufficient. Current mainstream Transformer-based brain tumor classification architectures often rely on their self-attention mechanism to establish global, long-distance regional interactions across the entire MRI image, attempting to capture potential connections between different brain regions. However, not all interactions between brain regions are beneficial for classification tasks; some interactions lack discriminative power or may even be redundant. These ineffective or irrelevant interactions easily introduce interference noise during propagation, weakening the model's focus on key information. Furthermore, the Transformer architecture itself lacks the ability to filter and dynamically adjust these interactions, making it difficult to adaptively highlight effective regional dependencies according to specific task requirements. Therefore, these methods often fail to fully extract truly valuable interaction features for brain tumor classification, thus limiting the model's overall discriminative performance.

[0005] Second, there is insufficient perception of tumor regions. In noisy and blurred-boundary brain MRI scenarios, achieving detailed perception and boundary representation of tumor regions is crucial. Subtle lesions are often small in area, with blurred edges and subtle differences from surrounding tissues. Therefore, higher-order feature interactions are needed to reveal key spatial distribution patterns, rather than simple additive aggregation. Classic graph neural network-based brain tumor classification architectures are prone to noise accumulation and over-smoothing of features during information transmission, thus masking subtle differences between lesion edges and normal tissues, and failing to effectively present spatial semantic information valuable for classification. This limits the model's ability to perceive the tumor as a whole and its peripheral regions in complex MRI environments.

[0006] Third, the computational complexity is high. Mainstream Transformer-based brain tumor classification architectures excel at modeling long-range dependencies in images, capturing global features of the entire brain image through self-attention mechanisms. However, the computational complexity of traditional Transformer self-attention increases quadratically with image size, posing a significant computational bottleneck when processing high-resolution brain tumor images commonly found in clinical settings. Therefore, such models typically rely on large-scale computing resources and are difficult to apply directly to real-world clinical scenarios. Summary of the Invention

[0007] In view of the above-mentioned deficiencies of the prior art, the present invention provides a brain tumor classification method, device and medium based on magnetic resonance images to solve the technical problems of insufficient global dependency modeling and high computational complexity in brain tumor classification.

[0008] To achieve the above and other related objectives, this invention provides a brain tumor classification method based on magnetic resonance imaging (MRI), comprising: acquiring an MRI image to be processed; preprocessing the MRI image to obtain a preprocessed first image; inputting the first image into a trained implicit high-dimensional state-space hybrid network to obtain a classification result corresponding to the first image, wherein the implicit high-dimensional state-space hybrid network comprises cascaded layers: a convolutional embedding layer for initial feature embedding and spatial downsampling of the first image to obtain a first feature map; a backbone network, the backbone network consisting of multiple stages, each stage including several stacked state-space multiplicative interaction blocks, the backbone network being used to perform multi-level, progressive feature abstraction and enhancement on the first feature map, extracting deep features containing global context information and high-order interaction features, and outputting them as a second feature map; and a classification module for processing the second feature map to obtain the classification result.

[0009] In one embodiment of the present invention, the backbone network consists of four stages, each stage containing at least one state-space multiplicative interaction block.

[0010] In one embodiment of the present invention, the four stages are respectively composed of 2, 2, 8, and 4 state space multiplicative interaction blocks stacked together.

[0011] In one embodiment of the present invention, the state-space multiplicative interaction block includes: a global scanning two-dimensional operation unit for selectively performing cross-regional context aggregation on the input features of the state-space multiplicative interaction block; a first processing submodule for normalizing and convolving the output features of the global scanning two-dimensional operation unit; an implicit multiplicative interaction feature transformation unit for performing high-dimensional feature interaction on the output features of the first processing submodule; a second processing submodule for normalizing and convolving the output features of the implicit multiplicative interaction feature transformation unit; and a residual connection unit for adding the output features of the second processing submodule to the input features of the state-space multiplicative interaction block to obtain the output features of the state-space multiplicative interaction block.

[0012] In one embodiment of the present invention, the global scanning two-dimensional operation unit achieves selective cross-regional context aggregation through the following steps: the input features of the global scanning two-dimensional operation unit are expanded into one-dimensional sequences along four complementary directions to obtain four sets of feature units; for each set of feature units, a selective state-space model is used to recursively process them in their sequence directions to generate a corresponding output sequence; each set of output sequences is reverse-mapped according to its original scanning order to reconstruct a two-dimensional feature map; the two-dimensional feature maps from the four directions are fused to obtain the output features of the global scanning two-dimensional operation unit.

