Image segmentation method for small and irregularly shaped pancreatic tumors based on deep learning

Abstract

This invention provides an image segmentation method for small, irregularly shaped pancreatic tumors based on deep learning, comprising the following steps: Step 1: Preprocessing the CT image dataset and then dividing it into training and test sets; Step 2: Constructing a deep segmentation neural network; Step 3: Training the deep segmentation neural network using the training set, selecting and saving the best-performing neural network model; Step 4: Verifying the usability of the saved neural network model using the test set; Step 5: Segmenting the CT images using the deep segmentation neural network trained in Step 4. This invention clarifies that pancreatic tumor segmentation is a challenging problem involving small targets. It introduces multi-scale feature extraction into the network architecture, extracting more detailed spatial features and more comprehensive channel features. Simultaneously, the proposed channel and spatial attention modules improve the network's ability to select grids and its response to lesion detection.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and specifically relates to an image segmentation method for small, irregularly shaped pancreatic tumors based on deep learning. Background Technology

[0002] Pancreatic tumors generally refer to pancreatic ductal tumors (PDAC), which are common malignant tumors of the digestive tract, most often occurring in the head of the pancreas. Pancreatic cancer is highly malignant and progresses rapidly, but its onset is insidious, with atypical early symptoms, and most patients are already in the middle or late stages when they seek clinical attention. The incidence of pancreatic tumors is on the rise both domestically and internationally. Pancreatic cancer ranks sixth in malignant tumor-related mortality, with a global five-year survival rate of 7.5% and an average one-year survival rate of only 18%. Surgical resection is the only effective method for pancreatic cancer patients to achieve a cure and long-term survival. Pre-surgical assessment of the tumor is of significant clinical importance. Imaging examinations are crucial tools for the preliminary diagnosis and accurate staging of pancreatic tumors, and the results of imaging studies guide surgical procedures and specific treatment plans.

[0003] To address the shortcomings of manual segmentation, some scholars are currently researching semantic segmentation of pancreatic tumors. Semantic segmentation involves classifying each pixel in an image. In the field of medical imaging, image segmentation can be applied to image-guided interventions, radiotherapy, and radiodiagnosis. By automatically and accurately segmenting pancreatic tumors semantically based on medical image processing, deep neural networks, and computer vision, it can not only efficiently assist doctors in locating tumors for initial diagnosis but also achieve the ultimate goal of accurately removing lesion boundaries, improving patient survival and prognosis. Automatic semantic segmentation of pancreatic tumors has significant implications, and some proposed methods for pancreatic tumor segmentation have achieved good performance. However, several difficulties currently exist in semantic segmentation of pancreatic tumors from medical images.

[0004] Challenge 1: The pancreas and pancreatic tumors are located deep in the abdomen, making their location inconspicuous and their edges irregular, thus making it difficult to discern a clear outline on imaging. The pancreas and pancreatic tumors have a complex three-dimensional structure, and their shapes vary greatly from individual to individual, mainly manifested by curvature of the pancreatic neck and irregular expansion of the pancreatic head, resulting in discontinuity on a single image slice. Pancreatic tumors adhere to the rich blood vessels and tissues surrounding them, making them difficult to distinguish from the surrounding liver and mucosal tissues on imaging, thus posing a significant technical challenge to the identification and segmentation of the pancreas from surrounding organs.

[0005] Challenge 2: Pancreatic tumors are relatively small, making segmentation difficult and resulting in severe class imbalance. Statistically, the median 3D tumor volume is 12.12 cm³. 3The maximum 2D diameter of tumors ranges from 1.7cm to 6.7cm, with a median of 3.1cm. Some tumors are less than 0.5cm in diameter. In imaging, this manifests as tumors occupying only about 20-50 pixels in a 512x512 pixel slice. Because of the small size of the tumors, not only is the proportion of positive classes in a single slice minimal, but also, in a large number of slices in abdominal images, only a few slices contain tumors as the positive class. Generally, the number of positive classes accounts for only 1 / 4 of the number of negative classes (background class), resulting in a severe imbalance between positive and negative classes, which poses a significant challenge to deep learning.

[0006] Challenge 3: Collecting datasets for pancreatic tumors is difficult, as it involves small-sample learning. A common problem in medical image processing is the difficulty in collecting datasets. Most teams use non-public datasets collected independently in hospitals, typically containing around 100 images. In contrast, ImageNet, a widely used dataset for deep learning, has 10,000 images. The small sample size of pancreatic tumors is not conducive to deep learning. The data format also suffers from inconsistencies due to varying imaging equipment and scan slice thicknesses. Data annotation requires manual drawing by professional doctors, which is time-consuming and costly, making it difficult to create large-scale labeled datasets.

