Liver tumor segmentation method based on expert attention and deformable cross-period fusion

Abstract

The present application relates to the technical field of medical image processing, in particular to a liver tumor segmentation method based on expert attention and deformable cross-period fusion; comprising: acquiring multi-period CT image data and preprocessing; independently extracting features of each period image to obtain single-period features; performing adaptive enhancement processing on each single-period feature to highlight key regions; taking the target period enhanced feature as the main branch, the auxiliary period enhanced feature as the auxiliary branch, performing spatial alignment through deformable sampling, and performing cross-period feature interaction; generating fusion reliability weight based on the structure information of the main branch and the cross-period feature similarity, and weighting and fusing the cross-period enhanced features; decoding the fusion features to generate a segmentation result. The present application combines deformable alignment and adaptive feature selection to realize robust fusion of multi-period features without relying on accurate registration, effectively improving the accuracy and stability of liver tumor segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a liver tumor segmentation method based on expert attention and deformable interphase fusion. Background Technology

[0002] Liver tumors are among the most common malignant diseases in clinical practice. Early detection and accurate segmentation are crucial for lesion assessment, surgical planning, and monitoring of treatment efficacy. Multiphase enhanced CT can reflect the blood supply characteristics of tumors from different perspectives, including the plain scan phase, arterial phase, venous phase, and delayed phase, providing rich diagnostic evidence for tumor identification.

[0003] Currently, automatic segmentation methods for liver tumors are mainly based on deep learning models. To fully utilize the complementary information in multi-phase enhanced CT scans, existing methods typically perform multi-phase modeling through multi-phase image stitching or feature-level fusion, but these methods still have the following limitations: (1) Spatial misalignment across phases. Due to factors such as respiratory movement and organ deformation, there is a non-rigid spatial offset between different phases. Existing methods are difficult to achieve precise alignment at the feature level, resulting in blurred lesion boundaries.

[0004] (2) The difference in time phase leads to inconsistent feature distribution. Images from different phases have significant differences in gray-scale distribution and contrast relationship. Simple fusion can easily introduce distribution shift and affect the stability of segmentation.

[0005] (3) Lack of adaptive fusion and reliability control. Existing methods are difficult to dynamically adjust the contribution weight of each phase according to spatial location, and also lack a filtering mechanism for low-quality interphase information, which is easy to introduce noise interference.

[0006] Therefore, there is an urgent need for a liver tumor segmentation method that can achieve adaptive alignment, dynamic feature selection, and reliable fusion at the feature level. Summary of the Invention

[0007] In view of this, the purpose of this invention is to propose a liver tumor segmentation method based on expert attention and deformable interphase fusion, so as to solve the problem of liver tumor segmentation that achieves adaptive alignment, dynamic feature selection and reliable fusion at the feature level.

[0008] To achieve the above objectives, this invention provides a liver tumor segmentation method based on multi-phase feature fusion, comprising the following steps: Acquire multi-phase CT image data; Independent feature extraction is performed on each phase image in the multi-phase CT image data to obtain the corresponding single-phase features; Adaptive enhancement processing is performed on the single-phase feature of each phase to obtain the enhanced single-phase feature; The enhanced features of the target phase are used as the main branch, and the enhanced features of the auxiliary phase are used as the auxiliary branches. The auxiliary branches are spatially aligned by deformable sampling, and the cross-phase feature interaction between the aligned auxiliary branch features and the main branch is performed to obtain cross-phase enhanced features. Based on the structural information of the main branch and the feature similarity in the inter-period feature interaction process, a fusion reliability weight is generated, and the inter-period enhancement feature is weighted and fused with the main branch using the fusion reliability weight to obtain the fused feature; The fusion features are decoded to generate segmentation results for the liver and tumor.

