A cross-period ct liver tumor segmentation method and device based on redundant perception

Abstract

The application discloses a kind of cross-period CT liver tumor segmentation method and device based on redundancy perception, which is aimed at realizing the accurate segmentation of liver tumor.The method comprises the following steps: S1, pre-processing and data set division are carried out on abdominal CT data of arterial phase and portal phase, local entropy guided response calibration is carried out on the pre-processed data, and a local entropy guided response calibration module LERC is constructed;S2, fusion partial optimal transmission theory and attention mechanism, construct selective phase fusion module SPF;S3, fusion feature decoupling and multi-branch convolution, construct progressive refinement decoding module PRD;S4, fusion local entropy guided response calibration module, selective phase fusion module and progressive refinement decoding module, construct RAPT-Net model;S5, the training, verification optimization and performance evaluation of the model are completed using experimental data set.The application can suppress redundant interference and utilize phase complementary information, effectively improve the segmentation accuracy of liver tumor, and is suitable for automatic segmentation of liver tumor in cross-period CT.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically to a method and apparatus for segmenting liver tumors across time-phase CT based on redundancy sensing. Background Technology

[0002] In recent years, medical image segmentation has played a crucial role in the clinical diagnosis, treatment planning, and efficacy evaluation of liver tumors. Computed tomography (CT) is widely used in liver tumor-related examinations due to its high spatial resolution and excellent imaging capabilities for liver lesions. For liver tumor segmentation, accurately obtaining the location, extent, and boundary information of the tumor region helps clinicians in lesion assessment, treatment planning, and postoperative follow-up. However, manual layer-by-layer delineation is not only time-consuming and labor-intensive but also easily influenced by the patient's experience in interpreting images, making it difficult to meet the demands of high efficiency and consistency in clinical applications.

[0003] Most existing automatic segmentation methods for liver tumors are based on deep learning networks, especially convolutional neural network-based segmentation models, which have achieved good results in this task. Most existing methods primarily segment single-phase CT images, while some studies have begun to focus on segmenting liver tumors across different CT phases. These methods aim to obtain richer lesion discrimination clues by jointly utilizing enhancement differences between arterial and portal venous phase images. However, liver tumors often exhibit characteristics such as small volume, low contrast, and blurred boundaries in different cases and regions, making it difficult for single-phase images to consistently present a complete lesion morphology. Furthermore, while existing cross-phase segmentation methods emphasize the utilization of complementary information between different phases, they do not pay sufficient attention to redundant responses in cross-phase features, which can easily lead to the propagation of low-value shared information during the fusion process, thereby weakening tumor-related discrimination clues.

[0004] Furthermore, in the decoding and feature recovery stages, existing methods typically focus on directly recovering fused features, lacking targeted refinement mechanisms for missegmentation and missed segmentation caused by redundant responses and weak boundaries. This results in the model still being prone to problems such as boundary overflow, incomplete contours, or missed lesions in scenarios with blurred boundaries and small lesions.

[0005] To address the aforementioned issues, it is necessary to explore a novel cross-phase CT liver tumor segmentation method that can simultaneously suppress intra-phase and inter-phase redundancy, and further enhance boundary restoration and small lesion characterization capabilities. Therefore, effectively suppressing low-information-content redundant responses within a single phase during cross-phase CT liver tumor segmentation, reducing the propagation of redundant information between different phases during the fusion process, and further improving the restoration capabilities for blurred boundaries and small lesion regions, thereby achieving accurate and automatic segmentation of liver tumor regions, has become a key technical problem urgently needing to be solved in the field of medical image segmentation. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and apparatus for segmenting liver tumors across time-phase CT based on redundancy sensing, which can provide users with a more accurate, efficient, and highly adaptable liver tumor segmentation solution.

[0007] Specifically, the method includes the following steps: S1: Preprocess and partition the abdominal CT data of the arterial and portal venous phases, perform local entropy-guided response calibration on the preprocessed data, and construct the Local Entropy-Guided Response Calibration (LERC) module. S2: Integrating partial optimal transport theory and attention mechanism, constructing a selective period phase fusion module SPF; S3: Integrate feature decoupling and multi-branch convolution to construct the progressively refined decoding module PRD; S4: Integrate the local entropy-guided response calibration module, the selective phase fusion module, and the progressive refinement decoding module to construct the RAPT-Net model; S5: Use the experimental dataset to complete the training, validation, optimization, and performance evaluation of the model.