[0007] Several researchers have proposed methods to address the aforementioned issues. Zhu et al. proposed a multi-scale coarse-to-fine segmentation method to screen PDAC in CT images, achieving a Dice score of 57.3% for PDAC cyst segmentation. Turecova et al. proposed a CNN method using deep supervision and attention gating to segment lesions such as liver tumors and pancreatic tumors, achieving a Dice score of 54.66% for pancreatic tumor segmentation. Furthermore, Zhang et al. used a large dataset of approximately 1,000 cases and the nnUNet network for PDAC segmentation using multi-phase CT images, achieving a Dice score of 0.709±0.159 on the multi-dataset, the highest known to date. Zhou et al. proposed using a super-paired network to integrate information from different stages for PDAC segmentation, achieving Dice scores of 63.94±22.74 for multi-phase and 53.08±27.06 for the venous phase. These studies demonstrate improved segmentation performance in PDAC patients. Chen et al. proposed a spiral transformation preprocessing method for pancreatic cancer, which uses a model-driven deep learning approach to segment pancreatic tumors, achieving a Dice score of 66.62±16.37 on the publicly available MSD-pancreas tumor dataset.

[0008] In recent years, scholars have proposed various solutions to address the imbalance between positive and negative samples in small object segmentation, thereby improving the accuracy of small object segmentation. Among these, multi-scale models are widely used due to their effectiveness in extracting features at different scales. He et al. proposed Spatial Pyramid Pooling to address the fixed input size caused by fully connected layers, and proposed parallel SPP layers for multi-level feature extraction, enabling inputs of different sizes to have fixed-size outputs. PSPNet applies multi-level feature extraction to the semantic segmentation field. In the design of the pyramid pooling module, four pools of different sizes are fused, followed by a linear interpolation and a 1×1 convolution. Google's DeepLabV series introduced ASPP, which uses dilated convolutions with different dilation factors to expand the receptive field without loss of resolution, thus fusing multi-scale contextual information. Furthermore, parallel 1×1 convolutions and global pooling are added. In the latest DeepLabV3, concatenated dilated convolutions with different dilation rates are proposed to extract multi-scale context, reducing the loss of detailed features caused by pooling or strided convolutions. Another example of using dilated convolution is the multi-scale enhancer proposed by Shao et al., which employs a multi-scale enhancer (MSB) with channel and spatial attention within the backbone feature pyramid network (FPN). The MSB uses hierarchical dilated convolution (HDC) to determine fine-grained scale changes, thereby improving the accuracy of detecting lesion sizes of different sizes. Summary of the Invention

[0009] This invention addresses the technical problems existing in the prior art by providing an image segmentation method for small and irregularly shaped pancreatic tumors based on deep learning. It clarifies that pancreatic tumor segmentation is a challenging task involving small targets. The invention introduces multi-scale feature extraction into the network architecture, extracting more detailed spatial features and more comprehensive channel features. At the same time, the proposed channel and spatial attention modules improve the network's ability to select grids and its response to lesion detection.

[0010] The technical solution adopted in this invention is: an image segmentation method for small, irregularly shaped pancreatic tumors based on deep learning, comprising the following steps:

[0011] Step 1: Collect CT image datasets of pancreatic tumors, preprocess the CT image datasets, and then divide them into training and test sets;

[0012] Step 2: Construct a deep segmentation neural network. The deep segmentation neural network is an encoder-decoder structure. The encoder part performs downsampling calculation, in which the width and height of the feature map are halved and the number of channels is doubled in each calculation. The decoder part performs upsampling calculation, in which the width and height are doubled and the number of channels is halved in each calculation. Both the encoder and decoder use multi-scale attention modules.

[0013] Step 3: Train a deep segmentation neural network using the training set, and select and save the best-performing neural network model;

[0014] Step 4: Verify the usability of the saved neural network model using a test set;

[0015] Step 5: Use the deep segmentation neural network trained in Step 4 to segment the CT images.

