Pelvic soft osteosarcoma MRI segmentation method based on edge-guided multi-scale fusion

Abstract

The invention discloses a pelvic soft osteosarcoma MRI (Magnetic Resonance Imaging) segmentation method based on edge-guided multi-scale fusion. The method comprises the following steps: firstly, acquiring an MRI image of a patient with soft osteosarcoma of pelvis, and preprocessing the image; and then, constructing a pelvic soft osteosarcoma MRI segmentation network model, namely an EMU-Net segmentation network, based on edge-guided multi-scale fusion. And training the network by using the preprocessed training data, inputting a pelvic soft osteosarcoma MRI image to be segmented into the trained network, outputting a corresponding segmentation probability graph, and obtaining a final tumor segmentation mask by setting a threshold value. The method not only effectively overcomes the clinical segmentation bottleneck, but also realizes full-automatic and high-efficiency precise segmentation by means of an end-to-end deep learning architecture, and has remarkable advantages in the aspects of precision and robustness.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence technology, and relates to an MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion. Background Technology

[0002] Chondrosarcoma is a malignant bone tumor originating from cartilage tissue. Pelvic chondrosarcoma, in particular, presents significant challenges to clinical diagnosis and treatment due to its complex anatomical location, insidious early symptoms, and indistinct tumor borders. Magnetic resonance imaging (MRI) technology can clearly visualize soft tissue and is an important tool for preoperative assessment and surgical planning of pelvic chondrosarcoma. Accurate tumor segmentation is crucial for determining the extent of surgical resection and avoiding tumor residue and postoperative recurrence.

[0003] Currently, segmentation of pelvic chondrosarcoma mainly relies on manual delineation by radiologists. This method is time-consuming, labor-intensive, heavily dependent on the doctor's clinical experience, and has low reproducibility and efficiency. In recent years, deep learning-based automatic segmentation methods have achieved great success in various medical image segmentation tasks. However, when these methods are applied to pelvic chondrosarcoma segmentation, their performance is often unsatisfactory. This is mainly due to several unique imaging features of this tumor:

[0004] 1. Blurred and multi-contour tumor boundaries: Pelvic chondrosarcoma grows in a "garland" pattern, with multiple inner and outer boundaries and low contrast with surrounding normal tissue. This makes it difficult for traditional segmentation models to accurately capture edge features, easily resulting in blurred or broken boundaries, such as... Figure 1 As shown in Figure a.

[0005] 2. Lesion morphological heterogeneity and spatial discontinuity: Tumors often contain multi-scale cavities and scattered small lesions, exhibiting highly irregular morphology and discontinuous distribution. This makes it difficult for general models to simultaneously capture both global morphology and local details, easily leading to missegmentation and undersegmentation. Figure 1 As shown in Figures b and c.

[0006] 3. Data scarcity and insufficient model generalization: The disease is relatively rare, which makes data-driven deep learning models prone to overfitting.

[0007] Existing solutions do not specifically address the core challenges of segmenting chondrosarcoma. Therefore, there is a need for an automated segmentation method designed specifically for the characteristics of pelvic chondrosarcoma MRI images, capable of accurately capturing blurred boundaries and multi-scale heterogeneous structures. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion. This method significantly improves the segmentation accuracy for blurred boundaries and multi-scale lesions by introducing two core innovative modules—Edge-Guided Coordinate Attention (EGCA) and Multi-Scale Dilated Convolution Fusion (MSDCF).

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, embodiments of this application provide an MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion, comprising the following steps:

[0011] Step 1: Data preprocessing.

[0012] MRI images of patients with pelvic chondrosarcoma were acquired, preprocessed, and divided into training and testing sets.

[0013] Step 2: Construct an edge-guided multi-scale U-Net (EMU-Net) segmentation network model for pelvic chondrosarcoma MRI, i.e., the EMU-Net segmentation network.

[0014] Step 3: Train the constructed EMU-Net segmentation network using the preprocessed training data.

[0015] Step 4: Input the MRI image of the pelvic chondrosarcoma to be segmented into the trained EMU-Net segmentation network, output the corresponding segmentation probability map, and obtain the final tumor segmentation mask by setting a threshold.

[0016] In one possible implementation, data preprocessing specifically includes:

[0017] 1-1. Data screening and slicing: 2D slices were made from the 3D MRI sequences, and only slices containing lesion areas were retained.

[0018] 1-2. Image size and intensity normalization: All images are scaled to the same size using bilinear interpolation, and the Min-Max normalization method is used to map pixel gray values ​​to the 0,1 range.

[0019] 1-3. Labeling and Division: The ITK-SNAP software was used to delineate the lesions layer by layer, generating lesion masks as ground truth. The dataset was randomly divided into training and test sets according to the cases.

