Pancreatic Image Segmentation Method and System Based on an Adaptive Learning Framework

Abstract

The present invention discloses a pancreatic image segmentation method and system based on an adaptive learning framework, belonging to the technical field of medical image segmentation. A pancreatic image to be segmented is obtained; the pancreatic image is input into the adaptive learning framework for segmentation to obtain a segmentation result; the adaptive learning framework includes a different kernel feature fusion module and an adaptive attention contrast learning module. The pancreatic image to be segmented is divided into multiple input image patches, and the multiple input image patches are input into the adaptive attention contrast learning module. After convolution operations with different kernel sizes, multiple feature maps are obtained. The multiple feature maps are input into the different kernel feature fusion module for between-class aggregation of the same feature and between-class separation of different features, and then multiple classification features are output through a feature contrast learning process, thereby obtaining the segmentation result. Feature information is obtained through convolution kernels of different sizes and a coordinated dual-attention branch, so as to enhance the learning of pancreatic detail and overall structure features.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and particularly relates to a pancreatic image segmentation method and system based on a mutual adaptation learning framework. Background Art

[0002] The statements in this part only provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] Pancreatic cancer is a highly lethal malignant tumor with a very low five-year survival rate. Early diagnosis and timely intervention can significantly delay the progression of the disease. Accurately segmenting the pancreas from CT images is crucial for clinicians to monitor pancreatic abnormalities, assist in prevention, early diagnosis, and surgical treatment of pancreatic cancer. Since it is time-consuming and laborious for radiologists to manually label the pancreatic boundary and there are inter-observer differences, it is urgent to explore deep learning methods that can automatically identify and segment the pancreas.

[0004] As a landmark architecture in the field of medical image segmentation, U-Net has been widely used in automatic pancreatic segmentation. However, due to the complex morphology of the pancreas, unclear boundaries with adjacent organs, and accounting for less than 0.5% of the total abdominal volume, the classic U-Net algorithm faces major challenges. Therefore, researchers are committed to enhancing the U-Net model and its variants to improve the feature extraction ability. Multi-scale feature fusion can be used to effectively capture multi-scale features, while large kernel convolution provides a more simplified structure and can also provide a larger receptive field and high-resolution features. However, directly performing multi-scale feature fusion may lead to information redundancy. The attention mechanism can not only alleviate the problem of redundant features but also reduce the uncertainty of pixel classification. However, the existing attention mechanisms do not fully utilize the similarity learning between different scale features, which is crucial for enhancing the model's ability to extract similar features and suppress non-critical differential features. Summary of the Invention

[0005] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a pancreatic image segmentation method and system based on a mutual adaptation learning framework, which obtains feature information through convolutional kernels of different sizes and cooperative dual-attention branches, thereby enhancing the learning of pancreatic detail and overall structure features.

[0006] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0007] In the first aspect, the present invention provides a pancreatic image segmentation method based on a mutual adaptation learning framework, including:

[0008] Obtain a pancreatic image to be segmented;

[0009] The pancreatic image is input into a co-adaptive learning framework for segmentation to obtain a segmentation result; the co-adaptive learning framework includes a different kernel feature fusion module and an adaptive attention contrast learning module, the pancreatic image to be segmented is divided into a plurality of input image blocks, the plurality of input image blocks are input into the adaptive attention contrast learning module, a plurality of feature maps are obtained after convolution operations with different kernel sizes, the plurality of feature maps are input into a different kernel feature fusion module for inter-class aggregation between the same features and inter-class separation between different features, and then a plurality of classification features are output through a feature contrast learning process to obtain a segmentation result.

[0010] According to a further technical solution, the different kernel feature fusion module includes a parallel small kernel convolution module and a parallel large kernel convolution module, a parallel global context perception branch and a detail sensitive feature enhancement branch.

[0011] A further technical solution is that after the global context-aware branch performs different convolution operations on the output features of the small-core convolution module and the output features of the large-core convolution module, complementary features are generated by element-by-element addition, and the complementary features are adaptively average pooled to generate multiple compressed features, and then a comprehensive weight map is generated by element-by-element addition.