[0009] Optionally, the adaptive enhancement processing for each single-phase feature specifically includes: Flatten the single-phase characteristics into a sequence form; The outputs of multiple parallel expert branches are weighted and fused based on gated routing weights, which are calculated using the following formula: ; in E represents the number of experts. W represents the noise term, used to enhance model robustness. r and W n A learnable weight matrix used to incorporate the features of the input sequence. Mapped to the routing space; Softplus is the activation function, defined as This is used to ensure that the output is positive.

[0010] Optionally, the spatial alignment of the auxiliary branch through deformable sampling includes enhancing features of the main branch. Enhancement features with the auxiliary branch The data is stitched together, and the intertemporal spatial migration field Δ is estimated through convolution operations: ; Enhancement features of the auxiliary branch based on the intertemporal spatial migration field Δ Deformable sampling is performed to obtain spatially aligned auxiliary features. : .

[0011] Optionally, the cross-period feature interaction includes: using the enhanced features of the main branch as query Q, and using the spatially aligned auxiliary features as key K and value V; Calculate the cosine similarity between the query Q and the key K, as the feature similarity S: ; The intertemporal enhancement feature F is generated by weighting the value V using feature similarity S. C : .

[0012] Optionally, the generation of fusion reliability weights includes: enhancement features for the main branch. After averaging along the channel dimension, the gradient information is extracted and used as the structural information G: ; Among them, the Mean C (•) indicates that the mean is calculated along the channel dimension; The structural information G is concatenated with the feature similarity S, and the fusion reliability weight α is generated through convolution and nonlinear mapping: ; Where σ is the Sigmoid function, and GELUGELU is the activation function.

[0013] Optionally, after acquiring the multi-phase CT image data, a preprocessing step is further included: spatially resampling the multi-phase CT image data to unify the voxel spacing to a preset resolution s. ; Where I represents the original image and X represents the resampled image; Intensity normalization is performed on the resampled image: ; Where μ and σ are the mean and standard deviation of the image, respectively. The image after normalization; Spatial alignment processing is performed on images from different phases.

[0014] Optionally, when extracting independent features from each phase image in the multi-phase CT image data, a hierarchical coding network is used to extract multi-scale feature representations: ; Where L represents the number of feature levels, with different levels corresponding to different receptive fields and semantic granularities.

[0015] Optionally, the step of using fusion reliability weights to weight and fuse the inter-period enhancement features with the main branch specifically involves: ; Where F fused As a fusion feature, the Enhancement features of the main branch, F C For cross-period enhancement features, α is the fusion reliability weight.

[0016] Optionally, the decoding of the fused features includes: employing a layer-by-layer upsampling structure, fusing encoder features from corresponding layers during the upsampling process to gradually restore spatial resolution. ; in This represents the features after upsampling at layer l, where Up(•) denotes the upsampling operation. These are the encoder features for the corresponding level; The segmentation probability map is obtained through the output mapping layer: ; Where P is the segmentation probability map, F out For the final high-resolution feature, σ is the Sigmoid function.

[0017] Optionally, the multi-phase CT image data includes at least two phases from the plain scan phase, arterial phase, venous phase, and delayed phase.

[0018] The beneficial effects of this invention are: (1) In view of the problem of spatial misalignment and difficulty in precise alignment in the prior art, the present invention uses deformable sampling to spatially align the auxiliary phase features, and achieves spatial consistency of cross-phase features without relying on precise registration, effectively reducing the impact of non-rigid offset caused by respiratory motion and organ deformation on segmentation accuracy.

[0019] (2) In view of the problems of inconsistent feature distribution caused by temporal differences and distribution shift introduced by simple fusion in the prior art, the present invention performs adaptive enhancement processing on the phase features of each phase. Through the weighted fusion mechanism based on gated routing and multiple expert branches, the key regional information related to tumor is highlighted, background redundancy is suppressed, and the robustness of the model to distribution differences is enhanced.