[0020] In one possible implementation, the EMU-Net segmentation network model adopts a symmetrical U-shaped structure, consisting of three parts: an encoder, a bottleneck layer, and a decoder. The encoder is used for multi-scale feature extraction, the bottleneck layer uses a multi-scale dilated convolution fusion (MSDCF) module to achieve global context modeling and feature fusion, and the decoder gradually restores the spatial resolution and generates the segmentation result.

[0021] In one possible implementation, step 2 is specifically performed as follows:

[0022] 2-1. Encoder Construction: An Enhanced Visual Geometry Group (EVGG) is used as the backbone feature extractor. The encoder contains 5 EVGG blocks, each consisting of 2 standard convolutional blocks. Each standard convolutional block comprises a 3x3 convolutional layer, a normalized layer, and a ReLU activation function. After the first standard convolutional block, an edge-guided coordinate attention module is embedded, together forming the EVGG block. Progressive downsampling is used between the 5 EVGG blocks to expand the receptive field and extract features from low-level edge texture to high-level semantics.

[0023] 2-2. Between the encoder and decoder, a multi-scale dilated convolutional fusion module is embedded to construct a bottleneck layer. It receives the feature map from the encoder, obtains multiple sets of features through multi-scale dilated convolutional branches, then performs adaptive weight allocation and summation, adds the fusion result to the original input feature map by residual addition, and finally generates the fused feature map.

[0024] 2-3. Decoder and Output Layer Construction:

[0025] 2-3-1. The decoder contains 4 EVGG blocks. The spatial resolution of the fused feature map extracted by the multi-scale dilated convolution module is gradually restored through upsampling operations between the 4 EVGG blocks. Skip connections are used to concatenate the features of the corresponding scale extracted by the EVGG blocks in the encoder path with the features of the decoder path to fuse deep semantic information and shallow positional detail information.

[0026] 2-3-2. A 1×1 convolutional layer is connected to the end of the decoder to map the number of channels to the number of target categories, and outputs a probability map of each pixel belonging to the tumor through the Sigmoid activation function.

[0027] In one possible implementation, the edge-guided coordinate attention module includes two branches, a coordinate attention branch and an edge-guided branch, which adaptively fuse the coordinate attention map and the edge attention map through a learnable weight parameter to generate the final edge-guided coordinate attention map.

[0028] Coordinate attention branch: for the input feature map Global average pooling is performed along the height and width directions respectively to obtain a pair of orientation-aware feature vectors, the calculation process of which is defined by the following formula:

[0029]

[0030]

[0031] in, and These represent the height and width of the input feature map, respectively. Indicates the input feature map of the th Dimension 1 Liede Elements, similarly, Indicates the input feature map of the th Dimension 1 Line number element, and These are feature vectors in the height and width directions, respectively. Subsequently, these two feature vectors are concatenated spatially, and dimensionality reduction and feature fusion are performed using a shared 1×1 convolution to obtain the concatenated spatial feature vector. The process can be described as follows:

[0032]

[0033] in, This represents the ReLU activation function. This represents a 1×1 convolution operation. Next, ... Split into compressed feature vectors in the height direction and the compressed feature vector in the width direction Attention weight maps in the height and width directions are generated separately using independent 1×1 convolutions and sigmoid activation functions:

[0034]

[0035] Edge-guided branch: for the input feature map Gaussian filtering is used to suppress noise, as shown in the following formula:

[0036]

[0037] in, This is the standard deviation of the Gaussian kernel. Subsequently, a Laplacian operator is used for convolution to extract edge features. The Laplacian convolution kernel is defined as:

[0038]

[0039] Then, a sub-network consisting of a 1×1 convolutional layer (containing BN and ReLU) is used for transformation, and finally, the edge attention map is generated by the Sigmoid function, as shown in the following formula:

[0040]

[0041] in, This represents the ReLU activation function. This represents a 1×1 convolution operation.

[0042] Attention fusion: through a learnable weight parameter Adaptively fuse the coordinate attention map and the edge attention map to generate the final edge-guided coordinate attention map: .

[0043] Feature enhancement: The fused attention map The original input feature map is multiplied element-wise to obtain an output feature map with enhanced boundary awareness.

[0044] In one possible implementation, the multi-scale dilated convolutional fusion module operates as follows:

[0045] Multi-scale feature extraction: The multi-scale dilated convolutional fusion module contains five parallel branches for a given input feature map. The three branches employ 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively, and are transformed using a sub-network consisting of a 1×1 convolutional layer (containing BN and ReLU) to obtain differentiated receptive fields, as shown in the following formula:

[0046]

[0047] in, Indicates the first A side road, This represents a 1×1 convolution operation. , , These are 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively.

[0048] The other two branches are cascaded to obtain denser sampling and more complex feature representations. They also use 1×1 convolutional layers for variation, as shown in the following formula:

[0049]

[0050]

[0051] in, This indicates the 4th branch road. This indicates the 5th branch road. This represents a 1×1 convolution operation.

