Multi-modal MRI (Magnetic Resonance Imaging) brain tumor segmentation method based on convolution attention

Abstract

The invention relates to a multi-modal MRI brain tumor segmentation method based on convolution attention, and the method comprises the steps: carrying out the parallel coding of multi-modal medical image data, and extracting the specific low-layer space features of each modal; performing feature fusion and high-level coding on the low-level spatial features to obtain shared features; carrying out convolution attention processing on the shared features to obtain enhanced features; and carrying out decoding processing on the enhanced features to obtain a segmentation result. According to the method, more accurate tumor boundary segmentation and tissue differentiation can be realized in medical image analysis such as brain MRI (Magnetic Resonance Imaging), meanwhile, the parameter efficiency, the calculation speed and the generalization ability under limited data of the model are remarkably improved, and the clinical practicability is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and medical image analysis, in particular to a multi-modal MRI brain tumor segmentation method based on convolution attention. BACKGROUND

[0002] Currently, multi-modal medical image analysis mainly relies on convolutional neural networks and standard Transformer architectures. Convolutional neural networks can effectively extract local features, but have inherent limitations in modeling long-range global dependencies. The self-attention mechanism of the standard Transformer makes up for this deficiency and can capture global context information, but its direct application to medical image processing faces significant challenges.

[0003] Attention mechanism is a core breakthrough in the field of deep learning, aiming to solve the information focusing problem in sequence modeling. This mechanism calculates the correlation weights between elements, allowing the model to dynamically focus on key information in the input. Its development has evolved from a task-specific tool to a general architecture. In the machine translation task, the attention mechanism first showed its value. The traditional encoder-decoder framework faces the problem of information loss in long sequences, and attention allows the model to adaptively select important information by establishing a direct connection between the decoding end and the encoding end, significantly improving the ability to process long sequences. The structure of the traditional Transformer attention module is shown in Figure 1 .

[0004] The linear transformation used by the core module of the standard Transformer when generating query, key and value vectors is point-wise independent. This process completely ignores the inherent spatial continuity and local correlation between pixels or voxels in medical images. For example, the smooth boundaries of organs, the morphological structure of lesions and other key visual patterns are not fully considered in the initial stage of feature extraction, resulting in insufficient model perception of subtle anatomical structures.

[0005] In addition, the standard self-attention mechanism is a data-driven method that lacks inductive bias. In the field of medical images, high-quality, large-scale labeled data is usually scarce and costly to obtain. This mechanism is difficult to learn effectively in the absence of data and is prone to overfitting, which limits the model's generalization performance and robustness in real clinical scenarios.

[0006] From the perspective of computational efficiency, the computational complexity of standard self-attention is proportional to the square of the length of the input sequence. When dealing with high-resolution three-dimensional medical volume data, the sequence length is extremely large, which brings heavy computational and memory burden, which poses a practical obstacle to its deployment and application on conventional hardware.

[0007] Existing technical improvements are mostly focused on introducing locality constraints after attention computation, or simply concatenating or parallelizing convolution modules with attention modules. These methods do not fundamentally innovate the way attention mechanisms are generated, and fail to efficiently and organically inject locality perception into the source of attention formation. Therefore, there is an urgent need for a new attention generation method that can fuse local context from the source, be more efficient, and be more suitable for the inherent characteristics of medical images.

[0008] Transformer model and self-attention mechanism: First, the success of the Transformer model in natural language processing and its core component, the self-attention mechanism, are introduced. The ability to capture global context by computing the correlation between all element pairs (dot product of Query and Key) is described. This mechanism projects the input sequence X ∈ RN×C through three independent linear transformation matrices WQ, WK, WV ∈ RC×d to generate corresponding query, key, and value vector sequences: , , ; Then, by computing the similarity of the dot product of the query and all keys, the attention weights are obtained after Softmax normalization, and the value vectors are weighted and summed to finally output a new sequence representation that integrates global context: ; This "full connection" interaction allows any two elements in the sequence to be directly associated, giving the model unparalleled global context modeling capabilities, which is the key to its success.

[0009] Pain points of traditional QKV generation and segmentation: High computational and memory complexity: The traditional method uses three independent linear layers to project the input sequence X ∈ R^(N×d) to Q, K, V ∈ R^(N×d)`. The self-attention computation complexity is O(N²·d), where N is the sequence length. When processing high-resolution images (sequence length N=H×W is large), this becomes a computational bottleneck.

