A liver tumor segmentation method based on multi-directional spatial features

Abstract

The application discloses a liver tumor segmentation method based on multi-directional spatial features, which aims to realize accurate liver tumor segmentation. The method comprises the following steps: S1: preprocessing abdominal CT data to obtain NPY matrix data, and dividing the data into a training set, a validation set and a test set; S2: extracting initial features of the NPY matrix data; S3: extracting multi-directional spatial features from the initial features; S4: constructing an adaptive gating attention module (AGAM) to extract features of the multi-directional spatial features and obtain local features; S5: constructing a six-directional Mamba module (HoM) to extract features of the multi-directional spatial features and obtain global features; S6: fusing and enhancing the local features and the global features to obtain reinforced features; S7: constructing a channel parsing entropy module (CPEM) to screen the reinforced features and obtain final features; and S8: splicing the multi-scale final features to obtain a liver tumor segmentation map. The application can effectively improve the liver tumor segmentation precision and can be used for liver tumor segmentation in CT images.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a liver tumor segmentation method based on multi-directional spatial features. Background Art

[0002] Liver tumors are abnormal cell clusters in the liver, posing a serious threat to patients' lives and health. Computed tomography (CT) is an indispensable tool in modern clinical medicine, playing a particularly crucial role in the diagnosis of liver tumors. However, reading and analyzing CT images often requires extensive clinical experience and considerable time and effort. Traditional manual reading methods can miss complex or subtle lesions, impacting the accuracy and timeliness of diagnosis.

[0003] Existing liver tumor segmentation methods fall into two main categories: traditional machine learning and deep learning. Traditional machine learning methods are limited by the diversity of tumor morphology, size, and location, as well as fuzzy boundaries. They struggle to cope with complex medical image features, limiting their application. In recent years, deep learning has made significant progress in segmentation tasks. Its deep modeling capabilities for complex features enable it to better capture complex semantic information, surpassing traditional machine learning algorithms in many respects.

[0004] In recent years, numerous studies on liver tumor segmentation have demonstrated that the use of attention mechanisms can enhance the recognition of tumor edge features and microtumors, effectively alleviating the problems of blurred boundaries and inaccurate segmentation. However, a single attention mechanism still has shortcomings in modeling global semantic information and extracting local edge features. It is difficult to capture global features while also fully accounting for the extraction of local details, limiting further improvements in segmentation performance.

[0005] On the other hand, when processing 3D medical imaging data, while directly using 3D convolution kernels can effectively extract spatial features, it consumes significant computational resources. Transformer-based methods, while adept at capturing global information, also require high computing power because their computational complexity increases quadratically with the feature dimension.

[0006] Therefore, developing an automated liver tumor segmentation method that can ensure high segmentation accuracy while also having rapid response capabilities and efficient computing performance is of great clinical significance in helping doctors detect lesions earlier and formulate treatment plans in a timely manner. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention provides a liver tumor segmentation method based on multi-directional spatial features, which can provide users with a more accurate, efficient and highly adaptable liver tumor segmentation solution.

[0008] Specifically, the method comprises the following steps:

[0009] S1: Preprocess abdominal CT data to obtain NPY matrix data, and divide the data into training set, validation set and test set;

[0010] S2: Extract the initial features of NPY matrix data;

[0011] S3: Extract multi-directional spatial features from the initial features;

[0012] S4: Construct an adaptive gated attention module AGAM to extract multi-directional spatial features and obtain local features;

[0013] S5: Construct a six-directional Mamba module HoM to extract multi-directional spatial features and obtain global features;

[0014] S6: Fuse and enhance local features and global features to obtain enhanced features;

[0015] S7: Construct a channel parsing entropy module (CPEM) to filter the enhanced features and obtain the final features.

[0016] S8: Concatenate the final features at multiple scales to obtain the segmentation map of the liver tumor.

[0017] Preferably, S1 comprises the following steps:

[0018] S1.1: Collect the original abdominal CT data, perform preprocessing of the data by standardization, cropping, and data enhancement, and convert it into NPY matrix data ;

[0019] S1.2: Divide the preprocessed data into training set, validation set and test set.