[0052] Adaptive Feature Fusion: The feature maps output from the five branches are concatenated, and adaptive normalized weights for each branch are generated using the Softmax function. Based on this, the multi-scale features are weighted and summed to obtain the fused features, as shown in the following formula:

[0053]

[0054]

[0055] in, Indicates the weights of different branches. Indicates different branches, This is the fused feature map.

[0056] Residual connection: The fused feature map is added to the input feature map of the module by residual addition to obtain the final fused feature map. .

[0057] In one possible implementation, step 3 is specifically performed as follows:

[0058] Binary Cross-Entropy Loss (BCE Loss) is used as the loss function. By minimizing the BCE Loss, the network parameters are updated using the backpropagation algorithm. Stochastic Gradient Descent (SGD) is employed during optimization, and a momentum term is used to enhance the stability of the training process and improve the model's convergence speed.

[0059] Secondly, embodiments of this application provide an MRI segmentation device for pelvic chondrosarcoma based on edge-guided multi-scale fusion, comprising the following modules:

[0060] Data preprocessing module: Acquire MRI images of patients with pelvic chondrosarcoma, preprocess the images, and divide them into training and test sets.

[0061] Model building module: Constructs an edge-guided multi-scale U-Net (EMU-Net) segmentation network model for pelvic chondrosarcoma MRI based on edge-guided multi-scale fusion, namely the EMU-Net segmentation network.

[0062] Training module: Trains the constructed EMU-Net segmentation network using preprocessed training data.

[0063] Segmentation module: The MRI image of pelvic chondrosarcoma to be segmented is input into the trained EMU-Net segmentation network, which outputs the corresponding segmentation probability map and obtains the final tumor segmentation mask by setting a threshold.

[0064] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory.

[0065] The memory is used to store computer programs.

[0066] When the processor executes the program stored in the memory, it implements any of the MRI segmentation methods for pelvic chondrosarcoma described in this application.

[0067] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the MRI segmentation methods for pelvic chondrosarcoma described in this application.

[0068] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to execute any of the MRI segmentation methods for pelvic chondrosarcoma described in this application.

[0069] The beneficial effects of this invention are as follows:

[0070] This invention addresses the core challenges of existing MRI image segmentation techniques for pelvic chondrosarcoma, including blurred and fragmented tumor boundaries, significant heterogeneity in lesion morphology, and difficulty in effectively capturing multi-scale features. It proposes an image segmentation method based on edge guidance and multi-scale contextual fusion. By designing an edge-guided coordinate attention module, the method explicitly enhances the perception and localization capabilities of blurred and multi-contour boundaries. Furthermore, by constructing a multi-scale dilated convolutional fusion module, it adaptively aggregates global and local contextual information to handle lesion regions with diverse morphologies. The proposed technique not only effectively overcomes the aforementioned clinical segmentation bottlenecks but also achieves fully automated, highly efficient, and accurate segmentation based on an end-to-end deep learning architecture, demonstrating significant advantages in accuracy and robustness. Algorithmically, this invention constructs a new segmentation paradigm specifically for this disease. Clinically, it provides reliable tumor location analysis for preoperative planning, significantly improving diagnostic efficiency and surgical accuracy, and has clear application prospects and promotional value in the field of smart healthcare. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is an MRI image and lesion delineation diagram of pelvic chondrosarcoma according to an embodiment of the present invention.

[0073] Figure 2 This is a structural diagram of the EMU-Net segmentation network model according to an embodiment of the present invention.

[0074] Figure 3 This is a structural diagram of the enhanced visual geometry group in an embodiment of the present invention.

[0075] Figure 4 This is a structural diagram of the edge-guided coordinate attention module in an embodiment of the present invention.

[0076] Figure 5 This is a structural diagram of the multi-scale dilated convolutional fusion module according to an embodiment of the present invention. Specific implementation methods

[0077] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. It should be noted that the described embodiments are actually only some embodiments of the present invention, not all embodiments, and are merely illustrative, and are in no way intended to limit the present invention or its application or use. The scope of protection of this application shall be determined by the appended claims.

[0078] An MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion includes the following steps:

[0079] Step 1: Data Preprocessing

[0080] In this embodiment, the MRI images used in the experiment were obtained from the Bone Tumor Center of the Second Affiliated Hospital of Zhejiang University School of Medicine, and were acquired between 2011 and 2023, containing imaging data of 237 patients. Among them, 95 cases were pathologically confirmed as pelvic chondrosarcoma, and the rest were other bone tumor cases with similar imaging characteristics. All images were obtained using T1-weighted sequences (T1WI), with a slice thickness of 5 mm and a pixel pitch of 0.5 mm per case.

[0081] 1-1. Data screening and slicing: 2D slices were made from the 3D MRI sequences, and only slices containing lesion areas were retained, resulting in a total of 2636 images.