[0010] Insensitive to local structure: Linear transformation is point-wise independent, and when generating Q, K, V for each position in the sequence, the local context of the surrounding pixels / tokens is not considered. For image, video and other data, this loses the crucial local patterns (such as edges, textures).

[0011] Parameter inefficiency: uses three independent, large linear layers, large parameter quantity, and prone to overfitting on small datasets.

[0012] Lack of inductive bias: self-attention itself has no built-in prior knowledge about the spatial structure of images, and requires a large amount of data to learn these basic rules.

[0013] Brain tumor segmentation is crucial for neurosurgical planning and prognosis evaluation. Clinically, multiple MRI sequences (modalities) are routinely used: T1c (T1-weighted enhancement): clearly shows the enhanced tumor core region.

[0014] T2 (T2-weighted): clearly shows the peritumoral edema zone.

[0015] FLAIR (Fluid Attenuated Inversion Recovery): most sensitive to show tumor invasion and edema range.

[0016] Different sequences provide complementary but different anatomical and pathological information. The boundaries, contrast, and morphology of the tumor on different sequences may not be consistent. Existing segmentation networks based on standard Transformers lack spatial local continuity prior when generating Q, K, V for each modality feature, resulting in insufficient ability to fuse multi-modal information to determine accurate, biologically meaningful tumor boundaries (such as distinguishing edema from tumor infiltration), and the segmentation results may be rough or inconsistent.

[0017] Existing solutions mainly fall into two categories, but both have obvious shortcomings.

[0018] One type of solution aims to improve the standard attention itself to improve efficiency, such as sparse attention and linear attention. They reduce complexity by limiting the calculation range or mathematical approximation, but often at the expense of global modeling ability or expression ability, and do not solve the fundamental problem of feature generation lacking locality.

[0019] Another type of solution attempts to combine convolution and attention, such as adding convolution layers in parallel or in series in the Transformer block. These methods essentially simply stack the two structures, with convolution modules and attention modules independent of each other, failing to achieve deep functional integration. The local features extracted by convolution and the global context generated by attention have a gap in information interaction, and the computational overhead is usually large.

[0020] In summary, the current solutions have not been able to efficiently and organically inject local perception ability from the generation source of the attention mechanism, i.e., the formation stage of the query, key, and value vectors, which is a core defect of existing technology. SUMMARY

[0021] The application aims to provide a multi-modal MRI brain tumor segmentation method based on convolutional attention, to solve three core problems of current Transformer models in medical image segmentation tasks: lack of local structure-sensitive feature extraction capability, low efficiency of global attention mechanism in multi-modal fusion, and insufficient model generalization performance due to data scarcity. By replacing the traditional linear transformation with a convolution operation to generate query, key and value vectors, the application aims to inject local spatial context and image induction bias into the attention calculation process from the source, thereby building a more efficient model, more accurate feature fusion, and more fully utilizing medical image prior knowledge, and ultimately achieving significant improvement in multi-modal medical image segmentation accuracy and robustness.

[0022] To achieve the above object, the application provides the following scheme: The multi-modal MRI brain tumor segmentation method based on convolutional attention comprises: Parallelly encoding the multi-modal medical image data to extract low-level spatial features specific to each modality; Performing feature fusion and high-level encoding on the low-level spatial features to obtain shared features; Performing convolutional attention processing on the shared features to obtain enhanced features; Performing decoding processing on the enhanced features to obtain a segmentation result.

[0023] Optionally, the multi-modal medical image data comprises at least two of a T1-weighted enhanced image, a T2-weighted image and a FLAIR image in brain magnetic resonance imaging.

[0024] Optionally, the extraction of low-level spatial features specific to each modality comprises: Inputting image data of each modality in the multi-modal medical image data into an independent encoder branch to extract low-level spatial features specific to each modality; wherein each encoder branch comprises a plurality of convolutional layers, an instance normalization layer and a LeakyReLU activation layer.

[0025] Optionally, the shared features are obtained by: Concatenating the low-level spatial features in the channel dimension at a preset depth of the encoder; Inputting the merged features into a shared encoding layer to further extract abstract features and obtain the shared features.