[0020] Preferably, S2 comprises the following steps:

[0021] Introduce a deep convolution to NPY matrix data Perform convolution operation to obtain initial features , the operation is described as:

[0022]

[0023] in, Representative stride length , the padding size is Depthwise convolution.

[0024] Preferably, S3 includes the following steps:

[0025] Construct a multi-directional spatial feature extractor MDFE (Multi Directional Feature Extractor). The operation is described as follows:

[0026] Introduce a set of spatial separation convolutions to the initial features Continuous convolution operations are performed along the horizontal, vertical and depth directions to obtain multi-directional spatial features , which is expressed as follows:

[0027]

[0028]

[0029]

[0030] In the above formula, represent The output after horizontal convolution, represent The output after vertical convolution, express The output after depthwise convolution represents multi-directional spatial features. Represents the horizontal convolution calculation, Represents the vertical convolution calculation, Represents the depth-wise convolution calculation.

[0031] Preferably, S4 includes the following steps:

[0032] S4.1: Extract features from cross-sectional slices of multi-directional spatial features;

[0033] S4.2: Calculate attention weights;

[0034] S4.3: Local attention enhancement of multi-directional spatial features to obtain local features .

[0035] Preferably, S4.1 includes the following steps:

[0036]

[0037] In the above formula, Represents cross-sectional characteristics;

[0038] Preferably, S4.2 includes the following steps:

[0039] The multi-directional spatial features are up-sampled and down-sampled, and attention weights are generated through convolution. The operation is described as:

[0040]

[0041]

[0042]

[0043]

[0044] In the above formula, represent The output after downsampling operation is represent The output after the convolution operation is represent The output after upsampling operation is represents the weight of attention, stands for downsampling, which means reducing the spatial resolution of the input. stands for upsampling, which means increasing the spatial resolution of the input. stands for normalization;

[0045] Preferably, S4.3 includes the following steps:

[0046] Using attention weights Adjust the multi-directional spatial features and combine them with the cross-sectional features Perform fusion, which is described as:

[0047]

[0048] In the above formula, Represents the use of attention weights to adjust multi-directional spatial features, Represents the output of the adaptive gated attention module AGAM.

[0049] Preferably, S5 includes the following steps:

[0050] S5.1: Construct Mamba long sequence modeling operation;

[0051] S5.2: Flatten the coronal, sagittal, and transverse sequences of multi-directional spatial features into Mamba long sequence models.

[0052] Preferably, S5.1 includes the following steps:

[0053] The dimension of the feature is ,definition , , ;

[0054] in, A 2D flattened sequence representing features, Representative features, Representative characteristics The output after Mamba long sequence modeling, represent A 2D flattened sequence, Represents Mamba long sequence modeling operation;

[0055] Mamba long sequence modeling operations , defined as follows:

[0056] right After linear transformation and convolution operation, its expression is:

[0057]

[0058] In the above formula, represents a linear transformation;

[0059] Introducing parameters and , whose expression is:

[0060]

[0061]

[0062] In the above formula, Representative basis The dynamically generated input transformation matrix, Representative basis The dynamically generated input transformation matrix, Representatives specifically for The linear transformation of Representatives specifically for Linear transformation of

[0063] Construct a state space model, which is expressed as:

[0064]

[0065]

[0066] In the above formula, represents the hidden state of the state-space model, represents the predicted output of the state-space model, represents the state transition matrix;

[0067] Based on the state space model, dynamic time step is introduced whose expression is:

[0068]

[0069] In the above formula, represents a linear transformation specially for ;

[0070] constructing a time-varying state space model, whose expression is:

[0071]

[0072]

[0073] In the above formula, , represents a matrix discretized by a time step , represents a diagonal matrix, represents a natural exponential function;

[0074]

[0075]

[0076] In the above formula, represents the hidden state of the time-varying state space model, represents the predicted output of the time-varying state space model;

[0077] The multi-directional space feature is subjected to a Mamba long sequence modeling operation to obtain whose expression is:

[0078]

[0079] In the above formula, represents the output of the multi-directional space feature after Mamba long sequence modeling;

