Medical image segmentation method for lower limb femoral superficial artery calcification

Abstract

The invention discloses a medical image segmentation method for lower limb femoral superficial artery calcification, and belongs to the technical field of calcification medical image recognizing.The medical image segmentation method for lower limb femoral superficial artery calcification comprises the steps that CTA scanning data of lower limb femoral superficial artery calcification of multiple patients are collected, and multiple slices are selected for discrete marking; a three-way xLSTM module is adopted to comprehensively capture the dynamic change and spatial features of the blood vessels of the multiple discretely marked slices in the forward direction, the reverse direction and the slice mixing direction, and 3D features are obtained; a wavelet transform module is adopted to decompose and process the 3D features to obtain a key feature map; a multi-scale aggregation module is adopted to enhance the key feature map and identify the calcification region and the relative position between the calcification region and the blood vessel, and a segmentation result is obtained; efficient and accurate calcification segmentation is realized, and the extraction quality of the calcification region is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of calcification medical image recognition, and in particular relates to a medical image segmentation method for calcification of the superficial femoral artery of the lower limb. Background Art

[0002] Lower extremity arterial occlusive disease (LEAOD) is a major global health threat. One of its hallmark features is the formation of calcified vascular plaques. These calcified plaques not only reduce vascular recanalization rates but also increase the difficulty of endovascular interventional therapy, significantly impacting treatment outcomes. Therefore, accurate pre-treatment assessment of calcification is crucial for effective treatment planning, optimizing patient outcomes, and mitigating treatment risks. However, current clinical assessment methods for lower extremity arterial calcification still primarily rely on manual analysis or fixed-threshold segmentation methods. These traditional methods often struggle to achieve high-precision segmentation of the complex morphology and irregular distribution of superficial femoral artery calcification due to the overlap of grayscale ranges between calcified plaques and the vascular lumen, limiting early diagnosis and precision treatment. Furthermore, manual analysis methods are time-consuming and labor-intensive, making them inadequate for efficient clinical diagnosis and treatment. With the rapid advancement of medical imaging technology and artificial intelligence methods, deep learning-driven automated medical image segmentation techniques offer new opportunities for the precise extraction and quantitative assessment of calcified plaques. CT angiography (CTA) combined with deep learning segmentation models has demonstrated excellent performance in detecting calcifications in coronary arteries and other vascular locations. However, research on superficial femoral artery calcification segmentation remains limited, lacking specifically optimized small-target segmentation methods and efficient modeling strategies. This study, through in-depth exploration of a deep learning-based superficial femoral artery calcification segmentation method, aims to address challenges such as difficult calcification target localization, blurred boundaries, and insufficient multi-scale feature capture, providing technical support for efficient and accurate segmentation of superficial femoral artery calcification. Summary of the Invention

[0003] In response to the above problems, the present invention aims to provide a medical image segmentation method for superficial femoral artery calcification of the lower limbs, which solves the problems of existing assessment methods for lower limb artery calcification, such as difficulty in locating calcification targets, blurred boundaries, and insufficient capture of multi-scale features.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is as follows:

[0005] A method for segmenting medical images of superficial femoral artery calcification of lower limbs is provided, comprising:

[0006] Step 1: Collect CTA scan data of superficial femoral artery calcification in the lower limbs of multiple patients, and select multiple slices from the CTA scan data at a certain interval for discrete marking;

[0007] Step 2: A three-way xLSTM module is used to comprehensively capture the dynamic changes and spatial characteristics of blood vessels in multiple discretely labeled slices from three directions: forward, backward, and slice mixing, to obtain 3D features.

[0008] Step 3: Use the wavelet transform module to decompose and process the 3D features to obtain the key feature map;

[0009] Step 4: Use the multi-scale aggregation module to enhance the key feature map and identify the calcified area and the relative position between the calcified area and the blood vessel to obtain the segmentation result.

[0010] Furthermore, in step 1, the labeling categories include background, aortic lumen, and aortic calcification. Discrete labeling is completed under the guidance of a radiologist to ensure labeling quality and accuracy.