[0008] Preferably, S1 includes the following steps: S1.1: Perform data format conversion, registration, region of interest cropping, grayscale truncation, normalization, and annotation processing on the abdominal CT data of the arterial and portal venous phases, and divide the processed data into training set, validation set, and test set; S1.2: Construct a local entropy distribution and generate response weights, use the response weights to calibrate the response of shallow features, and construct the Local Entropy Guided Response Calibration Module (LERC). Preferably, S1.2 includes the following steps: First, the superficial features of the arterial and portal venous phases are projected using a single channel, and the projection results are normalized using the Sigmoid function to obtain the response map. Then, a soft histogram is constructed at the centers of K preset bins, and the local probability distribution of the response map is estimated. Based on the local probability distribution, the normalized local entropy is calculated. Finally, the original superficial features are calibrated using the local entropy to obtain the calibrated superficial features, thus constructing the Local Entropy Guided Response Calibration (LERC) module. The specific calculation process is shown below: In the above formula, This represents the shallow features of the input. express Convolution projection operation, This represents the Sigmoid function. This represents the normalized response graph. Indicates the center of the k-th box. This represents the smoothing coefficient of the soft histogram. Represents a minimal constant. This indicates a local average pooling operation. This represents the local probability distribution corresponding to the center of the k-th box. This represents the normalized local entropy. This represents the shallow features after calibration using a local entropy-guided response. This indicates an element-wise multiplication operation. This represents the basic retention factor.

[0009] Preferably, S2 includes the following steps: S2.1: Generate saliency weights based on cross-attention and construct a cost matrix between phase features across different periods by combining cosine distance; S2.2: Using the cost matrix, combined with the Sinkhorn algorithm and entropy regularization constraints, a partially optimal transfer matrix is ​​constructed; S2.3: Utilize the partially optimal transfer matrix to perform cross-period phase feature fusion and construct a selective period phase fusion module (SPF); Preferably, S2.1 includes the following steps: The portal venous phase features and arterial phase features, after being calibrated by local entropy-guided response, are convolved and mapped, and corresponding token sequences are constructed using adaptive pooling. Subsequently, saliency weights for the two phase tokens are generated based on bidirectional cross-attention. Combining the cosine distance between tokens and a saliency penalty term, a cost matrix for cross-phase feature matching is constructed. The specific calculation process is shown below: In the above formula, and These represent the characteristics of the portal venous phase and the arterial phase, respectively. This indicates an adaptive pooling operation. This indicates a tokenization operation. and The token sequences representing the portal venous and arterial phases. This indicates the operation of calculating the mean. and This indicates the significance weight of the two phases. Indicates the matching cost. Indicates the significance penalty coefficient; Preferably, S2.2 includes the following steps: To achieve partial transmission, unmatched terms are introduced into the cost matrix to construct an extended cost matrix. Transmission quality for tokens in two phases is allocated based on significance weights, and residual transmission quality is reserved for unmatched terms, resulting in a marginal distribution. Then, a partially optimal transmission objective function is constructed using entropy regularization constraints, and the Sinkhorn algorithm is used to solve for the partially optimal transmission matrix. The specific calculation process is as follows: In the above formula, Represents the extended cost matrix. Indicates the cost of the unmatched item. and Let N and M represent vectors consisting entirely of 1s, respectively. Indicates the matching ratio. b and 'b' represent the marginal distribution of tokens in the portal venous phase and arterial phase, respectively. This represents the set of feasible transports that satisfy the edge distribution constraints. The entropy regularity coefficient is... Let be the transfer matrix to be solved. This is a partially optimal transfer matrix; Preferably, S2.3 includes the following steps: First, the effective transfer sub-matrices corresponding to the two phase tokens in the partial optimal transfer matrix are extracted. Then, the effective transfer sub-matrices are used to perform cross-phase feature aggregation on the two phase tokens respectively, and the aggregation results are restored to spatial feature maps, yielding cross-phase aggregated features. Finally, the cross-phase aggregated features are concatenated with the original features, and residual fusion is performed through convolution, batch normalization, and activation functions to construct the selective phase fusion module SPF. The specific calculation process is as follows: In the above formula, This is the effective transmission submatrix in the partially optimal transmission matrix. and These represent the inter-period aggregation tokens of the two phases, This indicates the feature map recovery operation. This indicates a batch normalization operation. This represents the activation function. and These represent the residual fusion coefficients of the two phases, respectively. and The portal venous phase and arterial phase characteristics after fusion. This is the encoder output feature.

