A MLP-like medical image segmentation method suitable for multiple modalities

Abstract

The application belongs to the field of medical image processing, and discloses a kind of MLP medical image segmentation method suitable for multiple modalities, feature embedding operation is carried out on input image to convert it into a group of characteristic vector sequences;Fully symmetrical U-shaped network structure is used to encode and decode the characteristic vector;The weighted sum of Dice loss and modified cross-entropy is used as the loss function, and the Adam optimizer is used to train the network.The application uses improved ChannelFC as the basic operator to extract image features, calculates the offset through a linear operation, and the FC operation deforms in the spatial dimension according to the offset, so that more detailed features can be extracted.The application effectively filters the interference of a large amount of background information in medical images, thereby improving the segmentation accuracy.Compared with the traditional convolution-based or self-attention method, it has lower computational complexity and fewer parameters, and is suitable for CT, MRI and other different modal medical image datasets.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and more particularly to a multi-modal MLP-like medical image segmentation method. Background Technology

[0002] Medical image segmentation is a crucial component of computer-aided diagnosis, especially since accurate and efficient medical image segmentation algorithms play a vital role in image-guided clinical surgery. Thanks to advancements in convolutional neural networks (CNNs) and visual Transformers, medical image segmentation has seen significant breakthroughs. Recently, several MLP-like methods have demonstrated excellent performance across various computer vision tasks, becoming an effective alternative in this field.

[0003] Early MLP-like methods were generally divided into two parts: Spatial Fully Connected (SpatialFC) for extracting spatial semantic information and Channel Fully Connected (ChannelFC) for fusing channel features. Since SpatialFC performs fully connected operations across the entire spatial dimension, it not only has considerable computational complexity but also limits the size of the input feature map, making it unable to handle variable-scale inputs. Therefore, this method does not perform ideally for tasks such as semantic segmentation. Furthermore, we note that medical images often exhibit a low proportion of foreground organs. Directly using SpatialFC for global operations results in the extracted features containing a large amount of interference from background information. One approach to address these issues is to abandon the global SpatialFC and improve ChannelFC to enable it to extract local spatial semantic information.

[0004] The paper "T.Yu, X.Li, Y.Cai, M.Sun and P.Li, S2-MLP: Spatial-Shift MLP Architecture for Vision, 2021" introduces a spatial-shift operation between two ChannelFCs to extract spatial semantic information through spatial shifting. The paper "Cyclemlp: A mlp-like architecture for dense prediction", S.Chen, E.Xie, C.Ge, D.Liang and P.Luo, arXiv preprint arXiv:2107.10224, 2021" uses a cyclic shifting of ChannelFC sampling positions to extract spatial features. While these methods address the computational complexity of SpatialFC, they are still susceptible to interference from a large amount of background information in datasets with sparse foregrounds, such as medical images. Therefore, a new improved approach is needed to solve this problem. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multi-modal MLP-like medical image segmentation method. It uses MLP-like operations to extract features of foreground organs in medical images of various modalities and segment them. An offset is obtained through a linear operation, which is used to deform the ChannelFC in the spatial dimension in order to extract more organ features. The method abandons the computationally complex SpacialFC to ensure segmentation efficiency.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: a multi-modal MLP-like medical image segmentation method, comprising the following steps:

[0007] Step 1: Collect medical image segmentation datasets and select a standardization method based on the modality of the medical image segmentation datasets;

[0008] Step 2: Perform data preprocessing on the standardized medical image segmentation dataset to obtain slices;

[0009] Step 3: Construct a fully symmetric U-shaped network for image segmentation. The overall structure includes an image embedding layer, an encoder, a decoder, and a segmentation head. The image embedding layer embeds pre-processed slices into feature maps. The encoder mainly consists of basic blocks and downsampling layers, which downsample the feature maps output by the image embedding layer step by step to output a set of feature maps at different scales. The decoder mainly consists of basic blocks and upsampling layers, which upsample the feature maps step by step and make skip connections with the feature maps output by the corresponding positions in the encoder in the channel dimension, finally obtaining a feature map with the same output scale as the image embedding layer. The segmentation head reshapes the feature maps output by the decoder into slice sizes and compresses the number of channels to be the same as the number of classes. SoftMax operation is performed in the channel dimension to obtain the class of each pixel. The basic blocks in the encoder and decoder use an improved Channel FC as the basic operator.

[0010] Step 4: Select a loss function and training strategy to optimize the weight parameters of the U-shaped network model, and use the trained U-shaped network for inference.

[0011] The image embedding layer takes pre-processed slices as input and outputs an embedding feature map. It divides the input slices into equal-sized image blocks using a sliding window, then converts these image blocks into embedding feature vectors. Finally, it rearranges these embedding feature vectors according to the relative positions of the image blocks to form the embedding feature map. The entire process is shown in the following formula.

[0012] IEL(img)=[patch1;…;(patch) n ]·

[0013] in, Let be the weight matrix of the image embedding layer; when the sliding window size is k and the stride is s, the process of the image embedding layer is equivalent to a convolution with a kernel size of k, a stride of s, an input channel of the number of channels in patch, and an output channel of D.