[0011] Furthermore, in step 2, the method for obtaining 3D features includes:

[0012] The multiple slices after discrete labeling are expanded into sequences along the following three directions, and the features are modeled and fused from the three directions of forward, reverse, and inter-slice mixing; in each direction, the xLSTM layer with shared weights is used to extract feature dependencies. The forward sequence z f Expand along the slice order direction (D, H, W), modeling forward feature dependency, reverse sequence z r Expand along the reverse direction of the slice (D, H, W), supplement the reverse feature information, and slice the mixed sequence z s Mix multiple slices in the channel dimension to perform feature interaction across slices;

[0013] The calculation formula of the three-way xLSTM module is:

[0014] T o M(z)=xLSTM(z f )+xLSTM(z r )+xLSTM(z s )#

[0015] Among them, T o M(z) represents the output after three-way feature modeling, and xLSTM represents the module for temporal feature extraction along a specific sequence;

[0016] The features of the three directions are weighted and fused to obtain the 3D feature z fused , 3D feature z fused The expression is:

[0017] z fused =αz f +βz r +γz s

[0018] Among them, α, β, and γ are the directional weights of the forward, reverse, and inter-slice mixing directions respectively.

[0019] Furthermore, in step 3, the wavelet transform module decomposes the 3D features to obtain high-frequency components and low-frequency components; the high-frequency components are used to capture detail information; the low-frequency components retain global context information, and the high-frequency components and low-frequency components are concatenated to obtain the 3D feature F I ;

[0020] 3D feature F I First, pass through a 1x1x1 convolution layer Conv fc1 The number of channels is compressed, and then passed through the GELU activation function, and then through another identical convolutional layer Conv fc2 Restore the number of channels to get the compressed feature F O , compressed feature F O The expression is:

[0021] F o =Conv fc2 (GELU(Conv fc1 (F I )));

[0022] For the compressed feature F O Perform average pooling and maximum pooling to obtain two feature maps F avg and F max , F avg and F max The expressions are:

[0023] F avg =GAP(F O );

[0024] F max =GMP(F O );

[0025] The two feature maps F avg and

[0026] F max Splice them together in the channel dimension to get a merged feature map F cat , feature map F cat The expression is:

[0027] F cat =Concat(F avg ,F max ,dim=1);

[0028] The spatial attention information of the feature map F_cat is extracted through convolution operation to obtain the spatial attention map M s , spatial attention map M s The expression is:

[0029] M s =σ(Ψ3(F cat ));

[0030] Finally, let the spatial attention map M s With 3D feature F I Multiply to get the key feature map F out , key feature graph F out The expression is:

[0031] F out =M s ·F in .

[0032] Furthermore, in step 4, the key feature graph F out Input to the multi-scale aggregation module and extract features through a set of deep convolutions. The output F of the mth hole convolution layer is (m) for:

[0033] F (m) =DWConv 3×3×3 (X, dilation = d m ),m=1,2,3,4

[0034] Where dilation = d m is the expansion rate;

[0035] The key feature graph F out With output F (m) Aggregate together to get the aggregate feature map F agg , aggregate feature map F agg The expression is:

[0036]

[0037] Apply point-by-point convolution to integrate the aggregated feature map F agg , realize cross-channel interaction and generate optimized features F out , optimized feature F out The expression is:

[0038] F out =ReLU(BN(Conv 1×1×1 (F agg )));

[0039] Among them, Conv 1×1×1Represents point-by-point convolution, BN is batch normalization, and ReLU introduces nonlinear activation;

[0040] Key Feature Map F out The context aggregation block input to the multi-scale aggregation module extracts the relative position feature map Y, and the expression of the relative position feature map Y is:

[0041] Y=F out +A·M(Softmax(K(F out ))·V(F out ));

[0042] Among them, A is the attention map generated by the sigmoid function; K(·), V(·), and M(·) are linear projections achieved by convolution.