[0010] Preferably, S3 includes the following steps: S3.1: Fuse the decoded features from the previous scale and the features from the current scale to obtain fused features, decouple the fused features, and construct foreground attention features and background attention features; S3.2: Use multi-branch convolution to explore the context of the decoupled features, and refine the fused features based on the context exploration results to construct a progressive refinement decoding module (PRD); Preferably, S3.1 includes the following steps: The prediction results from the previous scale are upsampled and mapped using the Sigmoid function to obtain a foreground probability map. The decoded features from the previous scale are convolved and upsampled to the same spatial size as the features at the current scale, then concatenated and fused with the encoder output features at the current scale to obtain fused features. Based on the foreground probability map, the fused features are decoupled to obtain foreground attention features and background attention features. The specific calculation process is as follows: In the above formula, This is the prediction result for the previous scale. Foreground probability plot, This represents the Sigmoid function. Indicates an upsampling operation. The encoder outputs features for the current scale. These are the decoding features of the previous scale. As a feature of fusion, As a feature of future focus, Background features to focus on This represents the element-wise multiplication operation; Preferably, S3.2 includes the following steps: Multi-branch convolution processing is performed on the foreground attention features and the background attention features respectively. Multi-scale contextual information is extracted through convolutional branches and pooling branches with different receptive fields, and the outputs of each branch are fused to obtain the corresponding context exploration results. When the foreground attention features are used as input, a suppression response for suppressing false activation regions is obtained; when the background attention features are used as input, a compensation response for enhancing missed detection regions is obtained. The specific calculation process is as follows: Among them, when hour, , , indicating a suppressed response; when hour, , , indicating a compensation response; , and These represent convolution operations for different receptive fields. This indicates a pooling operation. Indicates use Convolution performs channel mapping operations; Based on the suppression response, the false activation regions in the fused features are suppressed to obtain intermediate features; then, based on the compensation response, the missed detection regions in the intermediate features are enhanced to obtain refined features and prediction results at the current scale, thereby constructing the progressive refinement decoding module (PRD); the specific calculation process is as follows: In the above formula, The intermediate feature after suppression For the refined features, The prediction results are for the current scale. This represents batch normalization and ReLU activation function operations. and They represent the learnable coefficients, This represents the Sigmoid function. This represents the hyperbolic tangent function.

[0011] Preferably, S4 includes the following steps: Based on a dual-branch U-shaped network architecture, the aforementioned Local Entropy Guided Response Calibration (LERC) module, Selective Phase Fusion (SPF) module, and Progressive Refinement Decoding (PRD) module are integrated to construct the RAPT-Net model. First, a ResNet-34-based dual-branch parallel encoder is constructed to extract multi-level features from the arterial and portal venous phases, respectively. Then, the LERC module is embedded in the shallow encoding stage to suppress low-information-content redundant responses within phases and enhance structurally relevant responses. Next, the SPF module is embedded in the mid-to-high-level encoding stage to establish selective correspondences between phase features and perform cross-phase feature fusion. Finally, the PRD module is embedded in the decoding stage to progressively refine and restore the fused features, thereby outputting the final liver tumor segmentation result.

[0012] Preferably, S5 includes the following steps: The predefined training set is input into the RAPT-Net model for training. Forward propagation is performed based on the arterial phase image, portal venous phase image and their corresponding annotations of the training samples to obtain liver tumor segmentation results. The model parameters are updated according to the difference between the segmentation results and the true annotations. The model performance during the training process is validated and optimized using the validation set, and the optimal model weights are saved. Finally, the performance of the trained model is evaluated using the test set.