[0016] Further, in step 1, the preprocessing process is as follows: the images in the CT image dataset are sliced ​​to obtain two-dimensional CT slice images; the two-dimensional CT slice images are resampled to adjust the thickness of the CT slice images to be consistent; the Hu values ​​of the CT slice images are adjusted: Hu values ​​less than or equal to -128 are set to 0, Hu values ​​greater than or equal to 127 are set to 255, and Hu values ​​between -128 and 127 are first normalized and then mapped to 0-255; the resolution of the CT slice images is adjusted to 256x256 by center cropping; the labels corresponding to the CT slice images are binarized to contain only 1 and 0, where 1 represents the tumor area and 0 represents other organ and tissue areas.

[0017] Furthermore, both the encoder and decoder sections are multi-layered structures, with encoders and decoders of corresponding layers being interconnected. Each decoder upsamples the input feature map to twice its original size through a deconvolution operation, and then concatenates it with the feature map output by the corresponding encoder from the interconnection, before inputting it into the multi-scale attention module.

[0018] Furthermore, the encoder part includes 5 encoder layers. Each encoder layer includes a multi-scale attention module, two 3x3 convolutional layers and a 2x2 max pooling layer connected in sequence. Each convolution operation is followed by a ReLU layer. The 2x2 max pooling layer downsamples the feature map to half of its original size.

[0019] The decoder section includes four decoder layers. Each encoder layer includes a multi-scale attention module and two 3x3 convolutional layers connected in sequence. Each convolution operation is followed by a ReLU layer. The last decoder layer is followed by a 3x3 convolutional layer, a 1x1 convolutional layer, and a sigmoid function computation layer connected in sequence.

[0020] Furthermore, the working process of the multi-scale attention module is as follows:

[0021] Step 2.1: Generate four branches of feature maps by using 1X1, 1X1 and 3X3, 1X1 and 5X5, and 1X1 convolutions respectively;

[0022] Step 2.2: Concatenate the feature maps of the four branches, perform batch normalization, then perform 3x3 convolution, and then perform batch normalization again to obtain multi-scale feature maps.

[0023] Step 2.3: Input the multi-scale feature map into the channel attention module, calculate the attention weight of each channel feature through global average pooling and activation operations, and filter the channels.

[0024] Furthermore, the working process of the channel attention module is as follows:

[0025] Step 2.31: The multi-scale feature maps are compressed into 1-dimensional vectors using global average pooling;

[0026] Step 2.31: Then, perform a nonlinear transformation on the 1D vector using formula (1) to generate a set of channel weights.

[0027] θ=F ex (T l )=σ(W2φ(W1T l )) (1)

[0028] In the formula, θ represents the channel weight, σ represents the sigmoid function, W1 represents linear mapping 1, W2 represents linear mapping 2, and T l Represents a 1-dimensional vector;

[0029] Step 2.32: Multiply the channel weights with the multi-scale feature map to filter the channels of the multi-scale feature map.

[0030] Furthermore, in step 3, step 3.1: set the hyperparameters, learning rate, number of training epochs, decay factor, and network optimizer parameters during the training process;

[0031] Step 3.2: Feed the training set into the deep segmentation neural network model and perform a 5-fold cross-validation experiment;

[0032] Step 3.3: During the training process, record the performance of the neural network model on the validation set for each round, and save the neural network model with the best performance.

[0033] Furthermore, in step 4, five metrics are selected as evaluation metrics for the test set: Dess index, Jaccard similarity coefficient, recall, precision, and Hausdorff distance. Segmentation tests are performed on the test set to verify the actual segmentation effect of the neural network model saved in step 3.

[0034] Compared with the prior art, the beneficial effects of this invention are:

[0035] 1. This invention addresses the issue of blurred boundaries and small proportions of pancreatic tumors in CT imaging. We specifically preprocessed the dataset. First, we resampled to unify the physical size represented by pixels and effectively increased the number of slices. Then, we adjusted the CT values ​​of the resampled slices to fit the pancreas and tumor, effectively enhancing their boundaries. Finally, we cropped the center to unify the size and appropriately increase the proportion of the tumor in the overall image. We demonstrated the role and necessity of preprocessing, using ablation experiments to show that appropriate data preprocessing significantly improves the final segmentation results, increasing Dice by 0.15.