[0012] A further technical solution adopts a recalibration mechanism to adjust the importance of each feature element in the comprehensive weight map to obtain globally context-aware perceptual features.

[0013] According to a further technical solution, the complementary features in the detail-sensitive feature enhancement branch are sequentially subjected to the operations of GELU nonlinear activation function, convolution, and Sigmoid activation function to obtain interactive features, and the interactive features are element-wise multiplied with the complementary features to obtain detail-sensitive features.

[0014] A further technical solution is to design a contrast loss between the global context-aware branch and the detail-sensitive feature enhancement branch. The loss function is defined as follows:

[0015]

[0016]

[0017]

[0018] in, Represents contrast loss The weight of represents the contrast loss, Expressed as the loss weight of each layer encoder, Indicates sensory loss, represents the Euler number, Indicated in The value of the ith voxel represents the value of the ith voxel of the detail-sensitive feature in the ith layer encoder; The value of the ith voxel represents the value of the ith voxel of the perception feature in the

[0019] ith layer encoder that belongs to the

[0020]

[0021]

[0022] wherein, represents the weight coefficient, represents the Dice loss, represents the region loss, represents the number of iterations of model training, represents the total number of iterations of model training.

[0023] In a second aspect, the present invention provides a pancreatic image segmentation system based on a mutual adaptation learning framework, including:

[0024] A data acquisition module configured to: acquire a pancreatic image to be segmented;

[0025] An image segmentation module configured to: input the pancreatic image into the mutual adaptation learning framework for segmentation to obtain a segmentation result; the mutual adaptation learning framework includes a different kernel feature fusion module and an adaptive attention contrast learning module, divides the pancreatic image to be segmented into multiple input image patches, inputs the multiple input image patches into the adaptive attention contrast learning module, obtains multiple feature maps after convolution operations with different kernel sizes, inputs the multiple feature maps into the different kernel feature fusion module for between-class aggregation of the same features and between-class separation of different features, and then outputs multiple classification features through a feature contrast learning process, thereby obtaining the segmentation result.

[0026] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps in the pancreatic image segmentation method based on the mutual adaptation learning framework as described in the first aspect are implemented.

[0027] In a fourth aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps in the pancreatic image segmentation method based on the mutual adaptation learning framework as described in the first aspect are implemented.

[0028] The above one or more technical solutions have the following beneficial effects:

[0029] The present invention utilizes the collaborative dual-attention branches in the different kernel feature fusion module (AFDK) to fuse global context information and local detail features simultaneously, enhancing the feature extraction ability of the network. Through adaptive attention contrast learning (AACL), the present invention can dynamically adjust the similarity between the features extracted by different kernels and optimize the feature representation. In addition, the present invention designs a boundary-sensitive region optimization (BSRO) strategy, which balances the model's attention to the core region and the detail boundary by integrating boundary information and region optimization, thereby improving the segmentation accuracy.

[0030] The present invention proposes a different kernel feature fusion module (AFDK), which obtains feature information through convolutional kernels of different sizes and collaborative dual-attention branches, thereby enhancing the learning of pancreatic detail and overall structure features.

[0031] The present invention develops an adaptive attention contrast learning module (AACL), which promotes the dynamic similarity learning between the features extracted by different kernels and optimizes the representation of key features.