[0082] 1-2. Image Size and Intensity Normalization: All images are scaled to 512×512 pixels using bilinear interpolation, and Min-Max normalization is applied to map pixel grayscale values ​​to the 0,1 range to eliminate scanning differences between different devices. The specific calculation formula is as follows: Where I is the original image, The minimum pixel value in the entire image. This is the maximum pixel value of the entire image.

[0083] 1-3. Labeling and Division: Two radiologists with over 10 years of experience used ITK-SNAP software to draw the lesions layer by layer, generating lesion masks as ground truth. The dataset was randomly divided into training and test sets according to the cases, with a ratio of 4:1.

[0084] Step 2: Construct an MRI segmentation network model for pelvic chondrosarcoma based on edge-guided multi-scale fusion (EMU-Net).

[0085] like Figure 2 As shown, the EMU-Net segmentation network model adopts a symmetrical U-shaped structure, consisting of three parts: an encoder, a bottleneck layer, and a decoder. The encoder is used for multi-scale feature extraction, the bottleneck layer uses a multi-scale dilated convolution fusion (MSDCF) module to achieve global context modeling and feature fusion, and the decoder gradually restores the spatial resolution and generates the segmentation result. The specific implementation steps are as follows:

[0086] 2-1. The encoder consists of five levels, each using an EVGG block as a feature extraction unit. The EVGG structure diagram is shown below. Figure 3 As shown, each EVGG block includes two standard convolutional blocks, each consisting of a 3x3 convolutional layer, a normalized layer, and a ReLU activation function. An edge-guided coordinate attention module is embedded after the first standard convolutional block, together forming the EVGG block. The input image size is 512×512×3, with an initial convolutional channel count of 32. The channel count is subsequently doubled for each layer (32, 64, 128, 256, 512). Max pooling (2×2, stride of 2) is used between the five EVGG blocks to progressively downsample, expanding the receptive field and extracting features from low-level edge textures to high-level semantics.

[0087] The edge-guided coordinate attention module includes two branches: the coordinate attention branch and the edge-guided branch. It adaptively fuses the coordinate attention map and the edge attention map through a learnable weight parameter to generate the final edge-guided coordinate attention map.

[0088] Coordinate attention branch: such as Figure 4 As shown, for the input feature map Global average pooling is performed along the height and width directions respectively to obtain a pair of orientation-aware feature vectors, the calculation process of which is defined by the following formula:

[0089]

[0090]

[0091] in, and These represent the height and width of the input feature map, respectively. Indicates the input feature map of the th Dimension 1 Liede Elements, similarly, Indicates the input feature map of the th Dimension 1 Line number element, and These are feature vectors in the height and width directions, respectively. Subsequently, these two feature vectors are concatenated spatially, and dimensionality reduction and feature fusion are performed using a shared 1×1 convolution to obtain the concatenated spatial feature vector. The process can be described as follows:

[0092]

[0093] in, This represents the ReLU activation function. This represents a 1×1 convolution operation. Next, ... Split into compressed feature vectors in the height direction and the compressed feature vector in the width direction Attention weight maps in the height and width directions are generated separately using independent 1×1 convolutions and sigmoid activation functions:

[0094]

[0095] This map effectively captures long-range spatial dependencies and preserves precise location information.

[0096] Edge-guided branch: for the input feature map Gaussian filtering is used to suppress noise, as shown in the following formula:

[0097]

[0098] in, This is the standard deviation of the Gaussian kernel. Subsequently, a Laplacian operator is used for convolution to extract edge features. The Laplacian convolution kernel is defined as:

[0099]

[0100] Then, a sub-network consisting of a 1×1 convolutional layer (containing BN and ReLU) is used for transformation, and finally, the edge attention map is generated by the Sigmoid function, as shown in the following formula:

[0101]

[0102] in, This represents the ReLU activation function. This represents a 1×1 convolution operation.

[0103] Attention fusion: through a learnable weight parameter Adaptively fuse the coordinate attention map and the edge attention map to generate the final edge-guided coordinate attention map: .

[0104] Feature enhancement: The fused attention map The original input feature map is multiplied element-wise to obtain an output feature map with enhanced boundary awareness.

[0105] 2-2. A bottleneck layer is constructed by embedding a multi-scale dilated convolutional fusion module between the encoder and decoder. This module receives feature maps from the encoder, obtains multiple sets of features through multi-scale dilated convolutional branches, adaptively assigns weights to these features, and sums them. The fusion result is then added to the original input feature map as a residual to generate a fused feature map, which is then fed into the decoder for further processing. This design allows the network to obtain receptive fields of different sizes without significantly increasing computational complexity, thereby improving its ability to model heterogeneous lesion morphology and fine boundaries.

[0106] The specific operation of the multi-scale dilated convolutional fusion module is as follows:

[0107] Multi-scale feature extraction: such as Figure 5 As shown, the multi-scale dilated convolutional fusion module contains five parallel branches for a given input feature map. The three branches employ 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively, and are transformed using a sub-network consisting of a 1×1 convolutional layer (containing BN and ReLU) to obtain differentiated receptive fields. The formula is as follows:

[0108]

[0109] in, Indicates the first A side road, This represents a 1×1 convolution operation. , , These are 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively.