[0020] (3) In view of the problems of lack of adaptive fusion and reliability control and easy introduction of noise interference in the existing technology, the present invention generates fusion reliability weight based on the structural information of the main branch and the similarity of cross-period features. The weight is used to perform adaptive weighted fusion of cross-period enhancement features, enhance information fusion in regions with high cross-period consistency, suppress interference in inconsistent or noisy regions, and improve the stability and accuracy of segmentation results. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this 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 for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process of the liver tumor segmentation method based on expert attention and deformable interphase fusion according to an embodiment of the present invention; Figure 2 This is a detailed flowchart of the liver tumor segmentation method based on expert attention and deformable interphase fusion according to an embodiment of the present invention; Figure 3 This is a network model architecture diagram of the liver tumor segmentation method based on expert attention and deformable interphase fusion according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the multi-period expert attention module in an embodiment of the present invention; Figure 5 This is a schematic diagram of a deformable interphase fusion block according to an embodiment of the present invention; Figure 6 This is a deformable interphase fusion method according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0024] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] like Figures 1 to 6 As shown in the figure, a specific embodiment of the present invention provides a liver tumor segmentation method based on expert attention and deformable interphase fusion. The overall process includes the following steps: Step 1: Acquisition and preprocessing of multi-phase CT images; Step 2: Feature extraction from multiple image phases; Step 3: Adaptive enhancement of single-period features; Step 4: Alignment and fusion of cross-period features with deformable features; Step 5: Integrate reliability control; Step 6: Generation of segmentation results.

[0026] Step 1, multi-phase CT image acquisition and preprocessing, includes: CT image data of the same patient at different enhancement phases, including plain scan, arterial phase, venous phase, and delayed phase, were acquired and organized into multi-channel input data. To eliminate the influence of different scanning conditions, the original images were preprocessed.

[0027] First, spatial resampling is performed on CT images of each phase to unify the voxel spacing to a preset resolution, ensuring spatial consistency between different phases. This process can be represented as: ; Where I (P) Let represent the original image of the Pth phase, and s be the target spatial resolution.

[0028] Secondly, the images are intensity normalized to reduce grayscale differences caused by different scanning devices and imaging parameters, thereby improving the stability of model training. The normalization process is as follows: ; Where μ and σ are the mean and standard deviation of the image, respectively.

[0029] Furthermore, spatial alignment processing is performed on images of different phases to reduce displacement errors caused by respiratory movements or organ deformation, thereby enabling different phases to have spatial correspondence.

[0030] Finally, the preprocessed images from multiple periods are uniformly cropped or sampled to obtain fixed-size 3D input data blocks, which are then used as model input. ; Through the above processing, spatially consistent and intensity-standardized multi-phase CT image data are obtained, providing a reliable input basis for subsequent feature extraction and cross-phase fusion.

[0031] Step 2: Multi-phase image feature extraction includes: The multi-phase CT image data obtained in step 1 are input into the feature extraction network, and each phase image is independently encoded to obtain the corresponding multi-scale feature representation. This step uses a deep neural network to learn the representation of the original images, mapping low-level pixel information into high-dimensional semantic features, providing a foundation for subsequent fusion.

[0032] Specifically, for the input data of the p-th phase Feature extraction is performed through an encoding network: ; in This represents the three-dimensional feature map of the corresponding phase, where B is the batch size, C is the number of channels, and D, H, and W are the spatial dimensions, respectively.

[0033] Furthermore, to enhance the model's ability to perceive structures at different scales, the encoding network employs a hierarchical structure to extract multi-scale features: ; Where L represents the number of feature levels, and different levels correspond to different receptive fields and semantic granularities.

[0034] By extracting features independently from images of each phase, the differences in development features under different enhancement phases can be fully preserved, providing multi-source information support for subsequent single-phase feature enhancement and cross-phase feature fusion.

[0035] Step 3: Single-period feature adaptive enhancement includes: Adaptive enhancement processing is applied to the phase features obtained in step 2 to improve the expressive power of key regions and suppress redundant information. This step introduces a feature enhancement method based on an expert mechanism, which dynamically routes and selects different feature transformation paths to achieve spatially relevant feature optimization.