[0013] The beneficial effects of the method of this invention are as follows: First, in the data preprocessing stage, by registering, cropping regions of interest, truncating grayscale, normalizing, and dividing the abdominal CT data of the arterial and portal venous phases, a high-quality data foundation with relatively consistent correspondences is provided for subsequent cross-phase feature modeling. During model construction, by constructing the Local Entropy Guided Response Calibration (LERC) module, adaptive calibration of feature responses in a single phase is performed using local information distribution. This effectively suppresses low-information redundant responses generated in non-lesion areas while enhancing high-information responses related to tumor boundaries, small lesions, and local structural changes, thereby improving the discriminative power of single-phase feature representation. Next, by constructing the Selective Phase Fusion (SPF) module, combining cross-attention and partial optimal transport theory, in... By establishing selective correspondences between different phases, the low-value shared information propagation caused by forced full fusion is avoided. While fully utilizing the complementary information of the arterial and portal venous phases, cross-phase redundant interference is effectively reduced. Subsequently, by constructing a progressively refined decoding module (PRD) and employing feature decoupling and multi-branch context exploration mechanisms, falsely activated regions are suppressed, and missed detection regions are compensated and enhanced, thereby improving the recovery capability for ambiguous boundaries and small lesion regions. Overall, the constructed RAPT-Net model synergistically optimizes the cross-phase CT liver tumor segmentation process from three levels: intra-phase redundancy suppression, inter-phase selective fusion, and progressive refinement in the decoding stage. This significantly improves the accuracy and boundary integrity of liver tumor segmentation, providing effective technical support for clinical auxiliary diagnosis and treatment planning. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following is a brief introduction to the drawings used in the prior art and embodiments. The following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 is a flowchart illustrating a cross-phase CT liver tumor segmentation method based on redundancy sensing according to the present invention. Figure 2 is a diagram of the RAPT-Net model architecture of a cross-phase CT liver tumor segmentation method based on redundancy perception according to the present invention. Figure 3 is a schematic diagram of the LERC structure of a cross-phase CT liver tumor segmentation method based on redundancy sensing according to the present invention. Figure 4 This is a schematic diagram of the SPF structure of a cross-phase CT liver tumor segmentation method based on redundancy sensing according to the present invention. Figure 5 This is a schematic diagram of the PRD structure of a cross-phase CT liver tumor segmentation method based on redundancy sensing according to the present invention. Detailed Implementation

[0016] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0017] This application provides a method and apparatus for segmenting liver tumors across time-series CT based on redundancy sensing, which can achieve accurate segmentation of liver tumors, assist in clinical diagnosis and treatment decisions, and thereby improve the accuracy and efficiency of medical image processing.

[0018] Reference Figure 1 The method includes the following steps: S1: Preprocess and partition the abdominal CT data of the arterial and portal venous phases, perform local entropy-guided response calibration on the preprocessed data, and construct the Local Entropy-Guided Response Calibration (LERC) module. S2: Integrating partial optimal transport theory and attention mechanism, constructing a selective period phase fusion module SPF; S3: Integrate feature decoupling and multi-branch convolution to construct the progressively refined decoding module PRD; S4: Integrate the local entropy-guided response calibration module, the selective phase fusion module, and the progressive refinement decoding module to construct the RAPT-Net model; S5: Use the experimental dataset to complete the training, validation, optimization, and performance evaluation of the model.

[0019] Furthermore, S1 includes the following steps: S1.1: Perform data format conversion, registration, region of interest cropping, grayscale truncation, normalization, and annotation processing on the abdominal CT data of the arterial and portal venous phases, and divide the processed data into training set, validation set, and test set; S1.2: Construct a local entropy distribution and generate response weights, use the response weights to calibrate the response of shallow features, and construct the Local Entropy Guided Response Calibration Module (LERC). Furthermore, S1.1 includes the following steps: The acquired abdominal CT data from the arterial and portal venous phases were converted into a numerical matrix format that is readable and processable by the network. Then, using the portal venous phase image as a reference, spatial registration was performed on the arterial and portal venous phase images of the same case to correct spatial misalignment caused by respiration and body movement. Subsequently, the regions of interest (ROIs) were cropped based on the liver region to remove irrelevant background interference, and the cropped images were uniformly adjusted to 512×512 pixels. Next, the images underwent grayscale truncation and normalization, with CT grayscale values ​​truncated to the [-60, 200] HU range and the truncated pixel values ​​normalized to the [0, 1] range. Then, the liver tumor region was labeled, constructing binary segmentation labels corresponding to the input image, where pixels in the liver tumor region were labeled as 1, and pixels in non-tumor regions were labeled as 0. Finally, the preprocessed data was divided into training, validation, and test sets in an 8:1:1 ratio. Furthermore, refer to Figure 3 S1.2 includes the following steps: First, the superficial features of the arterial and portal venous phases are projected using a single channel, and the projection results are normalized using the Sigmoid function to obtain a response map. Then, a soft histogram is constructed on K preset bin centers (8 bin centers in total). The local probability distribution of the response map is estimated, and the normalized local entropy is calculated based on this distribution. Finally, the local entropy is used to calibrate the original superficial features, resulting in calibrated superficial features. This forms the basis of the Local Entropy Guided Response Calibration (LERC) module. The specific calculation process is shown below: In the above formula, This represents the shallow features of the input. express Convolution projection operation, This represents the Sigmoid function. This represents the normalized response graph. Indicates the center of the k-th box. This represents the smoothing coefficient of the soft histogram, set to 0.05. Represents a minimal constant. This indicates a local average pooling operation, with the pooling window size set to 5×5. This represents the local probability distribution corresponding to the center of the k-th box. This represents the normalized local entropy. This represents the shallow features after calibration using a local entropy-guided response. This indicates an element-wise multiplication operation. This represents the base retention factor, set to 0.5.