[0080] Preferably, S5.2 comprises the following steps:

[0081] The multi-directional space feature is flattened into 6 feature sequences, Mamba long sequence modeling is performed on the 6 sequences respectively, and then the modeled sequences are fused to realize global attention reinforcement of the multi-directional space feature, and the operation is described as:

[0082]

[0083] In the above formula, represents the output of the six-direction Mamba module HoM, Represents multidirectional spatial features Anterior flattening sequence of the sagittal plane, Represents multidirectional spatial features Posterior flattening sequence of the sagittal plane, Represents multidirectional spatial features The superior flattening sequence of the coronal plane, Represents multidirectional spatial features The inferior flattening sequence of the coronal plane, Represents multidirectional spatial features The left-hand flattened sequence of the cross section, Represents multidirectional spatial features Right-sided flattened sequence of cross sections.

[0084] Preferably, S6 includes the following steps:

[0085] S6.1: Fuse the outputs of the adaptive gated attention module AGAM and the six-way Mamba module HoM to obtain the fused features ;

[0086] S6.2: Enhance the fused features to obtain enhanced features.

[0087] Preferably, S6.1 includes the following steps:

[0088] The multi-directional spatial features enhanced by the six-way Mamba attention module HoM and the adaptive gated attention module AGAM are combined to obtain the fused features , the operation is described as:

[0089]

[0090] In the above formula, represents fusion features;

[0091] Preferably, S6.2 includes the following steps:

[0092] Fusion features Perform layer normalization to balance the distribution of features, and then use a multi-layer perceptron to perform nonlinear transformation and feature mapping to further enhance the expressiveness of features and obtain enhanced features. , the operation is described as:

[0093]

[0094] In the above formula, Represents the strengthening feature, represents a multilayer perceptron, Representative layer normalization.

[0095] Preferably, S7 includes the following steps:

[0096] S7.1: Calculate channel selectivity coefficients for enhanced features;

[0097] S7.2: Calculate uncertainty estimates for enhanced characteristics;

[0098] S7.3: Construct the channel parsing entropy module CPEM to screen and enhance features.

[0099] Preferably, S7.1 includes the following steps:

[0100] Perform global average pooling on the input reinforcement features, map the reinforcement features of each channel to a scalar, and then input it into the multi-layer perceptron. Model the importance of each channel through nonlinear transformation, further extract the dependency between channels, and finally perform normalization to map the important coefficient of each channel to the range of [0, 1] to generate the channel selection coefficient. This operation is described as:

[0101]

[0102] In the above formula, represents the channel selection coefficient, stands for global average pooling;

[0103] Preferably, S7.2 includes the following steps:

[0104] Given the current feature dimension C, calculate the channel average of the current dimension to obtain the normalized representation of the single-scale feature , the operation is described as:

[0105]

[0106] In the above formula, Representative Enhanced features of each channel;

[0107] Use the information entropy formula to calculate the uncertainty estimation coefficient of the feature , the operation is described as:

[0108]

[0109] In the above formula, is a constant, Represents information entropy, which measures the uncertainty of the feature. A large entropy value indicates high uncertainty of the feature.

[0110] Preferably, S7.3 includes the following steps:

[0111] Adjust each channel feature and integrate the importance weight of the channel And uncertainty estimation coefficient , get the screened features, which is described as:

[0112]

[0113] In the above formula, represent the final features.

[0114] Preferably, S8 comprises the following steps:

[0115] S8.1: Splice the final features of each scale to obtain a final feature map;

[0116] S8.2: Build a segmentation head to convert the final feature map into a segmentation map;

[0117] S8.3: Use the training set, validation set and test set to train the model, verify the optimization and test.