[0013] In one embodiment of the present invention, the four complementary directions include row forward, row reverse, column forward, and column reverse.

[0014] In one embodiment of the present invention, the implicit multiplicative interactive feature transformation unit realizes high-dimensional feature interaction through the following steps: performing two parallel linear projections on the input features of the implicit multiplicative interactive feature transformation unit to generate a first projected feature and a second projected feature, respectively; applying a nonlinear activation function to the first projected feature to generate an activated feature; multiplying the activated feature and the second projected feature element-wise to generate a high-dimensional interactive feature; and performing linear projection and channel dimension adjustment on the high-dimensional interactive feature to generate the output feature of the implicit multiplicative interactive feature transformation unit.

[0015] In one embodiment of the present invention, the classification module includes: a global pooling layer for compressing the spatial dimension of the second feature map to obtain a global feature vector; and a fully connected layer for projecting the global feature vector onto the category space to obtain the classification result.

[0016] To achieve the above and other related objectives, the present invention also provides an electronic device, including a processor, a memory, and a communication bus; the communication bus is used to connect the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the method provided in any of the above embodiments.

[0017] To achieve the above and other related objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program being used to cause a computer to perform the method provided in any of the above embodiments.

[0018] The beneficial effects of this invention are as follows: This invention proposes a brain tumor classification method, device, and medium based on magnetic resonance imaging (MRI). This method achieves efficient and accurate classification of brain tumor MRI images by constructing an implicit high-dimensional state-space hybrid network comprising convolutional embedding layers, a multi-stage backbone network, and a classification module. Specifically, the convolutional embedding layers perform initial feature embedding and spatial downsampling on the input image, effectively extracting basic features. The backbone network adopts a multi-layered state-space multiplicative interaction block stacking structure, which enhances the perception of tumor regions and their boundaries while reducing computational complexity through selective long-range aggregation and implicit high-order feature interaction. Finally, the classification module converts deep features into classification results. This method significantly improves the accuracy and generalization ability of multi-classification tasks for brain tumors while ensuring a lightweight model, providing reliable technical support for clinical auxiliary diagnosis. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The accompanying drawings are incorporated in and constitute a part of this specification, illustrating embodiments consistent with this application, and are used together with the description to explain the principles of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0020] Figure 1 A flowchart of a brain tumor classification method provided in an embodiment of the present invention;

[0021] Figure 2 This is an architecture diagram of an implicit high-dimensional state-space hybrid network provided in an embodiment of the present invention; Figure 3 This is an architecture diagram of a state-space multiplicative interaction block provided in an embodiment of the present invention; Figure 4 This is an architectural diagram of a global scanning two-dimensional operation unit provided in an embodiment of the present invention; Figure 5 This is an architectural diagram of an implicit multiplicative interactive feature transformation unit provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention; Figure 7 This is a diagram illustrating the effect of heatmap visualization in a brain tumor classification task according to an embodiment of the present invention.

[0022] Explanation of reference numerals in the attached figures: 601, processor; 602, memory. Detailed Implementation

[0023] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other. In addition to the specific methods, equipment, and materials used in the embodiments, based on the knowledge of the prior art and the description of the present invention by those skilled in the art, any prior art methods, equipment, and materials similar to or equivalent to the methods, equipment, and materials in the embodiments of the present invention can be used to implement the present invention.

[0024] It should be understood that the terminology used in the embodiments of this invention is for describing specific particular implementations and not for limiting the scope of protection of this invention. Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art.

[0025] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In some embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0026] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functions, and operations that may be implemented in the methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0027] Please see Figure 1 , Figure 1 A brain tumor classification method based on magnetic resonance imaging is provided in one embodiment of the present invention, including steps S101 to S103.

[0028] Step S101: Obtain the magnetic resonance image (also known as MRI image) to be processed.

[0029] Step S102: Preprocess the magnetic resonance image to obtain the preprocessed first image. Preprocessing is a key step to ensure the accuracy and effectiveness of the model. To maintain the consistency of data quality, the following conventional preprocessing operations are often used.

[0030] Normalization: Normalization can bring MRI images to a uniform standard in terms of resolution and contrast, thereby eliminating differences caused by different imaging sources.