[0020] Furthermore, refer to Figure 4 S2 includes the following steps: S2.1: Generate saliency weights based on cross-attention and construct a cost matrix between phase features across different periods by combining cosine distance; S2.2: Using the cost matrix, combined with the Sinkhorn algorithm and entropy regularization constraints, a partially optimal transfer matrix is ​​constructed; S2.3: Utilize the partially optimal transfer matrix to perform cross-period phase feature fusion and construct a selective period phase fusion module (SPF); Furthermore, S2.1 includes the following steps: The portal venous phase features and arterial phase features, after calibration by local entropy-guided response, are subjected to 1×1 convolution mapping, and corresponding token sequences are constructed using adaptive pooling. Subsequently, saliency weights for the two-phase tokens are generated based on bidirectional cross-attention, where the cross-attention employs a 4-head attention mechanism. Finally, a cost matrix for cross-phase feature matching is constructed by combining the cosine distance between tokens and a saliency penalty term. The specific calculation process is shown below: In the above formula, and These represent the characteristics of the portal venous phase and the arterial phase, respectively. This indicates an adaptive pooling operation. This indicates a tokenization operation. and The token sequences representing the portal venous and arterial phases. This indicates the operation of calculating the mean. and This indicates the significance weight of the two phases. Indicates the matching cost. Indicates the significance penalty coefficient; Furthermore, S2.2 includes the following steps: To achieve partial transmission, unmatched terms are introduced into the cost matrix to construct an extended cost matrix. Transmission quality for tokens in two phases is allocated based on significance weights, and residual transmission quality is reserved for unmatched terms, resulting in a marginal distribution. Then, a partially optimal transmission objective function is constructed using entropy regularization constraints, and the Sinkhorn algorithm is used to solve for the partially optimal transmission matrix. The number of Sinkhorn iterations was set to 18, 15, 12, and 8 at the four scales, respectively. The specific calculation process is shown below: In the above formula, Represents the extended cost matrix. This represents the cost of the unmatched item, set to 0.10, 0.15, 0.20, and 0.25 on the four scales, respectively. and Let N and M represent vectors consisting entirely of 1s, respectively. This represents the matching ratio, with initial values ​​set to 0.8, 0.7, 0.6, and 0.5 for the four scales, respectively. b and 'b' represent the marginal distribution of tokens in the portal venous phase and arterial phase, respectively. This represents the set of feasible transports that satisfy the edge distribution constraints. The entropy regularization coefficient is set to 0.07. Let be the transfer matrix to be solved. This is a partially optimal transfer matrix; Furthermore, S2.3 includes the following steps: First, the effective transfer sub-matrices corresponding to the two phase tokens in the partial optimal transfer matrix are extracted. Then, the effective transfer sub-matrices are used to perform cross-phase feature aggregation on the two phase tokens respectively, and the aggregation results are restored to spatial feature maps, yielding cross-phase aggregated features. Finally, the cross-phase aggregated features are concatenated with the original features, and residual fusion is performed through 1×1 convolution, batch normalization, and activation functions to construct the selective phase fusion module SPF. The specific calculation process is as follows: In the above formula, This is the effective transmission submatrix in the partially optimal transmission matrix. and These represent the inter-period aggregation tokens of the two phases, This indicates the feature map recovery operation. This indicates a batch normalization operation. This represents the activation function. and These represent the residual fusion coefficients for the two phases, with an initial value set to 0.0. and The portal venous phase and arterial phase characteristics after fusion. This is the encoder output feature.