[0043] The beneficial effects of the present invention are:

[0044] A medical image segmentation method for superficial femoral artery calcification of the lower limbs in the present invention utilizes a three-way xLSTM module to model from three directions: forward, reverse, and slice mixing, to comprehensively capture spatial features and global correlations. Accurate segmentation combines a wavelet transform module to decompose high- and low-frequency information, cooperates with channel and spatial attention mechanisms to focus on key features, and captures multi-scale spatial information through a multi-scale aggregation module, thereby improving the segmentation accuracy and robustness of the calcified area; achieving efficient and accurate calcification segmentation and improving the extraction quality of the calcified area. The implementation of this method can provide more reliable auxiliary decision support for clinicians, helping doctors to better diagnose and plan treatment when faced with complex lesions and image data; it solves the problems of existing assessment methods for lower limb artery calcification, such as difficulty in locating calcification targets, blurred boundaries, and insufficient capture of multi-scale features. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flowchart of a medical image segmentation method for superficial femoral artery calcification of the lower limb.

[0046] Figure 2 Schematic diagram of the three-way xLSTM module.

[0047] Figure 3 and Figure 4 Schematic diagram of the wavelet transform module.

[0048] Figure 5 Schematic diagram of the multi-scale aggregation module.

[0049] Figure 6 Illustration of replacing matrix multiplication between query and key with lightweight projection. DETAILED DESCRIPTION

[0050] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0051] like Figure 1 As shown, the present invention provides a medical image segmentation method for superficial femoral artery calcification of lower limbs, which includes:

[0052] Step 1: Collect CTA scan data of superficial femoral artery calcification in the lower limbs of multiple patients, and select multiple slices from the CTA scan data at a certain interval for discrete marking;

[0053] Step 2: A three-way xLSTM module is used to comprehensively capture the dynamic changes and spatial characteristics of blood vessels in multiple discretely labeled slices from three directions: forward, backward, and slice mixing, to obtain 3D features.

[0054] Step 3: Use the wavelet transform module to decompose and process the 3D features to obtain the key feature map;

[0055] Step 4: Use the multi-scale aggregation module to enhance the key feature map and identify the calcified area and the relative position between the calcified area and the blood vessel to obtain the segmentation result.

[0056] Specifically, in step 1, lower limb CTA scan data were collected from 60 patients at the hospital. After excluding arteries without calcification, CTA scan data from 44 patients were ultimately obtained. Each CTA scan contains approximately 500-600 slices. Due to the high cost of manual labeling and to prevent excessive similarity between adjacent slices from affecting network generalization performance, 100 slices were selected at regular intervals for discrete labeling. Labeling categories included background, aortic lumen, and aortic calcification. Discrete labeling was performed under the guidance of a radiologist to ensure labeling quality and accuracy.

[0057] In the process of medical image processing, especially in the task of segmenting calcification of the superficial femoral artery of the lower limbs, the direction of the blood vessels usually presents a complex three-dimensional structure, and its imaging slices often have obvious correlation between the front and back slices. For the blood vessels of the lower limbs, especially in areas with lesions (such as vascular obstruction, stenosis, etc.), the continuity and correlation of the front and back slices become particularly important. The traditional xLSTM network is mainly based on a single-direction sequence modeling method. Although it can capture temporal features, it is difficult to effectively process the complex temporal information and global correlation of blood vessels in medical images. The ability of single-direction modeling is insufficient, especially when dealing with complex lesions in the superficial femoral artery of the lower limbs, which can easily lead to the lack of vascular information in the front and back slices, affecting the segmentation accuracy. The structure of the three-way xLSTM module in the present invention is as follows: Figure 2 As shown in the figure, there are two xLSTM modules, which are the main components of the model. xLSTM may be an extension or improvement of the traditional LSTM, capable of processing and predicting time-based series data.

[0058] There is a "Flip" operation on the left side of the figure, which may be a data preprocessing or feature transformation operation used to adjust the order or direction of the input data.

[0059] "Original" represents the original input data, and "Sliced" represents the input data after the slice operation. These input data are fed into the xLSTM module for processing.