[0118] The beneficial effects of the present application are: first, in the data preprocessing stage, by standardizing, cropping and data enhancement operation on the collected liver tumor images, the data quality is improved, the irrelevant information interference is reduced, and the data is reasonably divided into training set, validation set and test set, which lays a good foundation for subsequent model training and testing. In the model construction process, a kind of spatial feature extractor MDFE is proposed, which strengthens the spatial connection of features. Then, an adaptive gating attention module AGAM is proposed, which realizes the capture of local detailed features while using a small amount of parameters. Next, a six-direction Mamba module HoM is proposed, which can capture global features while using a small amount of parameters. Subsequently, a channel parsing entropy module CPEM is proposed, which is used to remove invalid features. These innovative designs make the model significantly reduce the parameter amount, speed up the reasoning speed, while maintain a high level of segmentation accuracy, providing an efficient and reliable solution for liver tumor segmentation task, showing good clinical application potential. BRIEF DESCRIPTION OF DRAWINGS

[0119] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings needed to be used in the prior art and embodiments, the following drawings are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0120] Fig. 1 is a flowchart of a liver tumor segmentation method based on multi-directional spatial features according to the present application;

[0121] Fig. 2 is a model architecture diagram of a liver tumor segmentation method based on multi-directional spatial features according to the present application;

[0122] FIG3 is a schematic diagram of the structure of an adaptive gated attention module AGAM of a liver tumor segmentation method based on multi-directional spatial features of the present invention;

[0123] FIG4 is a schematic diagram of the six-directional Mamba module HoM structure of a liver tumor segmentation method based on multi-directional spatial features of the present invention;

[0124] FIG5 is a schematic diagram of the structure of a channel parsing entropy module (CPEM) of a liver tumor segmentation method based on multi-directional spatial features according to the present invention;

[0125] Specific implementation plan

[0126] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work are within the scope of protection of the present invention.

[0127] The embodiment of the present invention provides a liver tumor segmentation method based on multi-directional spatial features, which is used to achieve accurate segmentation of liver tumors, assist clinical diagnosis and treatment decision-making, and improve the accuracy and efficiency of medical image processing.

[0128] Reference Figure 1 , the method comprises the following steps:

[0129] S1: Preprocess abdominal CT data to obtain NPY matrix data, and divide the data into training set, validation set and test set;

[0130] S2: Extract the initial features of NPY matrix data;

[0131] S3: Extract multi-directional spatial features from the initial features;

[0132] S4: Construct an adaptive gated attention module AGAM to extract multi-directional spatial features and obtain local features;

[0133] S5: Construct a six-directional Mamba module HoM to extract multi-directional spatial features and obtain global features;

[0134] S6: Fuse and enhance local features and global features to obtain enhanced features;

[0135] S7: Construct a channel parsing entropy module (CPEM) to filter the enhanced features and obtain the final features.

[0136] S8: Concatenate the final features at multiple scales to obtain the segmentation map of the liver tumor.

[0137] Furthermore, S1 includes the following steps:

[0138] S1.1: Collect the original abdominal CT data, perform preprocessing of the data by standardization, cropping, and data enhancement, and convert it into NPY matrix data ;

[0139] Furthermore, the raw abdominal CT data are collected, the data are pre-processed by standardization, cropping, and data enhancement, and converted into NPY matrix data. This step specifically includes:

[0140] The collected data include public datasets and datasets provided by cooperative hospitals. Each case image completely covers the entire area of ​​the liver. The CT data are cropped to remove irrelevant areas, resampled to 512×512×512 size, and the pixel values ​​of each data are normalized to a uniform range (0 to 1). Then the data are converted into NPY matrix format. ;

[0141] S1.2: Divide the preprocessed data into training, validation, and test sets;

[0142] Furthermore, the step of dividing the preprocessed data into a training set, a validation set, and a test set specifically includes:

[0143] The data is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0144] Further, S2 includes the following steps:

[0145] Introduce a Convolution kernel, NPY matrix data The stride is , filled with Convolution operation obtains initial features , the operation is described as:

[0146]

[0147] in, Representative stride length , the padding size is Depthwise convolution.

[0148] Furthermore, S3 includes the following steps:

[0149] Construct a multi-directional spatial feature extractor MDFE (Multi Directional Feature Extractor). The operation is described as follows:

[0150] Introduce a set of spatially separated convolutions to the initial features Continuous convolution operations are performed along the horizontal, vertical and depth directions to obtain multi-directional spatial features , which is expressed as follows:

[0151]

[0152]

[0153]

[0154] In the above formula, represent The output after horizontal convolution, represent The output after vertical convolution, express The output after depthwise convolution represents multi-directional spatial features. Represents the horizontal convolution calculation, Represents the vertical convolution calculation, Represents the depth-wise convolution calculation.