[0036] Specifically, for the characteristic F of the p-th phase... (p) First, normalization is performed to stabilize the feature distribution: ; Subsequently, the 3D features are flattened into a sequence for routing calculation: ; Based on this sequence feature, a gated routing weight is constructed to measure the degree to which different experts adapt to the current feature: ; in E represents the number of experts. W represents the noise term, used to enhance model robustness. r and W n A learnable weight matrix used to incorporate the features of the input sequence. Mapped to the routing space; Softplus is the activation function, defined as This is used to ensure that the output is positive.

[0037] For each expert branch, perform a different nonlinear transformation on the input features: ; Weighted fusion of all expert outputs: ; Finally, the enhanced feature representation is obtained through residual connections: ; Through the above process, the model can adaptively select the most appropriate feature enhancement method according to the feature distribution of different spatial locations, thereby highlighting tumor-related regional information and suppressing redundant background features, providing a more discriminative feature basis for subsequent interphase fusion.

[0038] Step 4: Trans-period feature deformable alignment and fusion includes: The enhanced features of each phase obtained in step 3 are subjected to cross-phase fusion processing to fully utilize the complementary information between different phases. This step takes the target phase features as the main branch and the other phases as auxiliary branches, and achieves cross-phase information fusion through deformable alignment and feature interaction.

[0039] Specifically, let the characteristics of the target period be... The auxiliary phase characteristics are Both dimensions are: ; First, the intertemporal spatial offset is estimated through feature concatenation and convolution operations: ; Where Δ represents the offset field in three-dimensional space, used to describe the displacement relationship between the auxiliary phase and the target phase.

[0040] Subsequently, based on the offset field, deformable sampling of the auxiliary phase features is performed to achieve spatial alignment: ; This process mitigates the effects of non-rigid deformation by sampling the position offset of features to align different phases in the feature space.

[0041] After spatial alignment is completed, a cross-period feature interaction mechanism is constructed. First, query, key, and value representations are generated: ; Further calculations were performed to determine the characteristic similarity between the target phase and the auxiliary phase: ; This similarity is then used to weight auxiliary features to obtain cross-period enhancement features: ; Through the above steps, effective interaction of features from different phases is achieved based on spatial alignment, enabling the model to fully utilize complementary information from multiple phase images and improve its ability to represent tumor regions and boundary structures.

[0042] Step 5: Fusion reliability control includes: After completing the cross-period feature fusion in step 4, in order to avoid mismatched or low-quality cross-period information from interfering with the results, a fusion reliability control mechanism is introduced to adaptively weight the fusion results at different spatial locations.

[0043] Specifically, firstly, based on the phase characteristics of the target period Structural information is extracted, and by aggregating the channel dimensions and calculating their spatial changes, gradient information reflecting boundary and structural features is obtained. ; in It is used to characterize the structural salience at different spatial locations.

[0044] Simultaneously, using the inter-period feature similarity S obtained in step 4, the input features for the fusion weights are jointly constructed: ; Based on this, fused weights are generated through convolution and nonlinear mapping: ; Where α∈(0,1) represents the fusion reliability weight of each spatial location.

[0045] Finally, the inter-period enhancement features are adjusted according to the weights and fused with the target period features: ; Through the above mechanism, the model can adaptively evaluate the credibility of different regions during the fusion process, enhance information fusion in locations with clear structures and high consistency across periods, and suppress cross-period interference in inconsistent or noisy regions, thereby improving the stability and accuracy of the overall segmentation results.

[0046] Step 6: Generating the segmentation result includes: The fusion feature F obtained in step 5 fused The input decoder network progressively restores spatial resolution and generates segmentation results for the liver and tumor.

[0047] Specifically, the decoder employs a layer-by-layer upsampling structure to restore the spatial size of the fused features, while simultaneously combining shallow features from the encoding stage to enhance the representation of detailed information. This process can be represented as: ; in This represents the features after upsampling at layer l, where Up(•) denotes the upsampling operation. These are the encoder features for the corresponding level.