[0060] The middle module contains multiple operations, including "Norm" (normalization), "Linear" (linear transformation), "SiLU" (Sigmoid Linear Unit, an activation function), and "Multi-Head Attention" (multi-head attention mechanism). These operations work together to extract features and transform the input data.

[0061] The "Attention Block" on the right represents the attention mechanism module, which is used to weight the input data so that the model can focus on important features.

[0062] The "Multiplication" and "Add" operations in the figure represent the element-wise multiplication and addition operations performed in the model, which are used to combine and adjust the outputs of different parts.

[0063] Furthermore, in step 2, the method for obtaining 3D features includes:

[0064] The multiple slices after discrete labeling are expanded into sequences along the following three directions, and the features are modeled and fused from the three directions of forward, reverse, and inter-slice mixing; in each direction, the xLSTM layer with shared weights is used to extract feature dependencies. The forward sequence zf Expand along the slice order direction (D, H, W), modeling forward feature dependency, reverse sequence z r Expand along the reverse direction of the slice (D, H, W), supplement the reverse feature information, and slice the mixed sequence z s Mix multiple slices in the channel dimension to perform feature interaction across slices;

[0065] The calculation formula of the three-way xLSTM module is:

[0066] T o M(z)=xLSTM(z f )+xLSTM(z r )+xLSTM(z s )#

[0067] Among them, T o M(z) represents the output of three-dimensional feature modeling, and xLSTM represents a module that extracts temporal features along a specific sequence. This three-dimensional modeling approach allows the model to comprehensively capture the dynamic changes and spatial characteristics of blood vessels from different directions, improving the robustness and accuracy of the segmentation model for calcified areas of the superficial femoral artery in the lower extremities. In particular, in complex situations such as vascular occlusion, the normal vascular information from the preceding and following slices provides effective contextual support for the model, significantly improving the global consistency of vascular segmentation.

[0068] The features of the three directions are weighted and fused to obtain the 3D feature z fused , 3D feature z fused The expression is:

[0069] z fused =αz f +βz r +γz s

[0070] Here, α, β, and γ are the directional weights for the forward, reverse, and inter-slice mixing directions, respectively. By controlling the contribution of each direction to the global features and weightedly fusing feature representations from the three directions, we learn the directional weights and control the contribution of each directional feature to the global features. This not only enhances the model's feature representation capabilities but also further improves the coherence of vascular pathways and segmentation accuracy.

[0071] In traditional deep learning models, especially in multi-stage or skip connection structures, there is usually a problem of feature loss after downsampling. Although the downsampling process can reduce the amount of calculation, it often leads to loss of details of small lesions, especially in three-dimensional medical images. Although skip connections can partially alleviate this problem, they usually use a simple feature splicing method and lack effective separation and optimization of low-frequency global information and high-frequency local details, resulting in insufficient retention of fine-grained features. The wavelet transform module in the present invention is as follows: Figures 3 and 4 As shown, in Figure 3 In the figure, the multiple "Dilated Encoder" modules on the left represent dilated convolutional encoders. Dilated convolution is a special convolution operation that expands the receptive field by inserting holes (i.e., the dilation rate) between convolution kernel elements. This allows it to capture a wider range of contextual information without increasing the number of parameters or computational effort. These encoder modules extract features from the input data layer by layer. As the number of layers increases, the receptive field gradually increases, enabling the extraction of higher-level, more global features.

[0072] The "Wavelet Block" module in the center of the image represents the wavelet transform block. The wavelet transform is a mathematical tool that decomposes signals or images into subbands of varying frequencies, enabling analysis of data at multiple scales. Introducing wavelet blocks into neural networks leverages the multi-resolution analysis capabilities of the wavelet transform, enhancing the model's ability to capture features at different scales and improving its performance in processing complex data.

[0073] The "Decoder" module on the right side of the figure is responsible for gradually restoring the features extracted by the encoder into outputs related to the input data. The decoder typically uses the opposite operation of the encoder, such as deconvolution (transposed convolution), to gradually upsample the feature map and ultimately generate the target output.