[0026] Optionally, the convolutional attention processing of the shared features comprises: Performing convolutional projection on the shared features to obtain an intermediate feature tensor; Dividing the intermediate feature tensor in the channel dimension to obtain a query tensor, a key tensor and a value tensor; perform context-aware attention calculation based on the query tensor, the key tensor and the value tensor, to obtain enhanced features.

[0027] Optionally, the convolutional projection on the shared features comprises: inputting the shared features into a shared convolutional layer for convolution operation, wherein the shared convolutional layer uses a small convolutional kernel of 3*3*3 and sets padding to keep the spatial size unchanged, and the output channel number is set to 3*d, d being the planned projection dimension of the attention embedding.

[0028] Optionally, the context-aware attention calculation based on the query tensor, the key tensor and the value tensor comprises: reconstructing Q_conv, K_conv and V_conv into a sequence form [B, N, d], wherein Q_conv, K_conv and V_conv are the query tensor, the key tensor and the value tensor respectively, each dimension being [B, d, H', W', D'], N = H' W' D' is the sequence length, B is the batch size, d is the planned projection dimension of the attention embedding, H' is the height, W' is the width, and D' is the depth; performing standard multi-head self-attention calculation based on the reconstructed sequence form.

[0029] Optionally, the decoding processing on the enhanced features comprises: inputting the enhanced features into a decoder for upsampling and combining with low-level spatial features to generate a segmentation result map, wherein the decoder comprises a plurality of upsampling layers and convolutional layers, and each level of the decoder is connected through a jump connection to fuse the multi-modal low-level spatial features from the corresponding level of the encoder.

[0030] The present application has the following advantages: The present application realizes multiple beneficial effects in medical image segmentation through the core design of generating query, key and value vectors by convolution. The convolution operation injects local induction bias into the attention mechanism from the source, so that each Q, K and V vector generated fuses the multi-modal context information of the surrounding voxels. This forces the model to compare between local semantic regions when calculating the attention weight, rather than the traditional point-to-point comparison of isolated voxels, so that it can more accurately capture subtle structures such as tumor boundaries.

[0031] In terms of efficiency, a shared compact convolutional layer replaces the traditional three independent massive linear layers, significantly reducing the model parameter amount and mitigating the overfitting risk. At the same time, thanks to the highly parallel optimization characteristics of convolutional operations on hardware such as GPUs, this design can usually obtain better computational speed and memory efficiency than the theoretical complexity in actual deployment, especially for high-resolution three-dimensional medical images.

[0032] This method introduces the inherent advantages of convolution, such as translation invariance, into the attention mechanism, improving the robustness of the model to changes in target position in the image. This module itself is designed as plug-and-play and can be seamlessly integrated into existing Transformer-based architectures as a standardized component without modifying the overall framework, facilitating rapid deployment and application verification in various medical image analysis tasks. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0034] Figure 1 A schematic diagram of the structure of a traditional Transformer attention module; Figure 2 A schematic diagram of the structure of a convolution-based QKV generation module according to an embodiment of the present application; Figure 3 A detailed structure schematic diagram of scheme A according to an embodiment of the present application; Figure 4 A convolutional attention processing flow step schematic diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0036] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0037] The multi-modal MRI brain tumor segmentation method based on convolutional attention includes: Parallel encoding is performed on the multi-modal medical image data, and low-level spatial features specific to each modality are extracted; Feature fusion and high-level encoding are performed on the low-level spatial features to obtain shared features; Convolutional attention processing is performed on the shared features to obtain enhanced features; Decoding processing is performed on the enhanced features to obtain segmentation results.

[0038] Further, the multi-modal medical image data includes at least two of a T1-weighted enhanced image, a T2-weighted image, and a FLAIR image in brain magnetic resonance imaging.

[0039] Specifically, the entity of the embodiment is a deep neural network for multi-modal medical image (such as brain MRI) segmentation. The network takes U-Net++ as the basic skeleton, but makes a key improvement: replacing the bottleneck module at the bottom of the encoder with a self-developed Conv-QKV Transformer module. This module is the core of the network for multi-modal information fusion and global context modeling. The embodiment applies it to the fusion, analysis, and precise segmentation tasks of multi-modal medical images (such as magnetic resonance imaging), aiming to improve the recognition accuracy and boundary division capability of the model for regions with complex anatomical structures and pathological features (such as brain tumors and their sub-regions).