[0110] The other two branches are cascaded (the fourth branch uses a dilated convolution with a dilation rate of 1 followed by a dilated convolution with a dilation rate of 3, and the fifth branch uses a dilated convolution with a dilation rate of 1 followed by a dilated convolution with a dilation rate of 3 followed by a dilated convolution with a dilation rate of 5) to obtain denser sampling and more complex feature representations. They also use 1×1 convolutional layers for variation, as shown in the following formula:

[0111]

[0112]

[0113] in, This indicates the 4th branch road. This indicates the 5th branch road. This represents a 1×1 convolution operation.

[0114] Adaptive Feature Fusion: The feature maps output from the five branches are concatenated, and adaptive normalized weights for each branch are generated using the Softmax function. Based on this, the multi-scale features are weighted and summed to obtain the fused features, as shown in the following formula:

[0115]

[0116]

[0117] in, Indicates the weights of different branches. Indicates different branches, This is the fused feature map.

[0118] Residual connection: The fused features are added to the input features of the module by residual addition to obtain the final fused feature map. This preserves the original semantic information and ensures the stability of gradient propagation, effectively improving the model's ability to model multi-scale cavities and scattered lesions.

[0119] 2-3. Decoder and Output Layer Construction:

[0120] 2-3-1. The decoder is symmetrical to the encoder and contains 4 EVGG blocks. The fused feature map extracted by the multi-scale dilated convolutional module is gradually recovered through upsampling operations (using bilinear interpolation) between the 4 EVGG blocks. The spatial resolution is determined, and skip connections are used to concatenate the features at the corresponding scale extracted from the EVGG blocks in the encoder path with the features in the decoder path. The concatenated features are then processed by two layers of 3×3 convolution and ReLU activation, with the number of channels being halved layer by layer (512→256→128→64→32).

[0121] 2-3-2. A 1×1 convolutional layer is connected to the end of the decoder to map the number of channels to 1, and the probability map of each pixel belonging to the tumor is output through the Sigmoid activation function.

[0122] Step 3: Network model training. This is achieved by minimizing the binary cross-entropy loss. The network parameters are updated using the backpropagation algorithm.

[0123] 3-1. Binary Cross-Entropy Loss :

[0124] Binary Cross-Entropy Loss (BCE Loss) is used, which effectively measures the difference between the segmentation map predicted by the model and the ground truth labeling. The calculation formula is as follows:

[0125]

[0126] Where N is the total number of pixels in a batch. The true label (0 or 1) for pixel i. The model predicts the probability that this pixel is a tumor.

[0127] 3-2 Loss Function Optimization

[0128] By minimizing the binary cross-entropy loss The network parameters are updated using the backpropagation algorithm. Stochastic Gradient Descent (SGD) is employed during optimization, and a momentum term is introduced to enhance the stability of the training process and improve the model's convergence speed. The specific formula is as follows:

[0129]

[0130]

[0131] in, This represents the network parameters at the t-th iteration; Indicates the learning rate; Indicates the momentum coefficient; This represents the gradient of the loss function with respect to the network parameters; Let represent the momentum vector at the t-th iteration.

[0132] The specific training configuration parameters are as follows: learning rate is set to 0.001, momentum is 0.9, batch size is 4, and training epochs are 150. To quantitatively evaluate the performance of the proposed EMU-Net, we selected commonly used metrics for medical image segmentation: Intersection over Union (IoU), Dice similarity coefficient, recall, and precision as evaluation metrics.

[0133] Step 4: Performance Validation. To validate the effectiveness of the proposed model in the MRI image segmentation task of pelvic chondrosarcoma, we conducted a quantitative comparison with mainstream segmentation models, and the results are shown in the table below.

[0134] As shown in the table, the proposed EMU-Net consistently outperforms all baseline methods on all key evaluation metrics: IoU reaches 78.25%, Dice 87.02%, recall 87.73%, and precision 87.22%, representing improvements of 3.14%, 1.89%, 1.12%, and 2.86% respectively compared to the second-best results.

[0135] This application also provides an MRI segmentation device for pelvic chondrosarcoma based on edge-guided multi-scale fusion, comprising the following modules:

[0136] Data preprocessing module: Acquire MRI images of patients with pelvic chondrosarcoma, preprocess the images, and divide them into training and test sets.

[0137] Model building module: Constructs an edge-guided multi-scale U-Net (EMU-Net) segmentation network model for pelvic chondrosarcoma MRI based on edge-guided multi-scale fusion, namely the EMU-Net segmentation network.

[0138] Training module: Trains the constructed EMU-Net segmentation network using preprocessed training data.