[0074] exist Figure 4 On the left side of the image is the Wavelet Transform. The wavelet transform is a mathematical tool that decomposes a signal or image into subbands of different frequencies, allowing analysis of data at multiple scales. Introducing the wavelet transform into a neural network leverages its multi-resolution analysis capabilities, enhancing the model's ability to capture features at different scales.

[0075] "UP" usually represents an upsampling operation, which is used to increase the resolution of the feature map; "DOWN" represents a downsampling operation, which is used to reduce the resolution of the feature map. These two operations can be combined to perform feature processing and information fusion at different scales.

[0076] "GAP" stands for Global Average Pooling, which aggregates global information by averaging all pixels in the feature map. "GMP" stands for Global Max Pooling, which extracts the most representative features by taking the maximum value in the feature map. These two operations can extract global features from different perspectives.

[0077] CAP and CMP are specially designed attention mechanism modules that weight and select features, enhancing the model's ability to focus on important features. The modules are connected by arrows in the figure, indicating the direction of data flow. The outputs of different modules are fused (e.g., multiplication) to integrate multi-scale and multi-type feature information, and the final output is the processed result.

[0078] Specifically, in step 3, the wavelet transform module decomposes the 3D features to obtain high-frequency components and low-frequency components; the high-frequency components are used to capture detailed information, especially the changes in subtle structures such as the boundaries and textures of the lesions; the low-frequency components retain global context information, which can help the model understand the overall structure and context information in the image, and are very important for the overall recognition of calcified lesions. The high-frequency components and low-frequency components are spliced together to obtain the 3D feature F I ;

[0079] 3D feature F I First, pass through a 1x1x1 convolution layer Conv fc1 The number of channels is compressed, and then passed through the GELU activation function, and then through another identical convolutional layer Conv fc2 Restore the number of channels to get the compressed feature F O , compressed feature F O The expression is:

[0080] F O =Conv fc2 (GELU(Conv fc1 (F I )));

[0081] Compared to conventional fully connected layers, 3D convolutions naturally adapt to 3D data while reducing the number of parameters through shared weights, thereby improving computational efficiency while preserving spatial structural information. This design enhances nonlinear feature extraction capabilities through simple and efficient inter-channel mapping, making it suitable for processing 3D data such as medical images.

[0082] For the compressed feature F O Perform average pooling and maximum pooling to obtain two feature maps F avg and F max , F avg and Fmax The expressions are:

[0083] F avg =GAP(F O );

[0084] F max =GMP(F o );

[0085] The two feature maps F avg and

[0086] F max Splice them together in the channel dimension to get a merged feature map F cat , feature map F cat The expression is:

[0087] F cat =Concat(F avg ,F max ,dim=1);

[0088] The spatial attention information of the feature map Fcat is extracted through convolution operation to obtain the spatial attention map M s , spatial attention map M s The expression is:

[0089] M s =σ(Ψ3(F cat ));

[0090] Finally, let the spatial attention map M s With 3D feature F I Multiply to get the key feature map F out , key feature graph F out The expression is:

[0091] F out =M s ·F in .

[0092] The output of the squeeze enhancement branch is then multiplied pixel by pixel with the output of the attention enhancement branch, which can effectively combine the enhancement effects of channel and spatial information, ensuring that the model retains important information while suppressing redundant or irrelevant information during feature fusion, thereby better capturing complex boundaries and tiny lesions, and improving performance in the task of segmenting superficial femoral artery calcification in the lower limb.

[0093] like Figure 5 As shown in FIG, the multi-scale aggregation module includes a hole convolution and a context aggregation block; Figure 5In the figure, DWConv3x3 represents the 3x3 convolution operation in depthwise separable convolution. Depthwise separable convolution splits the standard convolution operation into two independent steps: depthwise convolution and pointwise convolution. The "3x3" here refers to the size of the convolution kernel. Depthwise convolution performs convolution on each input channel separately, while pointwise convolution uses a 1x1 convolution kernel to combine features from different channels. This convolution method can significantly reduce the amount of computation and the number of parameters compared to standard convolution.