[0031] Field correction: Due to hardware limitations or imaging principles, MRI scanners may capture images with slowly varying brightness inhomogeneities, known as "field inhomogeneity" or "intensity inhomogeneity." This causes differences in grayscale values ​​for the same tissue at different locations, affecting the accuracy of subsequent analysis. The purpose of field correction is to eliminate this inhomogeneity, ensuring that the signal intensity of the same tissue remains consistent across the entire image range, thereby significantly improving model performance and robustness. Field correction can be implemented using mature algorithms or software tools, such as the N4ITK algorithm, which is integrated into many commonly used medical image processing platforms (such as ANTs and SPM).

[0032] Reshaping and Registration: This operation aligns all images to a unified template. Reshaping primarily involves changing the spatial dimensions (resolution) and orientation of the images; for example, adjusting all input images to a uniform size required by the model (e.g., 224×224). Different individuals' brains naturally differ in shape, size, and location. Registration eliminates these anatomical differences, allowing images from different subjects to be analyzed comparablely within the same standard space. Registration typically aligns brain images from all individuals to a recognized standard brain atlas (e.g., the MNI space template). This ensures that each pixel represents roughly the same brain region location across different individuals, significantly enhancing the comparability between samples.

[0033] This invention utilizes a pre-trained implicit high-dimensional state-space hybrid network. Therefore, training this network is necessary in the initial stages. The training dataset contains images from different patients, which may have been acquired in different hospitals, using different MRI machines, and with different scanning parameters (such as slice thickness and resolution). After registration and alignment, the anatomical structures of the brain in different images are completely corresponding in pixel coordinates. This allows the model to accurately learn and fuse information from different patients.

[0034] Random shuffling: This operation is not necessary for the magnetic resonance images to be processed; however, during the training of the above network model, the samples in the training and test sets can be randomly shuffled to evaluate the model's generalization ability and robustness.

[0035] Step S103: Input the first image into the trained implicit high-dimensional state space hybrid network to obtain the classification result corresponding to the first image.

[0036] In this step, the architecture diagram of the Implicit High-Dimensional State-Space Hybrid Network (IHSHN) is as follows: Figure 2As shown, it includes cascaded convolutional embedding layers, a backbone network, and a classification module. The convolutional embedding layers are used to perform initial feature embedding and spatial downsampling on the first image to obtain a first feature map. The backbone network consists of multiple stages, each including several stacked state-space multiplicative interaction blocks. The backbone network is used to perform multi-level, progressive feature abstraction and enhancement on the first feature map, extracting deep features containing global context information and high-order interaction features, and outputting them as a second feature map. The classification module processes the second feature map to obtain the classification result. Compared with the traditional Transformer architecture, the implicit high-dimensional state-space hybrid network in this invention reduces the computational complexity of long-range dependency modeling from quadratic to linear, avoiding the computational bottleneck under high-resolution MRI images, and possessing greater potential for practical clinical deployment.

[0037] In a specific embodiment of the present invention, the convolutional embedding layer is a convolutional Stem, the structure of which can be, for example, composed of a 7×7 convolution, a batch normalization layer and a ReLU activation function connected in sequence; wherein, the stride of the 7×7 convolution is 2 and the padding is 3, used to downsample the first image with an input size of 224×224×3 to a first feature map of 56×56×32, significantly reducing the spatial resolution while expanding the channel dimension, providing the subsequent backbone network with an initial feature representation that combines local details and basic semantics.

[0038] In a specific embodiment of the present invention, the backbone network consists of four stages, each containing at least one state-space multiplicative interaction block. The first three stages also include downsampling layers. Features are modeled and abstracted progressively through these four stages. The first three stages continuously double the channel dimension while progressively downsampling, thereby enhancing semantic abstraction capabilities layer by layer. The fourth stage maintains the spatial resolution and focuses on further aggregating and strengthening discriminative feature representations. This structure ensures that the network can capture tumor features at different scales. The shallow stages focus on local texture and edges, while the deep stages integrate the global context to form a comprehensive understanding of the overall morphology of the tumor and its relationships with surrounding tissues, thus comprehensively improving the model's discriminative ability.