[0048] After the multi-layer features are gradually fused, the final high-resolution feature representation is obtained: ; Furthermore, the features are mapped to segmentation probability maps through an output mapping layer: ; in K represents the number of categories.

[0049] Finally, the final segmentation result is obtained through threshold segmentation or maximum probability selection: ; Through the above decoding and reconstruction process, accurate segmentation of the liver and tumor regions is achieved, resulting in spatially continuous and clearly defined segmentation results.

[0050] The working process of this invention: This invention first acquires multi-phase enhanced CT images of the same patient, and preprocesses the plain, arterial, venous, and delayed phase images through resampling, normalization, and spatial alignment to form unified multi-phase input data. Subsequently, the images of each phase are input into a feature extraction network to obtain corresponding multi-scale feature representations, thereby preserving the imaging differences of the tumor and surrounding tissues in different phases.

[0051] In the feature modeling stage, adaptive enhancement is first applied to the features of each phase. Specifically, an attention module based on an expert mechanism dynamically selects appropriate feature transformation paths for features at different spatial locations, enhancing key tumor-related regional information and suppressing redundant background interference. Then, with the target phase features as the main branch and the other phase features as auxiliary branches, cross-phase spatial alignment is performed using deformable sampling methods, and cross-phase information interaction is achieved by combining feature similarity, enabling complementary information from different phases to be fused under a unified spatial semantic framework.

[0052] During the fusion process, reliability weights are further generated based on the structural information of the target phase and the consistency of cross-phase features. These weights are then adaptively adjusted to reduce the impact of mismatched features or noise. Finally, the fused features are input into the decoder network to restore spatial resolution layer by layer and output the segmentation results for the liver and tumor regions. These steps enable the effective utilization of multi-phase imaging information, improving the accuracy and robustness of liver tumor segmentation.

[0053] Optional parameter settings: In the specific implementation of this invention, the following parameters can be flexibly set and adjusted according to the actual data scale, computing resources, and task requirements: (1) Input data related parameters include: resampling resolution s: for example, 1.0mm~2.0mm; input block size (D,H,W): for example, 64×128×128 or 96×96×96; number of phases P: can be 2, 3 or 4 phases (plain scan, artery, vein, delayed scan).

[0054] (2) Feature extraction network parameters include: number of encoder channels C: for example 48, 96, 192, 384; number of feature layers L: for example 4~5 layers; network type: optional convolutional network or Transformer-based coding structure.

[0055] (3) The parameters of the expert attention module include: number of experts E: for example, 4~8; number of top-k selections: for example, k=1~2; noise coefficient Used to control the randomness of routing, with values ​​ranging from 0.1 to 1.0.

[0056] (4) Deformable Fusion Module Parameters (DCPF) include: number of reference sampling points n ref For example, 4; Offset field convolution kernel size: for example, 3×3×3; Whether to use multi-scale alignment: optional to be on or off.

[0057] (5) Fusion control parameters include: similarity calculation method: cosine similarity or dot product similarity; weight function σ(•): optional Sigmoid function; gradient extraction convolution kernel size: for example 3×3×3.

[0058] (6) Training and optimization parameters include: optimizer: Adam or AdamW; learning rate: for example, 1×10 -4 Batch size B: e.g., 1~4; Loss function: Dice loss, cross-entropy loss or a combination thereof.

[0059] (7) Output setting parameters include: number of categories K: for example, 2 categories (liver, tumor) or multiple categories; threshold setting: for example, 0.5 for binarization of segmentation results.

[0060] The above parameters can be adjusted according to different dataset sizes, hardware conditions, and application scenarios without affecting the core technical solution and implementation method of this invention.