[0021] Furthermore, refer to Figure 5 S3 includes the following steps: S3.1: Fuse the decoded features from the previous scale and the features from the current scale to obtain fused features, decouple the fused features, and construct foreground attention features and background attention features; S3.2: Use multi-branch convolution to explore the context of the decoupled features, and refine the fused features based on the context exploration results to construct a progressive refinement decoding module (PRD); Furthermore, S3.1 includes the following steps: The prediction results from the previous scale are upsampled and mapped using the Sigmoid function to obtain a foreground probability map. The decoded features from the previous scale are convolved and upsampled to the same spatial size as the features at the current scale, then concatenated and fused with the encoder output features at the current scale to obtain fused features. Based on the foreground probability map, the fused features are decoupled to obtain foreground attention features and background attention features. The specific calculation process is as follows: In the above formula, This is the prediction result for the previous scale. Foreground probability plot, This represents the Sigmoid function. Indicates an upsampling operation. The encoder outputs features for the current scale. These are the decoding features of the previous scale. For the current scale fusion features, As a feature of future focus, Background features to focus on This represents the element-wise multiplication operation; Furthermore, S3.2 includes the following steps: Multi-branch convolution processing is performed on the foreground and background attention features respectively. Multi-scale contextual information is extracted through convolutional branches with different receptive fields and pooling branches. The outputs of each branch are then fused to obtain the corresponding context exploration results. The multi-branch convolution includes 3×3, 5×5, and 7×7 convolutional branches, as well as pooling branches. The number of output channels for each branch is set to one-quarter of the number of input channels, and channel compression and mapping are performed through 1×1 convolution. When the foreground attention features are used as input, a suppression response for suppressing false activation regions is obtained; when the background attention features are used as input, a compensation response for enhancing missed detection regions is obtained. The specific calculation process is as follows: Among them, when hour, , , indicating a suppressed response; when hour, , , indicating a compensation response; , and These represent convolution operations for different receptive fields. This indicates a pooling operation. Indicates use Convolution performs channel mapping operations; Based on the suppression response, the false activation regions in the fused features are suppressed to obtain intermediate features; then, based on the compensation response, the missed detection regions in the intermediate features are enhanced to obtain refined features and prediction results at the current scale, thereby constructing the progressive refinement decoding module (PRD); the specific calculation process is as follows: In the above formula, The intermediate feature after suppression For the refined features, The prediction results are for the current scale. This represents batch normalization and ReLU activation function operations. and These represent the learnable coefficients, both initialized to 1. This represents the Sigmoid function. This represents the hyperbolic tangent function.

[0022] Furthermore, refer to Figure 2 S4 includes the following steps: Based on a dual-branch U-shaped network architecture, the aforementioned Local Entropy Guided Response Calibration (LERC) module, Selective Phase Fusion (SPF) module, and Progressive Refinement Decoding (PRD) module are integrated to construct the RAPT-Net model. First, a ResNet-34-based dual-branch parallel encoder is constructed to extract multi-level features from the arterial and portal venous phases, respectively. Then, the LERC module is embedded in the shallow encoding stage to suppress low-information-content redundant responses within phases and enhance structurally relevant responses. Next, the SPF module is embedded in the mid-to-high-level encoding stage to establish selective correspondences between phase features and perform cross-phase feature fusion. Finally, the PRD module is embedded in the decoding stage to progressively refine and restore the fused features, thereby outputting the final liver tumor segmentation result.

[0023] Furthermore, S5 includes the following steps: The predefined training set is input into the RAPT-Net model for training. Forward propagation is performed based on the arterial phase image, portal venous phase image and their corresponding annotations of the training samples to obtain liver tumor segmentation results. The model parameters are updated according to the difference between the segmentation results and the true annotations. The model performance during the training process is validated and optimized using the validation set, and the optimal model weights are saved. Finally, the performance of the trained model is evaluated using the test set.

[0024] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make other equivalent modifications or substitutions without departing from the spirit of the invention. These equivalent modifications or substitutions are included within the scope defined by the claims.

Claims

1. A cross-phase CT liver tumor segmentation method based on redundancy sensing, characterized in that, The method includes the following steps: S1: Preprocess and partition the abdominal CT data of the arterial and portal venous phases, perform local entropy-guided response calibration on the preprocessed data, and construct the Local Entropy-Guided Response Calibration (LERC) module. S1 includes the following steps: S1.1: Perform data format conversion, registration, region of interest cropping, grayscale truncation, normalization, and annotation processing on the abdominal CT data of the arterial and portal venous phases, and divide the processed data into training set, validation set, and test set; S1.2: Construct a local entropy distribution and generate response weights, use the response weights to calibrate the response of shallow features, and construct the Local Entropy Guided Response Calibration Module (LERC). S2: Integrating partial optimal transport theory and attention mechanism, constructing a selective period phase fusion module SPF; S2 includes the following steps: S2.1: Generate saliency weights based on cross-attention and construct a cost matrix between phase features across different periods by combining cosine distance; S2.2: Using the cost matrix, combined with the Sinkhorn algorithm and entropy regularization constraints, a partially optimal transfer matrix is ​​constructed; S2.3: Utilize the partially optimal transfer matrix to perform cross-period phase feature fusion and construct a selective period phase fusion module (SPF); S3: Integrate feature decoupling and multi-branch convolution to construct the progressively refined decoding module PRD; S3 includes the following steps: S3.1: Fuse the decoded features from the previous scale and the features from the current scale to obtain fused features, decouple the fused features, and construct foreground attention features and background attention features; S3.2: Use multi-branch convolution to explore the context of the decoupled features, and refine the fused features based on the context exploration results to construct a progressive refinement decoding module (PRD); S4: Integrate the local entropy-guided response calibration module, the selective phase fusion module, and the progressive refinement decoding module to construct the RAPT-Net model; S5: Use the experimental dataset to complete the training, validation, optimization, and performance evaluation of the model.