[0039] In a specific embodiment of the present invention, the four stages are composed of stacked state-space multiplicative interaction blocks of 2, 2, 8, and 4, respectively. The number of state-space multiplicative interaction blocks in each stage is an optimized trade-off design. The first two stages use fewer blocks to quickly reduce dimensionality and capture basic features, avoiding overcomputation at low levels. The third stage significantly increases to 8 blocks to perform deep modeling at the current resolution, focusing on extracting and refining the most critical discriminative features. The final stage appropriately reduces to 4 blocks to further aggregate high-level semantic information, preparing a robust and compact feature representation for the final classification output, thereby achieving the best balance between model performance and computational efficiency.

[0040] Please see Figure 3 In a specific embodiment of the present invention, the state-space multiplicative interaction block includes a global scan 2D operation unit (Global Scan 2D), a first processing submodule (DW-Conv+BatchNorm), an implicit multiplicative interaction feature transformation unit (including dashed boxes of three fully connected functions), a second processing submodule (BatchNorm+DW-Conv), and a residual connection unit. The global scan 2D operation unit is used to selectively aggregate cross-regional context on the input features of the state-space multiplicative interaction block; the first processing submodule is used to normalize and convolve the output features of the global scan 2D operation unit; the implicit multiplicative interaction feature transformation unit is used to perform high-dimensional feature interaction on the output features of the first processing submodule; the second processing submodule is used to normalize and convolve the output features of the implicit multiplicative interaction feature transformation unit; and the residual connection unit is used to add the output features of the second processing submodule to the input features of the state-space multiplicative interaction block to obtain the output features of the state-space multiplicative interaction block.

[0041] In this embodiment, a state-space multiplicative interaction block constructs an efficient feature processing pipeline through a combination of cascading and residual connections. The two-dimensional selective scanning mechanism (i.e., the global scanning two-dimensional operation unit) can perform state-space recursion in four directions (rows / columns), adaptively filtering redundant interactions with linear complexity and efficiently aggregating global dependencies between brain regions, thereby avoiding interference from invalid features. Subsequently, the first processing submodule performs regularization and nonlinear transformation on the aggregated features. The implicit multiplicative interaction feature transformation unit maps the features to a multinomial space, capturing high-dimensional dependencies through element-wise multiplication, significantly improving the model's sensitivity to small lesions and ambiguous boundaries, while suppressing redundant anatomical signals. The second processing submodule further optimizes the feature representation, and finally, the residual connection unit fuses the processed features with the original input, effectively alleviating the gradient vanishing problem in deep networks, ensuring the smooth transmission of key information, and thus significantly enhancing the module's ability to model complex tumor patterns.

[0042] In this embodiment, the state-space multiplicative interaction block (SMB) composed of these units unifies selective long-range aggregation and implicit higher-order interaction in a single unit, taking into account both global dependencies and fine-grained discrimination modes, effectively improving the model's ability to model complex tumor morphologies.

[0043] Please see Figure 4 In a specific embodiment of the present invention, the global scanning two-dimensional operation unit achieves selective cross-region context aggregation through the following steps 4.1 to 4.4.

[0044] Step 4.1: Unfold the input features of the global scanning two-dimensional operation unit into a one-dimensional sequence along four complementary directions to obtain four sets of feature units, which can be expressed by the formula: , , , .

[0045] Step 4.2: For each set of feature units, recursively process them in their sequence direction using a selective state-space model to generate the corresponding output sequence.

[0046] For each path P, the sequence feature unit The process is handled by the core of the global scan two-dimensional operation unit—the selective state-space model. This model maintains a hidden state along the scan direction. This is used to gradually accumulate and update context information. , , In the formula, , , The input correlation coefficients are generated by the lightweight projection. This represents the context output at time step t on path P. When P is row-f, the output sequence obtained after this step is: By analogy, the output sequences corresponding to other paths can be obtained.

[0047] Step 4.3: Reverse the original scanning order of each group of output sequences to reconstruct a two-dimensional feature map. The reconstructed two-dimensional feature map can be denoted as Y. row-f Y row-b Y col-f Y col-b .

[0048] Step 4.4: Fuse the two-dimensional feature maps from the four directions to obtain the output features of the global scanning two-dimensional operation unit. This step can be expressed by the formula: .

[0049] In one specific embodiment of the present invention, the four complementary directions include row-forward, row-backward, column-forward, and column-backward. Through these four scanning methods, the model can selectively filter irrelevant responses along sequence directions, capture directional dependencies across brain regions, and aggregate global tumor context information with discriminative value.