[0094] Context Aggregation is a block that aggregates contextual information from the input data and enhances the feature representation of the current location by considering surrounding data points. This helps the model better understand the global structure and semantic information of the data.

[0095] Dilated Convolution is a dilated convolution. It inserts holes (i.e., dilation rate) between the elements of the standard convolution kernel, thereby expanding the receptive field without increasing the number of parameters and computational complexity. This enables the model to capture a wider range of contextual information, which is very useful for processing data with long-range dependencies.

[0096] Furthermore, in step 4, the key feature graph F out Input to the multi-scale aggregation module and extract features through a set of deep convolutions. The output F of the mth hole convolution layer is (m) for:

[0097] F (m) =DWConv 3×3×3 (X, dilation = d m ),m=1,2,3,4

[0098] Where dilation = d m is the expansion rate; is d m =m×3. This design enables each convolution to capture contextual features at different scales, corresponding to dilation rates of 3, 6, 9, and 12. Next, the outputs of these layers are aggregated with the input feature maps.

[0099] The key feature graph F out With output F (m) Aggregate together to get the aggregate feature map F agg , aggregate feature map F agg The expression is:

[0100]

[0101] Apply point-by-point convolution to integrate the aggregated feature map F agg , realize cross-channel interaction and generate optimized features F out , optimized feature F out The expression is:

[0102] F out =ReLU(BN(Conv 1×1×1 (F agg )));

[0103] Among them, Conv 1×1×1 Represents point-by-point convolution, BN is batch normalization, and ReLU introduces nonlinear activation;

[0104] Key Feature Map F out The context aggregation block input to the multi-scale aggregation module extracts the relative position feature map Y, and the expression of the relative position feature map Y is:

[0105] Y=F out +A·M(Softmax(K(F out ))·V(F out ));

[0106] Among them, A is the attention map generated by the sigmoid function; K(·), V(·), and M(·) are linear projections achieved by convolution.

[0107] like Figure 6 As shown, in Figure 6 In , the relative position feature map Y simplifies the traditional self-attention mechanism by replacing the matrix multiplication between the query and the key with a lightweight projection, significantly reducing the computational complexity while effectively preserving the spatial context information. Figure 6 In the figure, "Conv 1x1" represents a 1x1 convolution operation. The 1x1 convolution kernel is primarily used to adjust the number of channels. It linearly combines features from different channels without changing the spatial size of the feature map, achieving dimensionality reduction or upscaling, as well as information fusion between channels. The figure shows three "Conv 1x1" modules, each performing a different feature transformation on the input data.

[0108] The activation function Sigmoid maps the input value to between 0 and 1. In the attention mechanism, the output of the Sigmoid function is often used to generate attention weights, which indicate the degree of attention paid to different features or spatial positions.

[0109] The Softmax activation function is used in multi-classification tasks to convert the input vector into a probability distribution. In this figure, the multiplication symbol × represents a multiplication operation between features, which is used to fuse features from different branches. For example, features processed by sigmoid are multiplied with features processed by other convolution operations to achieve feature weighting. Through these technical approaches, our model not only effectively identifies and segments calcified areas in the superficial femoral artery of the lower extremity, but also enhances our understanding of the complex spatial relationship between blood vessels and calcified areas, promoting the development of arterial calcification imaging analysis technology.

[0110] To evaluate the segmentation methods for superficial femoral artery calcification in lower extremity medical images, DSC and NSD were used as evaluation metrics. DSC is a widely used overlap evaluation metric for segmentation tasks, quantifying the similarity between the predicted segmented region and the true segmented region. Its value range is [0, 1], with values closer to 1 indicating a higher degree of overlap between the segmentation result and the true region. A represents the set of voxels in the predicted segmented region, and B represents the set of voxels in the true segmented region.