[0139] Segmentation module: The MRI image of pelvic chondrosarcoma to be segmented is input into the trained EMU-Net segmentation network, which outputs the corresponding segmentation probability map and obtains the final tumor segmentation mask by setting a threshold.

[0140] In one possible implementation, the data preprocessing module operates as follows:

[0141] 1-1. Data screening and slicing: 2D slices were made from the 3D MRI sequences, and only slices containing lesion areas were retained.

[0142] 1-2. Image size and intensity normalization: All images are scaled to the same size using bilinear interpolation, and the Min-Max normalization method is used to map pixel gray values ​​to the 0,1 range.

[0143] 1-3. Labeling and Division: The ITK-SNAP software was used to delineate the lesions layer by layer, generating lesion masks as ground truth. The dataset was randomly divided into training and test sets according to the cases.

[0144] In one possible implementation, the EMU-Net segmentation network model adopts a symmetrical U-shaped structure, consisting of three parts: an encoder, a bottleneck layer, and a decoder. The encoder is used for multi-scale feature extraction, the bottleneck layer uses a multi-scale dilated convolution fusion (MSDCF) module to achieve global context modeling and feature fusion, and the decoder gradually restores the spatial resolution and generates the segmentation result.

[0145] In one possible implementation, the model building module operates as follows:

[0146] 2-1. Encoder Construction: An Enhanced Visual Geometry Group (EVGG) is used as the backbone feature extractor. The encoder contains 5 EVGG blocks, each consisting of 2 standard convolutional blocks. Each standard convolutional block comprises a 3x3 convolutional layer, a normalized layer, and a ReLU activation function. After the first standard convolutional block, an edge-guided coordinate attention module is embedded, together forming the EVGG block. Progressive downsampling is used between the 5 EVGG blocks to expand the receptive field and extract features from low-level edge texture to high-level semantics.

[0147] 2-2. Between the encoder and decoder, a multi-scale dilated convolutional fusion module is embedded to construct a bottleneck layer. It receives the feature map from the encoder, obtains multiple sets of features through multi-scale dilated convolutional branches, then performs adaptive weight allocation and summation, adds the fusion result to the original input feature map by residual addition, and finally generates the fused feature map.

[0148] 2-3. Decoder and Output Layer Construction:

[0149] 2-3-1. The decoder contains 4 EVGG blocks. The spatial resolution of the fused feature map extracted by the multi-scale dilated convolution module is gradually restored through upsampling operations between the 4 EVGG blocks. Skip connections are used to concatenate the features of the corresponding scale extracted by the EVGG blocks in the encoder path with the features of the decoder path to fuse deep semantic information and shallow positional detail information.

[0150] 2-3-2. A 1×1 convolutional layer is connected to the end of the decoder to map the number of channels to the number of target categories, and outputs a probability map of each pixel belonging to the tumor through the Sigmoid activation function.

[0151] In one possible implementation, the edge-guided coordinate attention module includes two branches, a coordinate attention branch and an edge-guided branch, which adaptively fuse the coordinate attention map and the edge attention map through a learnable weight parameter to generate the final edge-guided coordinate attention map.

[0152] Coordinate attention branch: for the input feature map Global average pooling is performed along the height and width directions respectively to obtain a pair of orientation-aware feature vectors, the calculation process of which is defined by the following formula:

[0153]

[0154]

[0155] in, and These represent the height and width of the input feature map, respectively. Indicates the input feature map of the th Dimension 1 Liede Elements, similarly, Indicates the input feature map of the th Dimension 1 Line number element, and These are feature vectors in the height and width directions, respectively. Subsequently, these two feature vectors are concatenated spatially, and dimensionality reduction and feature fusion are performed using a shared 1×1 convolution to obtain the concatenated spatial feature vector. The process can be described as follows:

[0156]

[0157] in, This represents the ReLU activation function. This represents a 1×1 convolution operation. Next, ... Split into compressed feature vectors in the height direction and the compressed feature vector in the width direction Attention weight maps in the height and width directions are generated separately using independent 1×1 convolutions and sigmoid activation functions:

[0158]

[0159] Edge-guided branch: for the input feature map Gaussian filtering is used to suppress noise, as shown in the following formula:

[0160]

[0161] in, This is the standard deviation of the Gaussian kernel. Subsequently, a Laplacian operator is used for convolution to extract edge features. The Laplacian convolution kernel is defined as:

[0162]

[0163] Then, a sub-network consisting of a 1×1 convolutional layer (containing BN and ReLU) is used for transformation, and finally, the edge attention map is generated by the Sigmoid function, as shown in the following formula:

[0164]

[0165] in, This represents the ReLU activation function. This represents a 1×1 convolution operation.

[0166] Attention fusion: through a learnable weight parameter Adaptively fuse the coordinate attention map and the edge attention map to generate the final edge-guided coordinate attention map: .

[0167] Feature enhancement: The fused attention map The original input feature map is multiplied element-wise to obtain an output feature map with enhanced boundary awareness.