[0061] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0062] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A liver tumor segmentation method based on expert attention and deformable interphase fusion, characterized in that, Includes the following steps: Acquire multi-phase CT image data; Independent feature extraction is performed on each phase image in the multi-phase CT image data to obtain the corresponding single-phase features; Adaptive enhancement processing is performed on each single-phase feature, including: flattening the single-phase feature into a sequence form, weighting and fusing the outputs of multiple expert branches based on gated routing weights, and adding the single-phase feature through residual connection to obtain the enhanced single-phase feature; The enhanced features of the target phase are used as the main branch, and the enhanced features of the auxiliary phase are used as the auxiliary branch. The auxiliary branch is spatially aligned by deformable sampling, and the features of the aligned auxiliary branch are interacted with the main branch across phases. This includes: generating a query with the enhanced features of the main branch, generating a key and a value with the features of the spatially aligned auxiliary branch, calculating the feature similarity between the query and the key, and using the feature similarity to weight the value to obtain the cross-phase enhanced features. Based on the structural information of the main branch and the feature similarity in the interaction process of cross-period features, a fusion reliability weight is generated, and the cross-period enhancement features are weighted and fused with the main branch using the fusion reliability weight to obtain fused features; the generation of fusion reliability weight includes: aggregating the enhancement features of the main branch in the channel dimension and extracting spatial change information to obtain structural information, and generating the fusion reliability weight based on the structural information and the feature similarity. The fusion features are decoded to generate segmentation results for the liver and tumor.

2. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, The outputs of multiple expert branches are weighted and fused based on gated routing weights, which are calculated using the following formula: ； in , For the number of experts, This represents the noise term, used to enhance the robustness of the model. and A learnable weight matrix used to incorporate the features of the input sequence. Mapped to the routing space; The activation function is defined as follows: This is used to ensure that the output is positive.

3. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, The spatial alignment of the auxiliary branches through deformable sampling includes enhancing features of the main branch. Enhancement features with the auxiliary branch The data is stitched together, and the intertemporal spatial migration field is estimated through convolution operations. : ； Based on the intertemporal spatial migration field Enhancement features for auxiliary branches Deformable sampling is performed to obtain spatially aligned auxiliary features. : 。 4. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, The cross-period feature interaction includes: using the enhanced features of the main branch as the query... Using spatially aligned auxiliary features as keys Sum ; Calculate the query AND key Cosine similarity between them, as a feature similarity : ； Utilizing the aforementioned feature similarity For the value Weighting is performed to generate the intertemporal enhancement features. : 。 5. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, The generated fusion reliability weights include: enhanced features for the main branch. After averaging along the channel dimension, the gradient information is extracted as structural information. : ； Among them, the This indicates that the mean is calculated along the channel dimension; structural information Similarity to the aforementioned features The data is concatenated, and the fusion reliability weights are generated through convolution and nonlinear mapping. : ； in, This is the Sigmoid function.

6. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, After acquiring multi-phase CT image data, a preprocessing step is also included: spatially resampling the multi-phase CT image data to unify the voxel spacing to a preset resolution s. ； in, Represents the original image. Resampled image; Intensity normalization is performed on the resampled image: ； in and These are the mean and standard deviation of the image, respectively. The image after normalization; Spatial alignment processing is performed on images from different phases.

7. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, When performing independent feature extraction on each phase image in the multi-phase CT image data, a hierarchical coding network is used to extract multi-scale feature representations: ； Where L represents the number of feature levels, with different levels corresponding to different receptive fields and semantic granularities.

8. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, The inter-period enhancement features are weighted and fused with the main branch using the fusion reliability weight, specifically as follows: ； in As a feature of fusion, Enhancement features of the main branch, For intertemporal enhancement features, To incorporate reliability weights.

9. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, Decoding the fused features includes: employing a layer-by-layer upsampling structure, fusing encoder features from corresponding layers during the upsampling process to gradually restore spatial resolution. ； in Indicates the first Features after layer upsampling Indicates an upsampling operation. These are the encoder features for the corresponding level; The segmentation probability map is obtained through the output mapping layer: ； in, For segmentation probability graphs, For the final high-resolution features, This is the Sigmoid function.

10. The liver tumor segmentation method based on expert attention and deformable interphase fusion according to claim 1, characterized in that, The multi-phase CT image data includes at least two phases from the plain scan phase, arterial phase, venous phase, and delayed phase.

Images