[0155] Further, refer to Figure 3 , S4 includes the following steps:

[0156] S4.1: Extract features from cross-sectional slices of multi-directional spatial features;

[0157] S4.2: Calculate attention weights;

[0158] S4.3: Local attention enhancement of multi-directional spatial features to obtain local features .

[0159] Furthermore, S4.1 includes the following steps:

[0160] Use size The convolution kernel is used for convolution calculation, which is described as:

[0161]

[0162] In the above formula, represents cross-sectional characteristics;

[0163] Furthermore, S4.2 includes the following steps:

[0164] The multi-directional spatial features are up-sampled and down-sampled, and attention weights are generated through convolution. The operation is described as:

[0165]

[0166]

[0167]

[0168]

[0169] In the above formula, represent The output after downsampling operation is represent The output after the convolution operation is represent The output after upsampling operation is represents the weight of attention, stands for downsampling, which means reducing the spatial resolution of the input. stands for upsampling, which means increasing the spatial resolution of the input. stands for normalization;

[0170] Furthermore, S4.3 includes the following steps:

[0171] Use attention weights to adjust multi-directional spatial features and combine them with cross-sectional features Fusion is performed to achieve local attention enhancement of multi-directional spatial features. The operation is described as:

[0172]

[0173] In the above formula, Represents the use of attention weights to adjust multi-directional spatial features, Represents the output of the adaptive gated attention module AGAM.

[0174] Further, refer to Figure 4 , S5 includes the following steps:

[0175] S5.1: Construct Mamba long sequence modeling operation;

[0176] S5.2: Model the coronal, sagittal, and transverse flattened sequences and the retrograde Mamba long sequence of multidirectional spatial features.

[0177] Furthermore, S5.1 includes the following steps:

[0178] The dimension of the feature is ,definition , , ;

[0179] in, A 2D flattened sequence representing features, Representative features, Representative characteristics The output after Mamba long sequence modeling, represent A 2D flattened sequence, Represents Mamba long sequence modeling operation;

[0180] Mamba long sequence modeling operations , defined as follows:

[0181] right After linear transformation and convolution operation, its expression is:

[0182]

[0183] In the above formula, represents a linear transformation;

[0184] Introducing parameters and , whose expression is:

[0185]

[0186]

[0187] In the above formula, Representative basis The dynamically generated input transformation matrix, Representative basis The dynamically generated input transformation matrix, Representatives specifically for The linear transformation of Representatives specifically for Linear transformation of

[0188] Construct a state space model, which is expressed as:

[0189]

[0190]

[0191] In the above formula, represents the hidden state of the state-space model, represents the predicted output of the state-space model, represents the state transition matrix;

[0192] Based on the state space model, dynamic time step is introduced , whose expression is:

[0193]

[0194] In the above formula, Representatives specifically for Linear transformation of

[0195] Construct a time-varying state space model, whose expression is:

[0196]

[0197]

[0198] In the above formula, , Represents the time step The discretized matrix is represents a diagonal matrix, represents the natural exponential function;

[0199]

[0200]

[0201] In the above formula, represents the hidden state of the time-varying state space model, represents the predicted output of the time-varying state-space model;

[0202] Multidirectional spatial features After Mamba long sequence modeling operation, we get , whose expression is:

[0203]

[0204] In the above formula, Represents multidirectional spatial features Output after Mamba long sequence modeling;

[0205] Furthermore, S5.2 includes the following steps:

[0206] The multi-directional spatial features are flattened into 6 feature sequences. Mamba long sequence modeling is performed on these 6 sequences respectively. Then, the modeled sequences are fused to achieve global attention enhancement of the multi-directional spatial features. The operation is described as follows:

[0207]

[0208] In the above formula, represents the output of the six-way Mamba module HoM, Represents multidirectional spatial features Anterior flattening sequence of the sagittal plane, Represents multidirectional spatial features Posterior flattening sequence of the sagittal plane, Represents multidirectional spatial features The superior flattening sequence of the coronal plane, Represents multidirectional spatial features The inferior flattening sequence of the coronal plane, Represents multidirectional spatial features The left-hand flattened sequence of the cross section, Represents multidirectional spatial features Right-sided flattened sequence of cross sections.