[0032] The present invention designs a boundary-sensitive region optimization strategy (BSRO) to balance the model's attention to the core region and the detail boundary during the training process, thereby improving the performance of pancreatic segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings forming a part of this invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0034] Figure 1 is the overall structure diagram of the mutual adaptation learning framework in the embodiment of the present invention;

[0035] Figure 2 is the structure diagram of the global context awareness branch in the different kernel feature fusion module in the embodiment of the present invention;

[0036] Figure 3It is the structure diagram of the detail-sensitive feature enhancement branch in the different nuclear feature fusion module of the embodiment of the present invention;

[0037] Figure 4 It is the structure diagram of the adaptive attention contrast learning module of the embodiment of the present invention;

[0038] Figure 5 It is the ground truth label in the pancreas segmentation label of the embodiment of the present invention;

[0039] Figure 6 It is the boundary position information in the visualization of the BLI corresponding to the pancreas label of the embodiment of the present invention;

[0040] Figure 7 It is three segmentation examples of pancreas segmentation on the NIH dataset of the embodiment of the present invention. Detailed implementation manners

[0041] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0042] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0043] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0044] Embodiment 1

[0045] As Figure 1 shown, this embodiment discloses a pancreas image segmentation method based on a mutual adaptation learning framework, and the method includes the following steps:

[0046] S1: Obtain the pancreas image to be segmented;

[0047] S2: Input the pancreatic image into the Dual-Kernel Mutual Adaptation Learning framework (DKMAL) for segmentation to obtain a segmentation result. The Dual-Kernel Mutual Adaptation Learning framework includes a Different Kernel Feature Fusion module and an Adaptive Attention Contrastive Learning module. The pancreatic image to be segmented is divided into multiple input image patches, which are input into the Adaptive Attention Contrastive Learning module. After convolution operations with different kernel sizes, multiple feature maps are obtained. The multiple feature maps are input into the Different Kernel Feature Fusion module for between-class aggregation of the same features and between-class separation of different features, and then multiple classification features are output through a feature contrastive learning process, thereby obtaining the segmentation result.

[0048] In this embodiment, a new end-to-end network is provided, namely the Dual-Kernel Mutual Adaptation Learning framework (DKMAL), which includes three core components: the Different Kernel Feature Fusion module (AFDK), the Adaptive Attention Contrastive Learning module (AACL), and the Boundary-Sensitive Region Optimization strategy (BSRO), for the joint segmentation of the pancreas in abdominal CT images.

[0049] In this embodiment, as Figure 1 shown, the Different Kernel Feature Fusion module AFDK sequentially includes a juxtaposed small kernel convolution module and a large kernel convolution module, and a juxtaposed global context-aware branch and a detail-sensitive feature enhancement branch. The small kernel convolution module is a convolution of size 3×3×3 (Conv2), and the large kernel convolution module is a convolution of size 7×7×7 (Conv1).

[0050] The AFDK module emphasizes the importance of feature extraction through convolution kernels of different sizes and feature fusion through collaborative dual-attention branches, and considers the interaction between global context information and local details in the segmentation task. The AFDK module uses two different sizes of convolution kernels to fully utilize the feature representation at multiple scales, thereby improving the model performance. Specifically, the convolution operation of size (7×7×7) is used to capture broader context information, which helps to learn the pancreatic structure. The convolution operation of size (3×3×3) is used to capture more detailed and localized features. Subsequently, the above features are fed into the collaborative dual-attention branches, namely the Global Context-Aware Branch (GCAB) and the Detail-Sensitive Feature Enhancement Branch (DFEB), for further information extraction.

[0051] As Figure 2 shown, the Global Context-Aware Branch (GCAB) uses global average pooling operations to obtain feature maps in three different spatial dimensions (height (H), width (W), and depth (D)), and then performs weighted fusion to achieve global context awareness of the input image patches. Specifically, the pancreatic image to be segmented is input into the small kernel convolution module and the large kernel convolution module respectively for feature extraction, and feature maps are obtained from the standard small kernel convolution (small kernel convolution module S) The feature map is obtained from the large-kernel convolution (large-kernel convolution module L). , ( (representing the number of channels), a convolution operation is performed on the global feature , and a convolution operation is performed on the local feature . After different convolution operations, complementary features are generated by element-wise addition . . .