[0036] 2. This invention addresses the challenge of segmenting small targets by designing a multi-scale attention module, which is applied to each layer of the network structure. The multi-scale approach is useful when the target size varies widely; in this segmentation task, the pancreatic tumor size ranges from 2 to 9000 mm. 2 The multi-scale approach, applied in the encoder, allows for the extraction of more information at different scales. For small, difficult-to-segment tumors, this network design incorporates two 1x1 convolutional channels in its multi-scale configuration. These two channels do not share weights, aiming to extract more feature information from small targets. In the decoder's multi-scale convolution, important information lost due to upsampling is expanded, and feature position shifts caused by upsampling and skip connections are corrected. This supplements decoding information and assists in localization, improving the final result. It effectively extracts features from the image, adding multi-dimensional features, particularly rich in feature extraction for small tumors in the target region.

[0037] 3. This invention employs a channel attention mechanism to weight small target segmentation networks. The channel attention mechanism assigns weights to the rich semantic information extracted. Indiscriminate feature fusion in multi-scale convolutions introduces redundant features or noise into the data stream, negatively impacting computational convergence. Channel attention is used after multi-scale convolutions to calculate the weights of these feature channels and emphasize channel information. Our results show that deeper channel attention helps the network locate pancreatic tumors and better identify their boundaries. This invention adds 3×3 and 1×1 convolutional layers at the end of the decoder to aggregate decoded features and depict more boundary details. Qualitative analysis revealed a significant reduction in false positives in segmentation results, while quantitative analysis showed an increase in the final segmentation score, improving the performance of Dice by approximately 4% compared to other methods. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the present invention;

[0039] Figure 2 This is a flowchart of the data preprocessing process according to an embodiment of the present invention;

[0040] Figure 3This is a diagram of the deep segmentation neural network structure according to an embodiment of the present invention;

[0041] Figure 4 This is a structural diagram of the multi-scale attention module according to an embodiment of the present invention;

[0042] Figure 5 This is a structural diagram of the channel attention module according to an embodiment of the present invention;

[0043] Figure 6 This is a diagram showing the test set segmentation effect of an embodiment of the present invention;

[0044] Figure 7 These are Dys index diagrams of pancreatic tumor segmentation at different sizes, as described in this embodiment of the invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Embodiments of the present invention provide an image segmentation method for small, irregularly shaped pancreatic tumors based on deep learning, such as... Figure 1 As shown, it includes the following steps:

[0047] Step 1: Collect the pancreatic tumor CT image dataset. The pancreatic tumor CT image dataset comes from the 2018 MSD Challenge Task 07 Pancreatic Dataset, which provides 281 CT images and pancreatic tumor segmentation labels.

[0048] Preprocessing of the pancreatic tumor CT image dataset: The ITK-SNAP software was used to read the DICOM files, and the 3D CT images were sliced ​​to obtain 2D CT slice images. Resampling was used to adjust the thickness of the CT slice images in the dataset to be consistent, while ensuring that the image size reflects the imaging size. In this embodiment, the original 512x512x44 CT slice image was adjusted to 389*389*220. The Hu value range of the original CT slice image is...

[0049] [-1007, 1007], setting Hu values ​​less than or equal to -128 to 0, and Hu values ​​greater than or equal to 127 to 255, normalizes Hu values ​​between -128 and 127 before mapping them to 0-255. Next, center cropping is performed to adjust the resolution of the CT slice image to 256x256, cropping edge-irrelevant information, such as... Figure 2 As shown, the labels corresponding to the CT slice images are binarized to contain only 1s and 0s, where 1 represents the tumor region and 0 represents other organ and tissue regions.

[0050] The preprocessed pancreatic tumor CT image dataset was divided into training and testing sets in a 7:3 ratio.

[0051] Step 2: Construct a deep segmentation neural network. For example... Figure 3 As shown, the deep segmentation neural network has an encoder-decoder structure. The encoder part performs downsampling calculations and consists of 5 encoder layers. Each encoder layer includes a multi-scale attention module, two 3x3 convolutional layers, and a 2x2 max-pooling layer connected in sequence. Each convolutional operation is followed by a ReLU layer. The number of channels doubles after each layer passes through the multi-scale attention module, and remains unchanged after passing through the two subsequent 3x3 convolutional layers. The 2x2 max-pooling layer downsamples the feature map to half its original size, meaning that the width and height of the feature map are halved with each calculation.

[0052] The decoder performs upsampling calculations and consists of four decoder layers. Each encoder layer includes a multi-scale attention module and two 3x3 convolutional layers connected sequentially, followed by a ReLU layer after each convolution operation. Encoders and decoders of corresponding layers are skipped. Each decoder upsamples the input feature map to twice its original size through deconvolution, then concatenates it with the feature map output from the corresponding encoder obtained from the skip connection, and inputs it into the multi-scale attention module. Each decoder calculation doubles the width and height while halving the number of channels. The final decoder layer is followed by a 3x3 convolutional layer, a 1x1 convolutional layer, and a sigmoid function calculation layer. The sigmoid function outputs the predicted segmentation result. The 3x3 and 1x1 convolutional layers are used to refine the details of the tumor contour.