[0209] Furthermore, S6 includes the following steps:

[0210] S6.1: Fuse the outputs of the adaptive gated attention module AGAM and the six-way Mamba module HoM to obtain the fused features ;

[0211] S6.2: Enhance the fused features to obtain enhanced features.

[0212] Furthermore, S6.1 includes the following steps:

[0213] The multi-directional spatial features enhanced by the six-way Mamba attention module HoM and the adaptive gated attention module AGAM are combined to obtain the fused features , the operation is described as:

[0214]

[0215] In the above formula, represents fusion features;

[0216] Furthermore, S6.2 includes the following steps:

[0217] Fusion features Perform layer normalization to balance the distribution of features, and then use a multi-layer perceptron to perform nonlinear transformation and feature mapping to further enhance the expressiveness of features and obtain enhanced features. , the operation is described as:

[0218]

[0219] In the above formula, Represents the strengthening feature, represents a multilayer perceptron, Representative layer normalization.

[0220] Further, refer to Figure 5 , S7 includes the following steps:

[0221] S7.1: Calculate channel selectivity coefficients for enhanced features;

[0222] S7.2: Calculate uncertainty estimates for enhanced characteristics;

[0223] S7.3: Construct channel parsing entropy module CPEM to screen and enhance features.

[0224] Furthermore, S7.1 includes the following steps:

[0225] Perform global average pooling on the input reinforcement features, map the reinforcement features of each channel to a scalar, and then input it into the multi-layer perceptron. Model the importance of each channel through nonlinear transformation, further extract the dependency between channels, and finally perform normalization to map the important coefficient of each channel to the range of [0, 1] to generate the channel selection coefficient. This operation is described as:

[0226]

[0227] In the above formula, represents the channel selection coefficient, stands for global average pooling;

[0228] Furthermore, S7.2 includes the following steps:

[0229] Given the current feature dimension C, calculate the channel average of the current dimension to obtain the normalized representation of the single-scale feature , the operation is described as:

[0230]

[0231] In the above formula, Representative Enhanced features of each channel;

[0232] Use the information entropy formula to calculate the uncertainty estimation coefficient of the feature , the operation is described as:

[0233]

[0234] In the above formula, is a constant, Represents information entropy, which measures the uncertainty of the feature. A large entropy value indicates high uncertainty of the feature.

[0235] Furthermore, S7.3 includes the following steps:

[0236] adjusting each channel feature, integrating importance weight of the channel and uncertainty estimation coefficient , obtaining the screened feature, and the operation is described as:

[0237]

[0238] In the above formula, represent the final feature.

[0239] Further, referring to Figure 2 , S8 includes the following steps:

[0240] S8.1: Splice the final features of each scale to obtain the final feature map;

[0241] S8.2: Build a segmentation head to convert the final feature map into a segmentation map;

[0242] S8.3: Use the training set, validation set and test set to train the model, verify the optimization and test.

[0243] Further, S8.3 includes the following steps:

[0244] Use the training set and the validation set to train the model and verify the optimization to obtain the optimal model weight, then import the optimal model weight into the test code, and use the test set to test the model.

[0245] The above is a specific description of the preferred implementation of the present application, but the present application is not limited to the described embodiments. Those skilled in the art can make other equivalent modifications or replacements without departing from the spirit of the application. These equivalent modifications or replacements are included in the scope defined by the claims.