[0052] To evaluate the importance weights of different spatial positions in the complementary feature , the present invention performs adaptive average pooling on to generate three compressed feature maps of different spatial dimensions (H, W, and D). As shown in Figure 2 , after these features are broadcast, a comprehensive weight map is generated by element-wise addition. The specific operations are as follows:

[0053] (1)

[0054] Among them, represents the element-wise addition operation, represents the weighted sum after global average pooling is performed separately on the H, W, and D dimensions, and its size is consistent with , represents the height position where the voxel is located, represents the width position where the voxel is located, represents the depth position where the voxel is located.

[0055] To model the complex interaction dependencies between the channel dimensions of different spatial positions in the comprehensive weight map , the present invention adopts a recalibration mechanism to adjust the importance of each feature element, thereby obtaining a globally context-aware perceptual feature . The recalibration mechanism is that the comprehensive weight map sequentially passes through , the GELU activation function, and the Sigmoid activation function to output features, and the output features are multiplied element-wise with the feature map to obtain the feature .

[0056] This process can be formally expressed as:

[0057] (2)

[0058] Among them, Denotes an element-wise multiplication operation; Corresponds to a (1x1x1) convolution, used to keep the size and number of channels of the input features unchanged; Are local features; Corresponds to a 3D convolution operation that doubles the number of feature channels, halving the number of channels while keeping the size of the input features unchanged; Corresponds to a 3D convolution operation that halves the number of feature channels, halving the number of channels while keeping the size of the input features unchanged; And Represent the GELU activation function and the Sigmoid activation function respectively.

[0059] Such as Figure 3 As shown, the Detail-Sensitive Feature Enhancement Branch (DFEB) is used as a supplement to the GCAB. It uses the GELU non-linear activation function to enhance the interaction of local details in the input image, improving the network's ability to capture small structures, edges, and other subtle but important details.

[0060] In the Detail-Sensitive Feature Enhancement Branch, the same operations in the GCAB are performed on the feature maps and to obtain complementary features . The complementary features are successively passed through the GELU non-linear activation function, , and the Sigmoid activation function to output corresponding features. The output features are multiplied element-wise with the complementary features to obtain the detail-sensitive features .

[0061] DFEB enhances the network's attention to local features, ensuring that the network does not ignore subtle local differences while grasping the global context. The operations of DFEB can be expressed as:

[0062] (3)

[0063] Where is the detail feature, is the complementary feature, and represent the GELU activation function and the Sigmoid activation function respectively, corresponds to a (1x1x1) convolution, used to keep the size and number of channels of the input features unchanged.

[0064] In this embodiment, such as Figure 4As shown, traditional contrastive learning methods obtain general feature representations by constructing positive and negative sample pairs, which requires explicitly defining positive and negative samples to make the learning objective clearer. However, in practical applications, defining and constructing appropriate positive and negative sample pairs usually poses certain challenges. For this reason, the present invention proposes a novel adaptive attention contrastive learning module (AACL module), which does not require explicit construction of positive and negative sample pairs and supports end-to-end training to optimize the pancreas segmentation task.

[0065] The AACL module introduces a new paradigm that uses two input image patches and as the initial inputs for contrastive learning. Through convolutions with different kernel sizes and a simple feature contrastive learning process, the model generates four different feature maps. Specifically, the two input image patches and are convolved by a convolution of size 3 (small kernel convolution module S) to obtain and , and the two input image patches and are convolved by a convolution of size 7 (large kernel convolution module L) to obtain and ; and , and perform intra-class aggregation between the features of the same image patch and inter-class separation between different image patches on the two attention branches of GCAB and DFEB, and then conduct contrastive learning to obtain four different feature maps.