[0111]

[0112] NSD is an indicator for evaluating the similarity of segmentation boundary results. It is particularly suitable for evaluating the boundary quality of target regions in medical images. NSD is calculated based on the distance between the predicted segmentation surface points and the true segmentation surface points, and measures the closeness of the segmentation surfaces through normalization. Where S_P represents the point set of the predicted segmentation surface, S_G represents the point set of the true segmentation surface, d(p,q) represents the Euclidean distance from point p to point q, τ is a predefined distance threshold, and 1 is an indicator function that takes the value 1 when the condition is met and 0 otherwise.

[0113]

[0114] In order to evaluate the medical image segmentation method of calcification of the superficial femoral artery of the lower limb in detail, this embodiment uses the CTA angiography image dataset of the superficial bony arteries of the lower limbs provided by the hospital, excluding arteries without calcification or implanted metal implants, totaling 40 cases. It contains arterial angiography sequences and label data corresponding to each image. The original size of each CTA slice is 512x512, containing approximately 300-400 CT slices. We discretely labeled 100 of them. Due to the limitation of manual labeling costs, the total amount of labeling is limited. If all slices in the CT volume are labeled, the similarity between adjacent slices will be too high, and the number of labeled cases will be reduced, which is not conducive to improving the generalization performance of the network. The labeled categories include background, aorta, and calcification. We randomly selected 8 cases in the dataset as the test set and the remaining 32 cases as the training set.

[0115] For comparative analysis, we selected several well-known competitors as baselines, including the CNN-based methods SegresNet and nnU-Net, the Transformer-based methods UNETR and SwinUNETR, and the Mamba-based method UMamba. The comparison results are shown in Table 1. Ablation studies of the design blocks are shown in Table 2. Specifically, the introduction of the three-way xLSTM module alone can bring improvements, demonstrating its excellent performance in comprehensively modeling temporal information and inter-slice dependencies, improving the precise localization of blood vessels and calcifications, and effectively alleviating the problems of vascular path interruption and calcification mislocation. Furthermore, the combination of the wavelet transform module further improves segmentation performance, alleviating the problem of information loss during downsampling that exacerbates the blurring of calcification boundaries. Notably, when the three-way xLSTM module is combined with the wavelet transform module and the multi-scale aggregation module, all indicators achieve the highest gains, highlighting the synergistic effect of these components in capturing complex spatial features and improving overall segmentation accuracy.

[0116] Table 1 Comparison of multi-view baseline models by skeleton network

[0117]

[0118] Table 2 Comparative experiments based on EMA attention module and other attention modules

[0119]

[0120]

[0121] In summary, a medical image segmentation method for superficial femoral artery calcification of the lower limbs in the present invention uses a three-way xLSTM module to model from three directions: forward, reverse, and slice mixing, to fully capture spatial features and global correlations. Accurate segmentation combines the wavelet transform module to decompose high- and low-frequency information, cooperates with the channel and spatial attention mechanism to focus on key features, and captures multi-scale spatial information through the multi-scale aggregation module, thereby improving the segmentation accuracy and robustness of the calcified area; achieving efficient and accurate calcification segmentation and improving the extraction quality of the calcified area. The implementation of this method can provide more reliable auxiliary decision support for clinical practice, helping doctors to better diagnose and plan treatment when faced with complex lesions and image data; it solves the problems of existing assessment methods for lower limb artery calcification, such as difficulty in locating calcification targets, blurred boundaries, and insufficient capture of multi-scale features.

Claims

1. A medical image segmentation method for superficial femoral artery calcification of lower limbs, characterized by: include: Step 1: Collect CTA scan data of superficial femoral artery calcification in the lower limbs of multiple patients, and select multiple slices from the CTA scan data at a certain interval for discrete marking; Step 2: A three-way xLSTM module is used to comprehensively capture the dynamic changes and spatial characteristics of blood vessels in multiple discretely labeled slices from three directions: forward, backward, and slice mixing, to obtain 3D features. Step 3: Use the wavelet transform module to decompose and process the 3D features to obtain the key feature map; Step 4: Use the multi-scale aggregation module to enhance the key feature map and identify the calcified area and the relative position between the calcified area and the blood vessel to obtain the segmentation result.