Claims

1. A liver tumor segmentation method based on multi-directional spatial features, characterized in that the method comprises the following steps: S1: Preprocess abdominal CT data to obtain NPY matrix data, and divide the data into training set, validation set and test set; S2: Extract the initial features of NPY matrix data; S3: Extract multi-directional spatial features from the initial features; S4: Construct an adaptive gated attention module AGAM to extract multi-directional spatial features and obtain local features; specifically, it includes the following steps: S4.1: Extract features from cross-sectional slices of multi-directional spatial features; S4.2: Calculate attention weights; S4.3: Local attention enhancement of multi-directional spatial features to obtain local features ; S5: Construct a six-directional Mamba module HoM to extract multi-directional spatial features and obtain global features; specifically, it includes the following steps: S5.1: Construct Mamba long sequence modeling operation; S5.2: Flatten the coronal, sagittal, and transverse sequences of multi-directional spatial features into Mamba long sequence models; Among them, S5.2 includes the following steps: The multi-directional spatial features are flattened into 6 feature sequences, and Mamba long sequence modeling is performed on these 6 sequences. The modeled sequences are then fused to achieve global attention enhancement of the multi-directional spatial features. S6: Fusing and enhancing local features and global features to obtain enhanced features; specifically, the following steps are included: S6.1: Fuse the outputs of the adaptive gated attention module AGAM and the six-way Mamba module HoM to obtain the fused features ; S6.2: Enhance the fused features to obtain enhanced features; S7: Construct a channel parsing entropy module (CPEM) to filter the enhanced features and obtain the final features. This includes the following steps: S7.1: Calculate channel selectivity coefficients for enhanced features; S7.2: Calculate uncertainty estimates for enhanced characteristics; S7.3: Construct channel parsing entropy module (CPEM) to screen and enhance features; S7.1 includes the following steps: Perform global average pooling on the input enhanced features, map the enhanced features of each channel to a scalar, and then input it into the multi-layer perceptron. Model the importance of each channel through nonlinear transformation, further extract the dependency between channels, and finally perform normalization to map the important coefficient of each channel to the range of [0, 1] to generate the channel selection coefficient. S7.3 includes the following steps: Adjust each channel feature and integrate the importance weight of the channel and uncertainty estimation coefficients , get the filtered features; S8: Concatenate the final features at multiple scales to obtain the segmentation map of the liver tumor.

2. The method for liver tumor segmentation based on multi-directional spatial features according to claim 1, characterized in that: S1 includes the following steps: S1.1: Collect the original abdominal CT data, perform preprocessing of the data by standardization, cropping, and data enhancement, and convert it into NPY matrix data ; S1.2: Divide the preprocessed data into training set, validation set and test set.

3. The method for liver tumor segmentation based on multi-directional spatial features according to claim 2, characterized in that: S2 includes the following steps: Introduce a deep convolution to NPY matrix data Perform convolution operation to obtain initial features , the operation is described as: In the above formula, represents NPY matrix data, Representative stride length , the padding size is Depthwise convolution.

4. The method for liver tumor segmentation based on multi-directional spatial features according to claim 3, characterized in that: S3 includes the following steps: Construct a multi-directional spatial feature extractor MDFE (Multi Directional Feature Extractor). The operation is described as follows: Introduce a set of spatial separation convolutions to the initial features Continuous convolution operations are performed along the horizontal, vertical and depth directions to obtain multi-directional spatial features , which is expressed as follows: In the above formula, represent The output after horizontal convolution, represent The output after vertical convolution, express The output after depthwise convolution, 、 and Collectively called multi-directional spatial features, Represents the horizontal convolution calculation, Represents the vertical convolution calculation, Represents the depth-wise convolution calculation.

5. The method for liver tumor segmentation based on multi-directional spatial features according to claim 4, characterized in that: Each sub-step in S4 further includes: S4.1: The calculation process of extracting features from the cross-sectional slices of the multi-directional spatial features is expressed by the following formula: In the above formula, represents cross-sectional characteristics; S4.2: Multi-directional spatial features Perform up and down sampling and generate attention weights through convolution. The operation is described as: In the above formula, represent The output after downsampling operation is represent The output after the convolution operation is represent The output after upsampling operation is represents the weight of attention, stands for downsampling, which means reducing the spatial resolution of the input. stands for upsampling, which means increasing the spatial resolution of the input. stands for normalization; S4.3: Using Attention Weights Adjust the multi-directional spatial features and combine them with the cross-sectional features Perform fusion, which is described as: In the above formula, Represents the use of attention weights to adjust multi-directional spatial features, Represents the output of the adaptive gated attention module AGAM.