[0066] These feature maps achieve intra-class aggregation between the features of the same image patch and inter-class separation between different image patches on the two attention branches of the global context awareness branch (GCAB) and the detail-sensitive feature enhancement branch (DFEB). Specifically, in order to enhance the contrast and complementarity of these four features generated by the two attention branches of GCAB and DFEB, the present invention designs a contrastive loss between GCAB and DFEB, focusing on capturing the similarity of the key features extracted through different convolutional kernels during the entire learning process. The AACL loss function is defined as follows:

[0067] (4)

[0068] (5)

[0069] (6)

[0070] Among them, represents the loss function of the AACL method, Represents the contrast loss, Represents the number of layers of the network encoder, Represents the loss weight of each layer of the encoder, Represents the serial number of the voxel point, Represents the Euler number, And Respectively represent the context-aware feature and the detail-sensitive feature mentioned above. Therefore, Represents at the The value of the i-th voxel of the context-aware feature in the i-th layer of the encoder, ; Represents at the The value of the j-th voxel of the detail-sensitive feature in the i-th layer of the encoder, ; Represents at the The value of the k-th voxel belonging to the m-th channel in the context-aware feature in the i-th layer of the encoder, ; Represents at the The value of the l-th voxel belonging to the n-th channel in the detail-sensitive feature in the i-th layer of the encoder, ; ; ; Represents the total number of elements, Represents the total number of channels, Represents the weight of the contrast loss, Represents the perceptual loss used to enhance the discrimination ability between And .

[0071] To better adapt to the needs of the model in different training stages, the present invention proposes an adaptive contrast learning strategy to further learn more accurate feature information. Specifically, the adaptive contrast learning strategy means that when the unimproved counting threshold Reaches the preset value, then Is added to the training stage. The weight of Can be self-adjusted through an adaptive strategy with the help of a smoothing reward factor And a parameter (as shown in the flowchart in Figure 4 ). Calculated by the exponential moving average (EMA) method:

[0072] (7)

[0073] (8)

[0074] (9)

[0075] Represents the finally calculated reward factor. Among them, is the number of training iterations; is an integer, indicating how many times the network iterates before performing a validation; is the smoothing coefficient in the EMA method; Represents at the th iteration, the overall network loss and the exponential moving average. is a combination of Dice loss (focusing on global structural consistency) and cross-entropy loss (sensitive to per-pixel classification), as shown in the following formula (13).

[0076] (10)

[0077] Among them, is the total number of model training iterations.

[0078] As Figure 4 shown in the flowchart of, remains zero until reaches the preset value, and then participates in the optimization process of DKMAL. When and is a uniformly distributed random number), a value within the range of is randomly generated and added to . When and or , is added to .

[0079] In this embodiment, the present invention designs a boundary-sensitive region optimization (BSRO strategy), which integrates pancreatic boundary information to guide the network to pay more attention to balancing the attention of the pancreatic core region and the detailed boundary. Considering that the pancreas is small in volume in three-dimensional abdominal CT images, the BSRO strategy aims to solve the inherent class imbalance problem in pancreatic segmentation, thereby improving the boundary sensitivity and segmentation accuracy of the model for small targets. The present invention defines the BSRO loss function (Formula 11), which not only focuses on the consistency of the global structure but also is sensitive to pixel-level classification.

[0080] (11)

[0081] (12)

[0082] Among them, represents The weight coefficient is used to balance during the training process and . This enables each component of to be adjusted at a finer granularity, enabling the model to better capture each feature according to individual data during the training process; represents the Dice loss, represents the region loss,

[0083] (13)

[0084] Among them, represents the probability that pixel is predicted to belong to class , represents the label value (0 or 1), indicating whether pixel belongs to class , while represents the total number of classes, represents the total number of pixels in the image.

[0085] (14)

[0086] (15)

[0087] Among them, represents the boundary pixel closest to pixel in the label, represents the value of the label map generated according to the label of the network input image, represents the class of the label value.

[0088] As Figure 6 shown, is used to encapsulate the boundary location information (BLI), which can enhance the model's ability to learn boundary details under the guidance of geometric prior knowledge.

[0089] Figure 5 and Figure 6 show visual examples of the label and its corresponding BLI. Figure 5 Only shows the black and white areas with clear regions, and the two colors represent the regions inside and outside the class respectively; Figure 6 shows the gradient color change with obvious boundaries, and the elements outside the class region are set to 1 (black).