[0168] In one possible implementation, the multi-scale dilated convolutional fusion module operates as follows:

[0169] Multi-scale feature extraction: The multi-scale dilated convolutional fusion module contains five parallel branches for a given input feature map. The three branches employ 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively, and are transformed using a sub-network consisting of a 1×1 convolutional layer (containing BN and ReLU) to obtain differentiated receptive fields, as shown in the following formula:

[0170]

[0171] in, Indicates the first A side road, This represents a 1×1 convolution operation. , , These are 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively.

[0172] The other two branches are cascaded to obtain denser sampling and more complex feature representations. They also use 1×1 convolutional layers for variation, as shown in the following formula:

[0173]

[0174]

[0175] in, This indicates the 4th branch road. This indicates the 5th branch road. This represents a 1×1 convolution operation.

[0176] Adaptive Feature Fusion: The feature maps output from the five branches are concatenated, and adaptive normalized weights for each branch are generated using the Softmax function. Based on this, the multi-scale features are weighted and summed to obtain the fused features, as shown in the following formula:

[0177]

[0178]

[0179] in, Indicates the weights of different branches. Indicates different branches, This is the fused feature map.

[0180] Residual connection: The fused feature map is added to the input feature map of the module by residual addition to obtain the final fused feature map. .

[0181] In one possible implementation, the training module operates as follows:

[0182] Binary Cross-Entropy Loss (BCE Loss) is used as the loss function. By minimizing the BCE Loss, the network parameters are updated using the backpropagation algorithm. Stochastic Gradient Descent (SGD) is employed during optimization, and a momentum term is used to enhance the stability of the training process and improve the model's convergence speed.

[0183] This application also provides an electronic device, including a processor and a memory.

[0184] The memory is used to store computer programs.

[0185] When the processor executes a program stored in the memory, it implements any of the methods described in this application.

[0186] In one possible implementation, the electronic device of this application embodiment further includes a communication interface and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0187] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0188] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0189] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0190] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0191] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements any of the methods described in this application.

[0192] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the methods described in this application.

[0193] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0194] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0195] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0196] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A method for MRI segmentation of pelvic chondrosarcoma based on edge-guided multi-scale fusion, characterized in that, Includes the following steps: Step 1: Data preprocessing; MRI images of patients with pelvic chondrosarcoma were acquired, the images were preprocessed, and the data were divided into training and testing sets. Step 2: Construct an MRI segmentation network model for pelvic chondrosarcoma based on edge-guided multi-scale fusion, namely the EMU-Net segmentation network; Step 3: Train the constructed EMU-Net segmentation network using the preprocessed training data; Step 4: Input the MRI image of the pelvic chondrosarcoma to be segmented into the trained EMU-Net segmentation network, output the corresponding segmentation probability map, and obtain the final tumor segmentation mask by setting a threshold.

2. The MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion according to claim 1, characterized in that, Data preprocessing specifically includes: 1-1. Data screening and slicing: 2D slices were made from the 3D MRI sequences, and only slices containing lesion areas were retained; 1-2. Image size and intensity normalization: All images are scaled to the same size using bilinear interpolation, and the Min-Max normalization method is used to map pixel gray values ​​to the 0,1 range; 1-3. Labeling and Division: The ITK-SNAP software was used to draw the lesions layer by layer and generate lesion masks as ground truth. The dataset was randomly divided into training set and test set according to the cases.

3. The MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion according to claim 1, characterized in that, The EMU-Net segmentation network model described above adopts a symmetrical U-shaped structure, consisting of three parts: an encoder, a bottleneck layer, and a decoder. The encoder is used for multi-scale feature extraction, the bottleneck layer uses a multi-scale dilated convolutional fusion module to realize global context modeling and feature fusion, and the decoder gradually restores the spatial resolution and generates the segmentation result.

4. The MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion according to claim 3, characterized in that, Step 2 is performed as follows: 2-1. Encoder Construction: Enhanced Visual Geometry Group (EVGG) is used as the backbone feature extractor. The encoder contains 5 EVGG blocks, each of which includes 2 standard convolutional blocks. The standard convolutional block consists of a 3x3 convolutional layer, a normalization layer, and a ReLU activation function. After the first standard convolutional block, an edge-guided coordinate attention module is embedded to form the EVGG block. Progressive downsampling is used between the 5 EVGG blocks to expand the receptive field and extract features from low-level edge texture to high-level semantics. 2-2. Between the encoder and decoder, a multi-scale dilated convolutional fusion module is embedded to construct a bottleneck layer. It receives the feature map from the encoder, obtains multiple sets of features through multi-scale dilated convolutional branches, then performs adaptive weight allocation and summation, adds the fusion result to the original input feature map by residual addition, and finally generates the fused feature map. 2-3. Decoder and Output Layer Construction: 2-3-1. The decoder contains 4 EVGG blocks. The spatial resolution of the fused feature map extracted by the multi-scale dilated convolution module is gradually restored through upsampling operations between the 4 EVGG blocks. Skip connections are used to concatenate the features of the corresponding scale extracted by the EVGG blocks in the encoder path with the features of the decoder path to fuse deep semantic information and shallow positional detail information. 2-3-2. A 1×1 convolutional layer is connected to the end of the decoder to map the number of channels to the number of target categories, and outputs a probability map of each pixel belonging to the tumor through the Sigmoid activation function.