[0050] Please see Figure 5In a specific embodiment of the present invention, the implicit multiplicative interaction feature transformation unit realizes high-dimensional feature interaction through the following steps: (1) performing two parallel linear projections (corresponding to) the input features of the implicit multiplicative interaction feature transformation unit. Figure 3 (1) Generate the first projection feature and the second projection feature respectively from the two FCs above; (2) Apply a nonlinear activation function to the first projection feature to generate the activated feature; (3) Multiply the activated feature and the second projection feature element by element to generate the high-dimensional interactive feature; (4) Perform linear projection and channel dimension adjustment on the high-dimensional interactive feature (corresponding to the two FCs above); Figure 3 The next FC in the model generates the output features of the implicit multiplicative interactive feature transformation unit.

[0051] The above steps can be expressed by the following formula: , U=ReLU6(W u Y'), V=W v Y', In the formula, , U is the activated feature, V is the second projected feature, and the feature Y' is obtained after the output feature Y of the global scanning two-dimensional operation unit is processed by the first processing submodule.

[0052] In a specific embodiment of the present invention, the classification module includes a global pooling layer and a fully connected layer. The global pooling layer is used to compress the spatial dimension of the second feature map to obtain a global feature vector. The fully connected layer is used to project the global feature vector onto the category space to obtain the classification result. The global pooling layer compresses the second feature map output by the backbone network into a one-dimensional global feature vector by performing global average pooling in the spatial dimension. This operation not only significantly reduces the number of parameters and prevents overfitting, but also makes the model translation invariant, focusing more on the global distribution of features rather than their specific locations. Subsequently, the fully connected layer acts as a classifier, linearly projecting the global feature vector onto a dimensional space with the same number of brain tumor categories, and finally outputting the predicted probability of each category, i.e., the classification result. This design greatly improves the simplicity and computational efficiency of the model while ensuring classification performance.

[0053] It should be noted that the steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they contain the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0054] Please see Figure 6 , Figure 6 An electronic device provided in one embodiment of the present invention includes a processor 601, a memory 602, and a communication bus; the communication bus is used to connect the processor 601 and the memory 602; the processor 601 is used to execute a computer program stored in the memory 602 to implement the above-mentioned brain tumor classification method based on magnetic resonance images.

[0055] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program being used to cause a computer to execute the above-described brain tumor classification method based on magnetic resonance images.

[0056] To verify the effectiveness of the implicit high-dimensional state-space hybrid network in this invention, extensive experimental evaluations were conducted on two publicly available benchmark datasets.

[0057] The Figshare dataset is a widely used brain tumor imaging dataset containing 3,064 T1-weighted enhanced brain MRI slices from 233 patients, covering three types of tumors: pituitary adenoma, glioma, and meningioma.

[0058] The Brain Tumor MRI Images (BTMI) database is a curated collection of brain tumor images, including T1, enhanced T1, and T2 MRI images, covering 44 different types of brain tumors.

[0059] The experiments were conducted on an NVIDIA GeForce RTX 3090 platform using PyTorch 1.10.2 and the AdamW optimizer. The optimal settings on the Figshare dataset were a learning rate of 0.00025 and 190 training epochs; while on the more challenging 44-class BTMI dataset, the optimal configuration was a learning rate of 0.00025 and 200 training epochs. All hyperparameters were tuned using a consistent strategy to ensure the comparability and fairness of the experiments.

[0060] The network model in this invention was compared with state-of-the-art methods on two benchmark brain tumor datasets. The comparison results on the Figshare dataset are shown in Table 1.

[0061] Table 1: Comparison of various methods on the Figshare dataset

[0062] On the Figshare dataset, our method outperforms 11 state-of-the-art deep learning models proposed in recent years (see Table 1). Compared to the best CNN model (Optimized CNN), the accuracy is improved by 1.30%; compared to Transformer-based Ensemble ViT, our method exhibits stronger discriminative power by dynamically and selectively modeling global dependencies. Compared to the state-of-the-art HLAE method, which relies on additive aggregation, our SMB module achieves superior performance by enhancing discriminative co-occurrence relationships and suppressing noise through multiplicative interactions.

[0063] The comparison results on the BTMI dataset are shown in Table 2.