[0053] like Figure 3 As shown, the first encoder layer of the encoder part is skip-connected to the fourth decoder layer (the last decoder layer) of the decoder part, and the remaining encoders and decoders are skip-connected sequentially. The feature map output by the fifth encoder layer is directly input into the first decoder layer, upsampled to twice its original size through deconvolution, and then concatenated with the feature map output by the fourth encoder layer before being input into the multi-scale attention module of the first decoder layer. The operation of the remaining decoder layers is similar.

[0054] like Figure 4 As shown, the working process of the multi-scale attention module is as follows:

[0055] Step 2.1: Generate four branches of feature maps by using 1X1, 1X1 and 3X3, 1X1 and 5X5, and 1X1 convolutions respectively;

[0056] Step 2.2: Concatenate the feature maps of the four branches, perform batch normalization, then perform 3x3 convolution, and then perform batch normalization again to obtain multi-scale feature maps.

[0057] Step 2.3: Input the multi-scale feature map into the channel attention module, calculate the attention weight of each channel feature through global average pooling and activation operations, and filter the channels.

[0058] like Figure 5 As shown, the working process of the channel attention module is as follows:

[0059] Step 2.31: The multi-scale feature maps are compressed into 1-dimensional vectors using global average pooling;

[0060] Step 2.31: Then, perform a nonlinear transformation on the 1D vector using formula (1) to generate a set of channel weights.

[0061] θ=F ex (T l )=σ(W2φ(W1T l )) (1)

[0062] In the formula, θ represents the channel weight, σ represents the sigmoid function, W1 represents linear mapping 1, W2 represents linear mapping 2, and T l Represents a 1-dimensional vector;

[0063] Step 2.32: Multiply the channel weights with the multi-scale feature map to filter the channels of the multi-scale feature map.

[0064] Step 3: Train the deep segmentation neural network using the training set. Training is performed on a server using an NVIDIA Tesla T4 graphics card and 15GB of RAM. The environment used is CUDA 10.9 and CDNN 1.0.2. The program is written in TensorFlow and runs on PyCharm. The training experiment is set to 400 generations with early stopping enabled; training stops if the optimal parameters on the validation set are not updated after 100 generations. The optimizer used is Adam. Learning rate decay is set; if the model performance does not improve after 10 generations, a 0.5x learning rate decay will be triggered, with a minimum learning rate of 0.00005. The training data consists of 28,608 2D images. Save the neural network model with the optimal parameters.

[0065] Step 4: Verify the usability of the saved neural network model using a test set; select five metrics as evaluation indicators for the test set: Desseauz index, Jaccard similarity coefficient, recall, precision, and Hausdorff distance. Perform segmentation tests on the test set to verify the actual segmentation performance of the neural network model saved in Step 3. The segmentation performance on the test set is as follows: Figure 6 As shown.

[0066] Dys index for segmentation of pancreatic tumors of different sizes, such as Figure 7 As shown, the network segmentation displays an area of ​​30 mm² on the CT slice.2 The tumor's Dysmark index reached 65.22%; its area was 170 mm. 2 The tumor segmentation index reached 91.55%; two parts of the tumor were shown on a single CT image, with a total area of ​​849 mm². 2 The Dess index for semantic segmentation of the network was 89.38%.

[0067] Step 5: Use the deep segmentation neural network trained in Step 4 to segment the CT images. Input the pancreatic tumor image to be segmented into the network and run the test program to achieve segmentation.

[0068] The present invention has been described in detail above through embodiments, but the content described is only an exemplary embodiment of the present invention and should not be considered as limiting the scope of the present invention. The scope of protection of the present invention is defined by the claims. Any technical solutions designed by those skilled in the art using the technical solutions described in the present invention, or designed by those skilled in the art under the inspiration of the technical solutions of the present invention, within the substance and protection scope of the present invention, to achieve the above-mentioned technical effects, or any equivalent changes and improvements made to the scope of the application, should still fall within the patent protection scope of the present invention.