[0040] Further, the extraction of low-level spatial features specific to each modality includes: The image data of each modality in the multi-modal medical image data is input into an independent encoder branch to extract low-level spatial features specific to each modality. Each encoder branch includes multiple convolutional layers, instance normalization layers, and LeakyReLU activation layers.

[0041] Further, obtaining the shared features includes: Channel dimension splicing is performed on the low-level spatial features at a preset depth of the encoder; The merged features are input into a shared encoding layer to further extract abstract features and obtain the shared features.

[0042] Further, the convolutional attention processing of the shared features includes: Convolutional projection is performed on the shared features to obtain an intermediate feature tensor; The intermediate feature tensor is divided in the channel dimension to obtain a query tensor, a key tensor, and a value tensor; Context-aware attention calculation is performed based on the query tensor, the key tensor, and the value tensor to obtain enhanced features.

[0043] Further, the convolutional projection of the shared features includes: The shared features are input into a shared convolutional layer for convolution operation; wherein the shared convolutional layer uses a small convolutional kernel of 3x3x3, and padding is set to keep the spatial size unchanged; the output channel number thereof is set to 3xd, and d is the planned projection attention embedding dimension.

[0044] Further, the context-aware attention calculation based on the query tensor, key tensor and value tensor comprises: Q_conv, K_conv, V_conv are reconstructed into a sequence form [B, N, d]; Based on the reconstructed sequence form, a standard multi-head self-attention calculation is performed.

[0045] The standard multi-head self-attention calculation is: Let the input sequence X∈Rn dmodel, wherein n is the sequence length, dmodel is the model width, a hyperparameter h (number of heads) is introduced and dmodel is required to be divisible by h, and a single head dk = dv = dmodel / h is defined. Linear projection: , , ; Multi-head split: Q, K, V are evenly divided along the feature dimension and reshaped to obtain: Q, K, V∈R h n dk ; Attention calculation: ; = concat (O1,...,Oh) Wo∈Rn dmodel.

[0046] Further, the decoding processing of the enhanced features comprises: The enhanced features are input into a decoder for upsampling, and a segmentation result map is generated in combination with low-level spatial features; wherein the decoder comprises a plurality of upsampling layers and convolutional layers, and each level of the decoder is connected by a jump connection for fusing multi-modal low-level spatial features from the corresponding level of the encoder.

[0047] Specifically, in the embodiment, the data processing and feature extraction process is: Input: The network receives registered multi-modal three-dimensional image patches (e.g., T1c, T2, FLAIR sequences of brain MRI). The input tensor has dimensions [B, M, H, W, D], where B is the batch size, M is the number of modalities (e.g., 3), H, W, D are the height, width, and depth, respectively.

[0048] Multi-modal parallel encoding: The images of each modality are first passed through an independent encoder branch. Each branch consists of multiple convolutional layers, instance normalization layers, and LeakyReLU activation layers, responsible for extracting low-level spatial features specific to each modality.

[0049] Feature fusion and high-level encoding: The feature maps extracted by each modality branch are concatenated in the channel dimension at a certain depth of the encoder (e.g., after four downsampling stages). The merged feature maps are then fed into subsequent shared encoding layers to further extract abstract features.

[0050] Specifically, the implementation details of the core module in this embodiment: Conv-QKV Transformer: Module input: The shared feature map from the deepest layer of the encoder, with dimensions [B, C, H', W', D'] (where H', W', D' are the spatial dimensions after downsampling).

[0051] Convolutional projection: The feature map is input into a shared three-dimensional convolutional layer. The convolutional layer uses a small 3x3x3 convolutional kernel and is padded to maintain the spatial dimensions. The number of output channels is set to 3xd, where d is the planned embedding dimension of the attention.

[0052] Generating Q, K, V: The feature map output by the above convolutional layer is evenly divided into three parts in the channel dimension. These three feature maps are directly used as query, key, and value tensors without any other linear transformation, denoted as Q_conv, K_conv, V_conv, each with dimensions [B, d, H', W', D'].

[0053] Context-aware attention calculation: Reconstruct Q_conv, K_conv, V_conv into sequence form [B, N, d], where N = H' W' D' is the sequence length.