5. The MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion according to claim 4, characterized in that, The edge-guided coordinate attention module includes two branches: a coordinate attention branch and an edge-guided branch. It adaptively fuses the coordinate attention map and the edge attention map through a learnable weight parameter to generate the final edge-guided coordinate attention map. Coordinate attention branch: for the input feature map Global average pooling is performed along the height and width directions respectively to obtain a pair of orientation-aware feature vectors, the calculation process of which is defined by the following formula: in, and These represent the height and width of the input feature map, respectively. Indicates the input feature map of the th Dimension 1 Liede Elements, similarly, Indicates the input feature map of the th Dimension 1 Line 1 element, and These are feature vectors in the height and width directions, respectively. Then, these two feature vectors are concatenated spatially, and dimensionality reduction and feature fusion are performed using a shared 1×1 convolution to obtain the concatenated spatial feature vector. The process can be described as follows: in, Represents the ReLU activation function; This represents a 1×1 convolution operation; then, Split into compressed feature vectors in the height direction and the compressed feature vector in the width direction Attention weight maps in the height and width directions are generated separately using independent 1×1 convolutions and sigmoid activation functions: Edge-guided branch: for the input feature map Gaussian filtering is used to suppress noise, as shown in the following formula: in, It is the standard deviation of the Gaussian kernel; subsequently, the Laplacian operator is used for convolution to extract edge features; the Laplacian convolution kernel is defined as: Then, a sub-network consisting of a 1×1 convolutional layer is used for transformation, and finally, an edge attention map is generated by the Sigmoid function, as shown in the following formula: in, Represents the ReLU activation function; This represents a 1×1 convolution operation; Attention fusion: through a learnable weight parameter Adaptively fuse the coordinate attention map and the edge attention map to generate the final edge-guided coordinate attention map: ; Feature enhancement: The fused attention map The original input feature map is multiplied element-wise to obtain an output feature map with enhanced boundary awareness.

6. The MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion according to claim 4, characterized in that, The multi-scale dilated convolutional fusion module operates as follows: Multi-scale feature extraction: The multi-scale dilated convolutional fusion module contains five parallel branches for a given input feature map. The three branches employ 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively, and are transformed using a sub-network consisting of a 1×1 convolutional layer to obtain differentiated receptive fields, as shown in the following formula: in, Indicates the first A side road, This represents a 1×1 convolution operation. , , These are 3×3 dilated convolutions with dilation rates of 1, 3, and 5, respectively. The other two branches are cascaded to obtain denser sampling and more complex feature representations. They also use 1×1 convolutional layers for variation, as shown in the following formula: in, This indicates the 4th branch road. This indicates the 5th branch road. This represents a 1×1 convolution operation; Adaptive Feature Fusion: The feature maps output from the five branches are concatenated, and adaptive normalized weights for each branch are generated using the Softmax function. Based on this, the multi-scale features are weighted and summed to obtain the fused features, as shown in the following formula: in, Indicates the weights of different branches. Indicates different branches, The fused feature map; Residual connection: The fused feature map is added to the input feature map of the module by residual addition to obtain the final fused feature map. .

7. The MRI segmentation method for pelvic chondrosarcoma based on edge-guided multi-scale fusion according to claim 1, characterized in that, Step 3 is performed as follows: Binary cross-entropy loss is used as the loss function. By minimizing the binary cross-entropy loss, the backpropagation algorithm is used to update the network parameters. During the optimization process, stochastic gradient descent algorithm is used, and a momentum term is used to enhance the stability of the training process and improve the convergence speed of the model.

8. A pelvic chondrosarcoma MRI segmentation device based on edge-guided multi-scale fusion, characterized in that, Includes the following modules: Data preprocessing module: Acquire MRI images of patients with pelvic chondrosarcoma, preprocess the images, and divide them into training and test sets; Model building module: Constructs an edge-guided multi-scale fusion-based segmentation network model for pelvic chondrosarcoma, namely the EMU-Net segmentation network; Training module: Trains the constructed EMU-Net segmentation network using preprocessed training data; Segmentation module: The MRI image of pelvic chondrosarcoma to be segmented is input into the trained EMU-Net segmentation network, which outputs the corresponding segmentation probability map and obtains the final tumor segmentation mask by setting a threshold.

9. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the MRI segmentation method for pelvic chondrosarcoma as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the MRI segmentation method for pelvic chondrosarcoma according to any one of claims 1-7.

Images