Claims

1. A method for image segmentation of small and irregularly shaped pancreatic tumors based on deep learning, characterized by: Includes the following steps: Step 1: Collect CT image datasets of pancreatic tumors, preprocess the CT image datasets, and then divide them into training and test sets; Step 2: Construct a deep segmentation neural network. The deep segmentation neural network is an encoder-decoder structure. The encoder part performs downsampling calculation, in which the width and height of the feature map are halved and the number of channels is doubled in each calculation. The decoder part performs upsampling calculation, in which the width and height are doubled and the number of channels is halved in each calculation. Both the encoder and decoder use multi-scale attention modules. Step 3: Train a deep segmentation neural network using the training set, and select and save the neural network models with good performance; Step 4: Verify the usability of the saved neural network model using a test set; Step 5: Use the deep segmentation neural network trained in Step 4 to segment the CT images; Both the encoder and decoder parts are multi-layer structures, with corresponding layers of encoders and decoders connected in a skip connection. Each decoder upsamples the input feature map to twice its original size through a deconvolution operation, and then concatenates it with the feature map output by the corresponding encoder from the skip connection, and inputs it into the multi-scale attention module. The encoder part consists of 5 encoder layers. Each encoder layer includes a multi-scale attention module, two 3x3 convolutional layers and a 2x2 max pooling layer connected in sequence. After each convolution operation, a ReLU layer is connected. The 2x2 max pooling layer downsamples the feature map to half of its original size. The decoder section includes four decoder layers. Each encoder layer includes a multi-scale attention module and two 3x3 convolutional layers connected in sequence. Each convolution operation is followed by a ReLU layer. The last decoder layer is followed by a 3x3 convolutional layer, a 1x1 convolutional layer, and a sigmoid function computation layer connected in sequence. 2.The deep learning-based image segmentation method for small and shape-irregular pancreatic tumors according to claim 1, wherein: In step 1, the preprocessing process is as follows: the images in the CT image dataset are sliced ​​to obtain two-dimensional CT slice images; the two-dimensional CT slice images are resampled to adjust the thickness of the CT slice images to be consistent. Adjust the Hu value of the CT slice image: set Hu values ​​less than or equal to -128 to 0, set Hu values ​​greater than or equal to 127 to 255, and normalize Hu values ​​between -128 and 127 before mapping them to 0-255; perform center cropping to adjust the resolution of the CT slice image to 256x256; binarize the labels corresponding to the CT slice image to contain only 1 and 0, where 1 represents the tumor area and 0 represents other organ and tissue areas. 3.The deep learning-based image segmentation method for small and shape-irregular pancreatic tumors according to claim 1, wherein: The working process of the multi-scale attention module is as follows: Step 2.1: Generate four branches of feature maps by using 1X1, 1X1 and 3X3, 1X1 and 5X5, and 1X1 convolutions respectively; Step 2.2: Concatenate the feature maps of the four branches, perform batch normalization, then perform 3x3 convolution, and then perform batch normalization again to obtain multi-scale feature maps. Step 2.3: Input the multi-scale feature map into the channel attention module, calculate the attention weight of each channel feature through global average pooling and activation operations, and filter the channels. 4.The deep learning-based image segmentation method for small and shape-irregular pancreatic tumors according to claim 3, wherein: The working process of the channel attention module is as follows: Step 2.31: The multi-scale feature map is compressed into a 1-dimensional vector using global average pooling; Step 2.31: Then, perform a nonlinear transformation on the 1D vector using formula (1) to generate a set of channel weights. ； where θ represents the channel weight, represents a sigmoid function, W1 represents a linear mapping 1, W2 represents a linear mapping 2, T l represents a 1-dimensional vector; Step 2.32: Multiply the channel weights with the multi-scale feature map to filter the channels of the multi-scale feature map. 5.The deep learning-based image segmentation method for small and shape-irregular pancreatic tumors according to claim 1, wherein: In step 3, step 3.1: Set the hyperparameters, learning rate, number of training epochs, decay factor, and network optimizer parameters during the training process; Step 3.2: Feed the training set into the deep segmentation neural network model and perform a 5-fold cross-validation experiment; Step 3.3: During the training process, record the performance of the neural network model on the validation set for each round, and save the neural network model with the best results. 6.The deep learning-based image segmentation method for small and shape-irregular pancreatic tumors according to claim 1, wherein: In step 4, five metrics are selected as evaluation metrics for the test set: Dess index, Jaccard similarity coefficient, recall, precision, and Hausdorff distance. Segmentation tests are performed on the test set to verify the actual segmentation effect of the neural network model saved in step 3.

Images