2. The medical image segmentation method for superficial femoral artery calcification of lower limbs according to claim 1, characterized in that: In step 1, the labeled categories include background, aortic lumen, and aortic calcification.

3. The medical image segmentation method for superficial femoral artery calcification of lower limbs according to claim 2, characterized in that: In step 2, the method of obtaining 3D features includes: The multiple slices after discrete labeling are expanded into sequences along the following three directions, and the features are modeled and fused from the three directions of forward, reverse, and inter-slice mixing; in each direction, the xLSTM layer with shared weights is used to extract feature dependencies. The forward sequence z f Expand along the slice order direction (D, H, W), modeling forward feature dependency, reverse sequence z r Expand along the reverse direction of the slice (D, H, W), supplement the reverse feature information, and slice the mixed sequence z s Mix multiple slices in the channel dimension to perform feature interaction across slices; The calculation formula of the three-way xLSTM module is: T o M(z)=xLSTM(z f )+xLSTM(z r )+xLSTM(z s )# Among them, T o M(z) represents the output after three-way feature modeling, and xLSTM represents the module for temporal feature extraction along a specific sequence; The features of the three directions are weighted and fused to obtain the 3D feature z fused , 3D feature z fused The expression is: With fused =αz f +βz r +γz s Among them, α, β, and γ are the directional weights of the forward, reverse, and inter-slice mixing directions respectively.

4. The medical image segmentation method for superficial femoral artery calcification of lower limbs according to claim 3, characterized in that: In step 3, the wavelet transform module decomposes the 3D features to obtain high-frequency components and low-frequency components; the high-frequency components are used to capture detail information; the low-frequency components retain global context information, and the high-frequency components and low-frequency components are concatenated to obtain the 3D feature F I ; 3D feature F I First, pass through a 1x1x1 convolution layer Conv fc1 The number of channels is compressed, and then passed through the GELU activation function, and then through another identical convolutional layer Conv fc2 Restore the number of channels to get the compressed feature F O , compressed feature F O The expression is: F o =Conv fc2 (GELU(Conv fc1 (F I ))); For the compressed feature F O Perform average pooling and maximum pooling to obtain two feature maps F avg and F max , F avg and F max The expressions are: F avg =GAP(F O ); F max =GMP(F O ); The two feature maps F avg and F max Splice them together in the channel dimension to get a merged feature map F cat , feature map F cat The expression is: F cat =Concat(F avg ,F max ,dim=1); The spatial attention information of the feature map F_cat is extracted through convolution operation to obtain the spatial attention map M s , spatial attention map M s The expression is: M s =σ(Ψ3(F cat )); Finally, let the spatial attention map M s With 3D feature F I Multiply and get the key feature map F out , key feature map M out The expression is: F out =M s ·F in 。 5. The medical image segmentation method for superficial femoral artery calcification of lower limbs according to claim 4, characterized in that: In step 4, the multi-scale aggregation module includes a dilated convolution and a context aggregation block; Key Feature Map F out Input to the multi-scale aggregation module and extract features through a set of deep convolutions. The output F of the mth hole convolution layer is (m) for: F (m) =DWConv 3×3×3 (X,dilation=d m ),m=1,2,3,4 Where dilation = d m is the expansion rate; The key feature graph F out With output F (m) Aggregate together to get the aggregate feature map F agg , aggregate feature map F agg The expression is: Apply point-by-point convolution to integrate the aggregated feature map F agg , realize cross-channel interaction and generate optimized features F out , optimized feature F out The expression is: F out =ReLU(BN(Conv 1×1×1 (F agg ))); Among them, Conv 1×1×1 Represents point-by-point convolution, BN is batch normalization, and ReLU introduces nonlinear activation; Key Feature Map F out The context aggregation block input to the multi-scale aggregation module extracts the relative position feature map Y, and the expression of the relative position feature map Y is: Y=F out +A·M(Softmax(K(F out ))·V(F out )); Among them, A is the attention map generated by the sigmoid function; K(·), V(·), and M(·) are linear projections achieved by convolution.

Images