[0054] Perform standard multi-head self-attention calculation. Since each position in Q, K, V already contains the context information of its local 3x3x3 neighborhood, the attention mechanism at this time is essentially calculating the semantic association between different local regions, rather than the association of original pixels.

[0055] Specifically, in this embodiment, the decoding and outputting are as follows: Feature decoding: the features enhanced by the Conv-QKV Transformer module are fed into the decoder. The decoder is composed of multiple up-sampling layers and convolutional layers, which gradually restore the spatial resolution of the feature map.

[0056] Skip connection: each level of the decoder is connected by a skip connection, which fuses the multi-modal shallow features from the corresponding level of the encoder to recover the detailed information.

[0057] Segmentation map generation: the last layer of the network is a 1x1x1 convolutional layer followed by a Softmax activation function. It maps the feature map to the same number of channels as the number of segmentation classes, outputting a probability map for each voxel belonging to each class (such as background, edema, tumor core, etc.).

[0058] Model training and validation scheme: Loss function: the weighted sum of Dice loss and cross-entropy loss is used as the objective function to address the problem of severe imbalance between foreground and background voxels in medical images.

[0059] Training strategy: use the AdamW optimizer for end-to-end training. Comparative experiments can be designed to compare with baseline models using standard Transformer modules or pure CNNs on public brain tumor segmentation datasets.

[0060] Evaluation indicators: mainly use Dice similarity coefficient and 95% Hausdorff distance to quantitatively evaluate the segmentation accuracy of the model on tumor core, enhanced tumor, edema, etc. sub-regions.

[0061] To address the problems of weak local perception, low computational efficiency, and strong data dependence of existing Transformer models in medical image segmentation, the core of this embodiment is to propose a convolutional attention mechanism, as shown in Figure 4 The mechanism uses a shared convolutional layer to directly process the input features and splits the output in the channel dimension, generating query, key, and value vectors that integrate local spatial context. This design introduces strong inductive bias such as locality and shift invariance from the source. When embedding this module as the core into multi-modal segmentation networks such as U-Net++, it can achieve more accurate tumor boundary segmentation and tissue differentiation in medical image analysis such as brain MRI, while significantly improving the parameter efficiency, computational speed, and generalization ability on limited data, enhancing the clinical practicability.

[0062] The improved U-Net++ architecture in the embodiment is described in detail below. The core is to replace the bottleneck module at the deepest layer of the network with a Conv-QKV Transformer module to achieve more efficient local context modeling in the multi-modal feature fusion stage. The entire processing procedure is divided into three steps: Step one: multi-modal feature extraction and deep layer fusion coding: First, input the registered T1c, T2 and FLAIR three-modal brain images. At the shallow layer, each modality extracts low-level features through an independent convolutional path. At the deepest layer of the encoder, the feature maps of the three modalities are channel spliced and input into the core Conv-QKV Transformer module.

[0063] The core operation of this module is to use a shared 3x3x3 three-dimensional convolutional layer to directly generate the query, key and value tensors for the spliced multi-modal features. This is not a simple linear transformation, but forces the model to integrate the context information of all modalities in its three-dimensional spatial neighborhood (about 1-2 voxels) when generating the attention vector of each voxel. For example, to distinguish a voxel that is high signal on FLAIR and not enhanced on T1c, the model must integrate the signal contrast patterns of each modality in its small neighborhood, thereby simulating the diagnostic logic of an image doctor and laying the foundation for subsequent accurate judgment.

[0064] Step two: attention fusion based on local context: The Conv-based QKV generation module structure proposed in the embodiment is shown in Figure 2 .

[0065] The Conv-QKV module generates attention vectors through two efficient methods: Scheme A (mainstream implementation): A shared convolutional layer outputs an intermediate tensor with 3 d channels, which is then uniformly divided along the channel dimension to directly obtain Q_conv, K_conv and V_conv. The detailed structure of scheme A is shown in Figure 3 .

[0066] Scheme B (flexible variant): Use a group convolution with a group number of 3 to generate more independent feature representations for Q, K and V.

[0067] The key optimization of self-attention calculation is that since each vector in K_conv and V_conv carries the fusion information of a local small area, the calculation of attention weights is upgraded from the traditional "voxel-to-voxel" isolated dot product to "local area-to-local area" semantic similarity comparison. This brings two core advantages: More accurate boundary fusion: In key areas such as tumor boundaries, the model can intelligently integrate the enhanced edges of T1c and the high signal edges of FLAIR to generate smoother segmentation boundaries that conform to the anatomical structure.