2. The method for segmenting liver tumors across time-phase CT based on redundancy sensing according to claim 1, characterized in that, S1.2 includes the following steps: First, the superficial features of the arterial and portal venous phases are projected using a single channel, and the projection results are normalized using the Sigmoid function to obtain the response map. Then, a soft histogram is constructed at the centers of K preset bins, and the local probability distribution of the response map is estimated. Based on the local probability distribution, the normalized local entropy is calculated. Finally, the original superficial features are calibrated using the local entropy to obtain the calibrated superficial features, thus constructing the Local Entropy Guided Response Calibration (LERC) module. The specific calculation process is shown below: In the above formula, This represents the shallow features of the input. This represents a 1×1 convolution projection operation. This represents the Sigmoid function. This represents the normalized response graph. Indicates the center of the k-th box. This represents the smoothing coefficient of the soft histogram. Represents a minimal constant. This indicates a local average pooling operation. This represents the local probability distribution corresponding to the center of the k-th box. This represents the normalized local entropy. This represents the shallow features after calibration using a local entropy-guided response. This indicates an element-wise multiplication operation. This represents the basic retention factor.

3. The method for segmenting liver tumors across time-phase CT based on redundancy sensing according to claim 1, characterized in that, S2.1 includes the following steps: The portal venous phase features and arterial phase features, after being calibrated by local entropy-guided response, are convolved and mapped, and corresponding token sequences are constructed using adaptive pooling. Subsequently, saliency weights for the two-phase tokens are generated based on bidirectional cross-attention. Combining the cosine distance between tokens and a saliency penalty term, a cost matrix for cross-phase feature matching is constructed. The specific calculation process is shown below: In the above formula, and These represent the characteristics of the portal venous phase and the arterial phase, respectively. This indicates an adaptive pooling operation. This indicates a tokenization operation. and The token sequences representing the portal venous and arterial phases. This indicates the operation of calculating the mean. and This indicates the significance weight of the two phases. Indicates the matching cost. Indicates the significance penalty coefficient; S2.2 includes the following steps: To achieve partial transmission, unmatched terms are introduced into the cost matrix to construct an extended cost matrix. Transmission quality for tokens in two phases is allocated based on significance weights, and residual transmission quality is reserved for unmatched terms, resulting in a marginal distribution. Then, a partially optimal transmission objective function is constructed using entropy regularization constraints, and the Sinkhorn algorithm is used to solve for the partially optimal transmission matrix. The specific calculation process is as follows: In the above formula, Represents the extended cost matrix. Indicates the cost of the unmatched item. and Let N and M represent vectors consisting entirely of 1s, respectively. Indicates the matching ratio. b and 'b' represent the marginal distribution of tokens in the portal venous phase and arterial phase, respectively. This represents the set of feasible transports that satisfy the edge distribution constraints. The entropy regularity coefficient is... Let be the transfer matrix to be solved. This is a partially optimal transfer matrix; S2.3 includes the following steps: First, the effective transfer sub-matrices corresponding to the two phase tokens in the partial optimal transfer matrix are extracted. Then, the effective transfer sub-matrices are used to perform cross-phase feature aggregation on the two phase tokens respectively, and the aggregation results are restored to spatial feature maps, yielding cross-phase aggregated features. Finally, the cross-phase aggregated features are concatenated with the original features, and residual fusion is performed through convolution, batch normalization, and activation functions to construct the selective phase fusion module SPF. The specific calculation process is as follows: In the above formula, This is the effective transmission submatrix in the partially optimal transmission matrix. and These represent the inter-period aggregation tokens of the two phases, This indicates the feature map recovery operation. This indicates a batch normalization operation. This represents the activation function. and These represent the residual fusion coefficients of the two phases, respectively. and The portal venous phase and arterial phase characteristics after fusion. This is the encoder output feature.