[0090] Dataset introduction and experimental parameter settings:

[0091] The method of the present invention was evaluated on the NIH Pancreas dataset (NIHPancreas), which contains 82 non-contrast-enhanced abdominal CT scans and their corresponding pancreatic labels manually annotated by NIH radiologists using ITK-SNAP software. The size of each image is 512×512×[94 - 198], and the slice thickness is generally between 1 mm and 2.5 mm, providing high-resolution details. The present invention used a five-fold cross-validation method for experiments, splitting the dataset into five subsets, with each subset containing 17, 17, 16, 16, and 16 samples respectively. In each round of cross-validation, the present invention selected four subsets for model training and used the remaining one subset for model testing.

[0092] The DKMAL framework of the present invention is implemented based on the PyTorch framework and runs on a Linux system equipped with an NVIDIA GeForce RTX 3090 GPU with 24GB of memory. The maximum number of training iterations is set to 20,000, and one batch of data is processed in each iteration. The present invention uses the Adam optimizer with an initial learning rate of 0.0001 to optimize the network parameters. In terms of data preprocessing, the present invention normalizes the image intensity values from -100 to 240 to the range of 0.0 to 1.0, and values outside this range are clipped. The balance of the dataset is achieved through random cropping, ensuring that each crop contains one positive sample (foreground) and one negative sample (background), and the cropped volume is fixed at 96×96×96. To enhance data diversity, the present invention introduces random intensity changes of up to 10% with a probability of 50% and applies random affine transformations such as rotation and scaling. The present invention uses four evaluation metrics to evaluate the segmentation performance: Dice Similarity Coefficient (DSC), Average Symmetric Surface Distance (ASSD), Positive Predictive Value (PPV), and Sensitivity (SEN).

[0093] Experimental results:

[0094] The DKMAL framework was tested on the NIH Pancreas dataset. The average DSC of DKMAL for pancreas segmentation was 88.03±2.60%, and it performed excellently with a minimum DSC of 74.91%, indicating that DKMAL has high robustness in dealing with the most challenging segmentation cases. In addition, the standard deviation of the average DSC of DKMAL in different pancreatic cases was 2.60%, further demonstrating the stability of the method on the NIH dataset.

[0095] Figure 7 Three segmentation examples (Example 1, Example 2, Example 3) on the pancreatic region of the NIHPancreas dataset are shown. The pancreatic region is marked by the position of the solid arrow, and the segmented region is marked by the dashed arrow.

[0096] Example Two

[0097] This embodiment discloses a pancreatic image segmentation system based on a mutual adaptation learning framework, including:

[0098] A data acquisition module, which is configured to: acquire a pancreatic image to be segmented;

[0099] An image segmentation module, which is configured to: input the pancreatic image into the mutual adaptation learning framework for segmentation to obtain a segmentation result; the mutual adaptation learning framework includes a different kernel feature fusion module and an adaptive attention contrast learning module, divide the pancreatic image to be segmented into multiple input image patches, input the multiple input image patches into the adaptive attention contrast learning module, obtain multiple feature maps after convolution operations with different kernel sizes, input the multiple feature maps into the different kernel feature fusion module for between-class aggregation of the same feature and between-class separation of different features, and then output multiple classification features through a feature contrast learning process, thereby obtaining the segmentation result.

[0100] Embodiment III

[0101] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the method in Embodiment I are implemented.

[0102] Embodiment IV

[0103] The purpose of this embodiment is to provide a computer-readable storage medium. A computer-readable storage medium stores a computer program, and when the program is executed by a processor, the steps of the method in Embodiment I are executed.

[0104] The steps involved in the devices in the above Embodiments III and IV correspond to those in Method Embodiment I. For specific implementation manners, reference may be made to the relevant description part of Embodiment I. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0105] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple of them can be fabricated into a single integrated circuit module. The present invention is not limited to any specific combination of hardware and software.