6. The method for liver tumor segmentation based on multi-directional spatial features according to claim 5, characterized in that: Each sub-step in S5 further includes: S5.1: Let the dimension of the feature be ,definition , , ; in, A 2D flattened sequence representing features, Representative features, Representative characteristics The output after Mamba long sequence modeling, represent A 2D flattened sequence, Represents Mamba long sequence modeling operation; Mamba long sequence modeling operations , defined as follows: right After linear transformation and convolution operation, its expression is: In the above formula, represents a linear transformation; Introducing parameters and , whose expression is: In the above formula, Representative basis The dynamically generated input transformation matrix, Representative basis The dynamically generated input transformation matrix, Representatives specifically for The linear transformation of Representatives specifically for Linear transformation of Construct a state space model, which is expressed as: In the above formula, represents the hidden state of the state-space model, represents the predicted output of the state-space model, represents the state transition matrix; Based on the state space model, dynamic time step is introduced , whose expression is: In the above formula, Representatives specifically for Linear transformation of Construct a time-varying state space model, whose expression is: In the above formula, , Represents the time step The discretized matrix is represents a diagonal matrix, represents the natural exponential function; In the above formula, represents the hidden state of the time-varying state space model, represents the predicted output of the time-varying state-space model; Multidirectional spatial features After Mamba long sequence modeling operation, we get , whose expression is: In the above formula, Represents multidirectional spatial features Output after Mamba long sequence modeling; S5.2: The calculation process of performing Mamba long sequence modeling on the flattened sequences of the coronal, sagittal, and transverse planes of the multi-directional spatial features is represented by the following formula: In the above formula, represents the output of the six-way Mamba module HoM, Represents multidirectional spatial features Anterior flattening sequence of the sagittal plane, Represents multidirectional spatial features Posterior flattening sequence of the sagittal plane, Represents multidirectional spatial features The superior flattening sequence of the coronal plane, Represents multidirectional spatial features The inferior flattening sequence of the coronal plane, Represents multidirectional spatial features The left-hand flattened sequence of the cross section, Represents multidirectional spatial features Right-sided flattened sequence of cross sections.

7. The method for liver tumor segmentation based on multi-directional spatial features according to claim 6, characterized in that: Each sub-step in S6 further includes: S6.1: Fuse the multi-directional spatial features enhanced by the six-way Mamba attention module HoM and the adaptive gated attention module AGAM to obtain the fused features , the operation is described as: In the above formula, represents fusion features; S6.2: Fusion Features Perform layer normalization to balance the distribution of features, and then use a multi-layer perceptron to perform nonlinear transformation and feature mapping to further enhance the expressiveness of features and obtain enhanced features. , the operation is described as: In the above formula, Represents the strengthening feature, represents a multilayer perceptron, Representative layer normalization.

8. The method for liver tumor segmentation based on multi-directional spatial features according to claim 7, characterized in that: Each sub-step in S7 further includes: S7.1: The calculation process of the channel selection coefficient for the enhanced feature is expressed by the following formula: In the above formula, represents the channel selection coefficient, stands for global average pooling; S7.2: Given the current feature dimension C, calculate the channel average of the current dimension to obtain the normalized representation of the single-scale feature , the operation is described as: In the above formula, Representative Enhanced features of each channel; Use the information entropy formula to calculate the uncertainty estimation coefficient of the feature , the operation is described as: In the above formula, is a constant, Represents information entropy, which measures the uncertainty of the feature. A large entropy value indicates high uncertainty of the feature. S7.3: The calculation process of constructing the channel parsing entropy module (CPEM) to screen and enhance features is represented by the following formula: In the above formula, Represents the final feature.

9. The method for liver tumor segmentation based on multi-directional spatial features according to claim 8, characterized in that: S8 includes the following steps: S8.1: Concatenate the final features of each scale to obtain the final feature map; S8.2: Build the Segmentation Head to convert the final feature map into a segmentation map; S8.3: Use the training set, validation set, and test set to train, validate, optimize, and test the model.

Images