4. The method for segmenting liver tumors across time-phase CT based on redundancy sensing according to claim 1, characterized in that, S3.1 includes the following steps: The prediction results from the previous scale are upsampled and mapped using the Sigmoid function to obtain a foreground probability map. The decoded features from the previous scale are convolved and upsampled to the same spatial size as the features at the current scale, then concatenated and fused with the encoder output features at the current scale to obtain fused features. Based on the foreground probability map, the fused features are decoupled to obtain foreground attention features and background attention features. The specific calculation process is as follows: In the above formula, This is the prediction result for the previous scale. Foreground probability plot, This represents the Sigmoid function. Indicates an upsampling operation. The encoder outputs features for the current scale. These are the decoding features of the previous scale. As a feature of fusion, As a feature of future focus, Background features to focus on This represents the element-wise multiplication operation; S3.2 includes the following steps: Multi-branch convolution processing is performed on the foreground attention features and the background attention features respectively. Multi-scale contextual information is extracted through convolutional branches and pooling branches with different receptive fields, and the outputs of each branch are fused to obtain the corresponding context exploration results. When the foreground attention features are used as input, a suppression response for suppressing false activation regions is obtained; when the background attention features are used as input, a compensation response for enhancing missed detection regions is obtained. The specific calculation process is as follows: Among them, when hour, , , indicating a suppressed response; when hour, , , indicating a compensation response; , and These represent convolution operations for different receptive fields. This indicates a pooling operation. Indicates use Convolution performs channel mapping operations; Based on the suppression response, the false activation regions in the fused features are suppressed to obtain intermediate features; then, based on the compensation response, the missed detection regions in the intermediate features are enhanced to obtain refined features and prediction results at the current scale, thereby constructing the progressive refinement decoding module (PRD); the specific calculation process is as follows: In the above formula, The intermediate feature after suppression For the refined features, The prediction results are for the current scale. This represents batch normalization and ReLU activation function operations. and They represent the learnable coefficients, This represents the Sigmoid function. This represents the hyperbolic tangent function.

5. A method for segmenting liver tumors across time-phase CT based on redundancy sensing according to claim 1, characterized in that, S4 includes the following steps: Based on a dual-branch U-shaped network architecture, the aforementioned Local Entropy Guided Response Calibration (LERC) module, Selective Phase Fusion (SPF) module, and Progressive Refinement Decoding (PRD) module are integrated to construct the RAPT-Net model. First, a ResNet-34-based dual-branch parallel encoder is constructed to extract multi-level features from the arterial and portal venous phases, respectively. Then, the LERC module is embedded in the shallow encoding stage to suppress low-information-content redundant responses within phases and enhance structurally relevant responses. Next, the SPF module is embedded in the mid-to-high-level encoding stage to establish selective correspondences between phase features and perform cross-phase feature fusion. Finally, the PRD module is embedded in the decoding stage to progressively refine and restore the fused features, thereby outputting the final liver tumor segmentation result.

6. The method for segmenting liver tumors across time-phase CT based on redundancy sensing according to claim 1, characterized in that, S5 includes the following steps: The predefined training set is input into the RAPT-Net model for training. Forward propagation is performed based on the arterial phase image, portal venous phase image and their corresponding annotations of the training samples to obtain the liver tumor segmentation result. The model parameters are then updated based on the difference between the segmentation result and the actual annotation. The model performance during training is validated and optimized using the validation set, and the optimal model weights based on the validation results are saved. Finally, the performance of the trained model is evaluated using the test set.

7. A cross-phase CT liver tumor segmentation device based on redundancy sensing, characterized in that, include: Data preprocessing module: Performs data format conversion, registration, region of interest cropping, grayscale truncation, normalization and annotation on the acquired abdominal CT data of the arterial and portal venous phases, and completes the data division of training set, validation set and test set; Local entropy-guided response calibration module: Constructs a local entropy distribution and generates response weights, and uses the response weights to calibrate the response of shallow features in order to suppress low-information-content redundant responses within phases and enhance structure-related responses; Selective phase fusion module: Generates significant weights based on cross-attention, constructs a cross-phase feature matching cost matrix by combining cosine distance, further constructs a partially optimal transmission matrix by combining partial optimal transmission theory, Sinkhorn algorithm and entropy regularization constraint, and performs cross-phase feature fusion using the partially optimal transmission matrix; Progressive refinement decoding module: The decoded features of the previous scale and the features of the current scale are fused to obtain fused features. The fused features are decoupled to construct foreground attention features and background attention features. Multi-branch convolution is used to explore the context of the decoupled features. The fused features are refined based on the context exploration results. Model training and evaluation module: Uses experimental datasets to train, validate, optimize and evaluate the performance of RAPT-Net model, and realize automatic segmentation of liver tumors in cross-stage CT.

Images