[0106] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0107] Although the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, they are not limitations on the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A pancreatic image segmentation method based on a co-adaptive learning framework, characterized in that: include: Acquire a pancreas image to be segmented; The pancreatic image is input into a co-adaptive learning framework for segmentation to obtain a segmentation result; the co-adaptive learning framework includes a different kernel feature fusion module and an adaptive attention contrast learning module, the pancreatic image to be segmented is divided into a plurality of input image blocks, the plurality of input image blocks are input into the adaptive attention contrast learning module, a plurality of feature maps are obtained after convolution operations with different kernel sizes, the plurality of feature maps are input into a different kernel feature fusion module for inter-class aggregation between the same features and inter-class separation between different features, and then a plurality of classification features are output through a feature contrast learning process to obtain a segmentation result.

2. The pancreatic image segmentation method based on the co-adaptive learning framework according to claim 1, characterized in that: The different kernel feature fusion module includes a parallel small kernel convolution module and a parallel large kernel convolution module, a parallel global context perception branch and a detail sensitive feature enhancement branch.

3. The pancreatic image segmentation method based on the co-adaptive learning framework as claimed in claim 2, characterized in that: The global context-aware branch performs different convolution operations on the output features of the small-core convolution module and the output features of the large-core convolution module, generates complementary features by element-by-element addition, performs adaptive average pooling on the complementary features to generate multiple compressed features, and then generates a comprehensive weight map by element-by-element addition.

4. The pancreatic image segmentation method based on the co-adaptive learning framework as claimed in claim 3, characterized in that: A recalibration mechanism is adopted to adjust the importance of each feature element in the comprehensive weight map to obtain global context-aware perceptual features.

5. The pancreatic image segmentation method based on the co-adaptive learning framework as claimed in claim 2, characterized in that: The complementary features in the detail-sensitive feature enhancement branch are sequentially subjected to the operations of the GELU nonlinear activation function, convolution, and Sigmoid activation function to obtain interactive features, and the interactive features are element-wise multiplied with the complementary features to obtain detail-sensitive features.

6. The pancreatic image segmentation method based on the co-adaptive learning framework according to claim 2, characterized in that: A contrast loss is designed between the global context-aware branch and the detail-sensitive feature enhancement branch, and the loss function is defined as follows: in, Represents contrast loss The weight of represents the contrast loss, Expressed as the loss weight of each layer encoder, Indicates sensory loss, represents the Euler number, Indicated in The perceptual features in the layer encoder The value of a voxel, Indicated in The first layer of detail-sensitive features in the encoder The value of a voxel, Indicated in The perceptual features in the layer encoder belong to Channel No. The value of a voxel, Indicated in The detail-sensitive features in the layer encoder belong to Channel No. The value of the voxel.

7. The pancreatic image segmentation method based on the co-adaptive learning framework according to claim 1, characterized in that: The co-adaptive learning framework uses a boundary-sensitive region optimization strategy and defines a boundary-sensitive region optimization loss function, which is expressed as: in, represents the weight coefficient, represents the Dice loss, represents the area loss, represents the number of model training iterations, Represents the total number of model training iterations.

8. A pancreatic image segmentation system based on a co-adaptive learning framework, characterized in that: include: A data acquisition module is configured to: acquire a pancreas image to be segmented; The image segmentation module is configured as follows: the pancreatic image is input into a co-adaptive learning framework for segmentation to obtain a segmentation result; the co-adaptive learning framework includes a different kernel feature fusion module and an adaptive attention contrast learning module, the pancreatic image to be segmented is divided into a plurality of input image blocks, the plurality of input image blocks are input into the adaptive attention contrast learning module, a plurality of feature maps are obtained after convolution operations with different kernel sizes, the plurality of feature maps are input into a different kernel feature fusion module for inter-class aggregation between the same features and inter-class separation between different features, and then a plurality of classification features are output through a feature contrast learning process to obtain a segmentation result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the pancreatic image segmentation method based on the co-adaptive learning framework as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the pancreatic image segmentation method based on the co-adaptive learning framework as described in any one of claims 1 to 7 are implemented.

Images