[0064] Table 2: Comparison of various methods on the BTMI dataset

[0065] On the more challenging BTMI dataset (containing 44 imbalanced classes), we compared our method with seven models proposed in recent years. As shown in Table 2, our method achieved a top accuracy of 98.09%, a 1.44% improvement over EfficientNetB3 (96.65%), and remained the leader among all compared methods. The results demonstrate that IHSHN maintains robust accuracy and strong generalization ability even with a significant increase in the number of tumor classes.

[0066] Figure 7 The image shows the effect of the present invention in heatmap visualization for brain tumor classification tasks. Two key conclusions can be drawn from this image: (1) the model can accurately highlight tumor-related regions and effectively suppress irrelevant regions; (2) the highlighted regions contain rich semantic information necessary for classification. It is worth noting that some normal tissue regions are also activated, which suggests that the model may use surrounding tissues as contextual references to better characterize tumor boundaries and their structural features.

[0067] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A brain tumor classification method based on magnetic resonance imaging, characterized in that, include: Acquire the magnetic resonance image to be processed; The magnetic resonance image is preprocessed to obtain a preprocessed first image; The first image is input into a trained implicit high-dimensional state-space hybrid network to obtain the classification result corresponding to the first image. The implicit high-dimensional state-space hybrid network comprises cascaded components: A convolutional embedding layer is used to perform initial feature embedding and spatial downsampling on the first image to obtain a first feature map; The backbone network, consisting of multiple stages, each including several stacked state-space multiplicative interaction blocks, is used to perform multi-level, progressive feature abstraction and enhancement on the first feature map, extracting deep features containing global context information and high-order interaction features, and outputting them as a second feature map; and The classification module is used to process the second feature map to obtain the classification result; The state-space multiplicative interaction block includes: A global scanning two-dimensional operation unit is used to selectively aggregate cross-regional contextual features of the input features of the state space multiplicative interaction block; The first processing submodule is used to normalize and convolution the output features of the global scanning two-dimensional operation unit. An implicit multiplicative interactive feature transformation unit is used to perform high-dimensional feature interaction on the output features of the first processing submodule. The second processing submodule is used to normalize and convolutionally process the output features of the implicit multiplicative interactive feature transformation unit; and The residual connection unit is used to add the output features of the second processing submodule to the input features of the state space multiplicative interaction block to obtain the output features of the state space multiplicative interaction block. Furthermore, the global scanning two-dimensional operation unit achieves selective cross-regional context aggregation through the following steps: the input features of the global scanning two-dimensional operation unit are expanded into one-dimensional sequences along four complementary directions: row forward, row backward, column forward, and column backward, respectively, to obtain four sets of feature units; for each set of feature units, a selective state-space model is used to recursively process them in their sequence directions to generate a corresponding output sequence; each set of output sequences is reverse-mapped according to its original scanning order to reconstruct a two-dimensional feature map; the two-dimensional feature maps from the four directions are fused to obtain the output features of the global scanning two-dimensional operation unit. The implicit multiplicative interactive feature transformation unit achieves high-dimensional feature interaction through the following steps: performing two parallel linear projections on the input features of the implicit multiplicative interactive feature transformation unit to generate a first projected feature and a second projected feature, respectively; applying a nonlinear activation function to the first projected feature to generate an activated feature; multiplying the activated feature and the second projected feature element-wise to generate a high-dimensional interactive feature; and performing linear projection and channel dimension adjustment on the high-dimensional interactive feature to generate the output feature of the implicit multiplicative interactive feature transformation unit.

2. The brain tumor classification method based on magnetic resonance imaging according to claim 1, characterized in that, The backbone network consists of four stages, each stage containing at least one state-space multiplicative interaction block.

3. The brain tumor classification method based on magnetic resonance imaging according to claim 2, characterized in that, The four stages are respectively composed of 2, 2, 8, and 4 multiplicative interaction blocks in the state space.

4. The brain tumor classification method based on magnetic resonance imaging according to claim 1, characterized in that, The classification module includes: A global pooling layer is used to compress the spatial dimension of the second feature map to obtain a global feature vector; and A fully connected layer is used to project the global feature vector onto the category space to obtain the classification result.

5. An electronic device, characterized in that, It includes a processor, a memory, and a communication bus; the communication bus is used to connect the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the method as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that, It stores a computer program thereon, the computer program being used to cause the computer to perform the method as described in any one of claims 1 to 4.

Images