[0068] Better distinguish similar tissues: For tissues with similar signals but different pathological properties on a single modality (such as peritumoral edema and tumor infiltration), the model can effectively distinguish them based on their unique local multi-modal "fingerprint" patterns.

[0069] Step three: feature decoding and segmentation map generation; The enhanced fusion features are upsampled by the decoder and combined with the high-resolution features of the shallow layers of the encoder through a skip connection to gradually restore spatial details. The network finally outputs precise probability maps for each voxel belonging to different sub-regions such as tumor core, edema, and enhanced tumor, completing the segmentation task.

[0070] This embodiment introduces locality and inductive bias from the source of attention generation mechanism by placing Conv-QKV Transformer at the fusion key node of U-Net++, achieving more efficient and accurate multi-modal information fusion and segmentation that conforms to the characteristics of medical images.

[0071] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for multi-modal MRI brain tumor segmentation based on convolutional attention, characterized in that, The method comprises the following steps: Parallelly encoding multi-modal medical image data, extracting low-level spatial features specific to each modality; Feature fusion and high-level encoding are performed on the low-level spatial features to obtain shared features; Convolutional attention processing is performed on the shared features to obtain enhanced features; The enhanced features are decoded to obtain segmentation results.

2. The convolutional attention-based multi-modal MRI brain tumor segmentation method according to claim 1, characterized in that, The multi-modal medical image data includes at least two of the T1 weighted enhancement image, the T2 weighted image and the FLAIR image in brain magnetic resonance imaging.

3. The convolutional attention-based multi-modal MRI brain tumor segmentation method of claim 1, wherein, Extracting low-level spatial features specific to each modality includes: Inputting image data of each modality in the multi-modal medical image data into an independent encoder branch to extract low-level spatial features specific to each modality; wherein each encoder branch includes a plurality of convolutional layers, an instance normalization layer and a LeakyReLU activation layer.

4. The convolutional attention-based multi-modal MRI brain tumor segmentation method of claim 1, wherein, Obtaining the shared features includes: Channel dimension splicing of the low-level spatial features at a preset depth of the encoder; The combined features are input into a shared encoding layer to further extract abstract features and obtain the shared features.

5. The convolutional attention-based multi-modal MRI brain tumor segmentation method according to claim 1, characterized in that, Convolutional attention processing of the shared features includes: Convolutional projection of the shared features to obtain an intermediate feature tensor; Dividing the intermediate feature tensor in the channel dimension to obtain a query tensor, a key tensor and a value tensor; Based on the query tensor, the key tensor and the value tensor, context-aware attention calculation is performed to obtain enhanced features.

6. The convolutional attention-based multi-modal MRI brain tumor segmentation method according to claim 5, characterized in that, Convolutional projection of the shared features includes: Inputting the shared features into a shared convolutional layer for convolution operation; wherein the shared convolutional layer uses a small convolutional kernel of 3*3*3 and sets padding to keep the spatial size unchanged; the output channel number is set to 3*d, and d is the planned projection attention embedding dimension.

7. The convolutional attention-based multi-modal MRI brain tumor segmentation method according to claim 5, characterized in that, Based on the query tensor, the key tensor and the value tensor, context-aware attention calculation includes: Q_conv, K_conv, V_conv are reshaped into a sequence form [B, N, d], where Q_conv, K_conv, V_conv are query tensor, key tensor and value tensor respectively, each dimension is [B, d, H', W', D'], N = H' W' D' is the sequence length, B is the batch size, d is the attention embedding dimension projected to, H' is the height, W' is the width, and D' is the depth. Based on the reconstructed sequence form, standard multi-head self-attention calculation is performed.

8. The convolutional attention-based multi-modal MRI brain tumor segmentation method of claim 1, wherein, Decoding processing of the enhanced features includes: Inputting the enhanced features into a decoder for upsampling and combining with low-level spatial features to generate a segmentation result image; wherein the decoder includes a plurality of upsampling layers and convolutional layers, and each level of the decoder is connected by a jump to fuse multi-modal low-level spatial features from the corresponding level of the encoder.

Images