[0014] The basic operator of the basic block is an improved Channel FC, with both input and output being a sequence of feature vectors of the same dimension;

[0015] by This represents the input feature vector sequence. This represents the offset corresponding to each feature vector calculated from the feature map output by the image embedding layer. It is the weight matrix of the improved Channel FC. It is an improved Channel FC bias; the shape of the Channel FC is changed according to the offset, so that it is more biased towards the foreground image in the spatial dimension, thereby aggregating more contextual information at the feature points. If only floating-point multiplication is considered, the FLOPs of this operator are linearly related to the length N of the input feature vector sequence, just like ordinary Channel FC.

[0016] The improved Channel FC operator formula is as follows:

[0017]

[0018] C in The input feature vector has 10 channels; c represents the number of channels; p represents the number of channels. i (c,O i,c () indicates the sampling position on channel c when the Channel FC operator calculates the i-th eigenvector;

[0019] In order to calculate p at a lower cost i The feature vector is alternately divided into g groups along the channel dimension, with each group sharing a position. The offset O is then shifted from... Compress to

[0020] p i It mainly consists of two parts, and the specific algorithm formula is shown below;

[0021]

[0022] The basic block is constructed using the improved Channel FC as the basic operator. The input and output are feature maps of the same size. The basic block consists of two stages. The first stage extracts spatial and channel feature information through the improved Channel FC. The second stage introduces nonlinear transformation to further enhance the feature representation capability. Each stage includes a Layer Norm and a residual connection.

[0023] (1): In the first stage, a linear operation is performed on the feature map output by the image embedding layer to obtain a set of offsets. This set of offsets is then compared with the feature map. Figure 1 The input is fed into the improved Channel FC for execution; since the improved Channel FC operation only accepts two-dimensional tensors, the feature map output from the image embedding layer is flattened along the spatial dimension to convert it into a sequence of feature vectors. An asymmetric parallel design is adopted, in which F is expanded along the x-axis and y-axis and FC operation is performed in parallel; in order to further integrate the features of the feature map channel dimension, an additional ChannelFC branch is added, and finally it is summed with the FC operation results of F in both directions and output.

[0024] (2): In the second stage, nonlinear operations are introduced using the Channel FC→GeLU→Channel FC method to improve feature representation capability; the activation function is GeLU, and its expression is as follows.

[0025]

[0026] x represents the value at a certain location in the feature map;

[0027] The standardization method for selecting the modality of the medical image segmentation dataset in step 1 is as follows:

[0028] (1) When the modality of the medical image segmentation dataset is CT image, the pixel values ​​of the CT image are first cropped to the range of [-125, 275], and then standardized to [0, 1].

[0029] (2) When the modality of the medical image segmentation dataset is MRI, no cropping is performed, and the entire MRI image is directly standardized to [0,1].

[0030] The specific method for preprocessing in step 2 is as follows:

[0031] The standardized medical image segmentation dataset was divided into training and validation sets in an 8:2 ratio. In the training set, the entire 3D image was cut into cross-sectional slices along the vertical direction. The cross-sectional slices were scaled to 224×224 pixels using bilinear interpolation. Finally, data augmentation methods were used to improve the generalization ability of the U-shaped network model.

[0032] In step 3, a completely symmetrical U-shaped network is constructed, with the following specific structure: image embedding layer → encoder → decoder → segmentation head; the encoder has the following specific structure: basic block × 2 → downsampling layer → basic block × 2 → downsampling layer → basic block × 6 → downsampling layer → basic block × 2; the decoder has the following specific structure: upsampling layer → skip connection → basic block × 6 → upsampling layer → skip connection → basic block × 2 → upsampling layer → skip connection → basic block × 2 → upsampling layer.

[0033] The specific method for step 4 is as follows:

[0034] (1) The weighted sum of Dice loss and modified cross-entropy is used as the loss function, and its calculation formula is as follows:

[0035]

[0036] X represents the true label, and Y represents the predicted label. ic The sign function is y when the true class of sample i is equal to c. ic Set to 1 otherwise set to 0, where β represents the weight of Dice and N represents the number of samples;

[0037] The Adam optimizer is used for backward gradient propagation to optimize the weight parameters, and the learning rate decay strategy is Poly decay, as shown in the following formula.

[0038]

[0039] lr represents the learning rate, base lr The initial learning rate is represented by , epoch represents the current iteration number, and power represents the decay rate.

[0040] The weight parameters of the U-shaped model are initialized using a Gaussian distribution, eliminating the need for pre-training on a large-scale dataset.

[0041] (2) During verification, a standardized 3D instance is input into the U-shaped network model at one time. The U-shaped network model reasones layer by layer along the z-axis. Before the slices are fed into the U-shaped network model, they are scaled to the same size as the training set to ensure the maximum segmentation accuracy. Finally, the results of each slice are reassembled into a 3D segmentation result output. The Dice coefficient (DSC) and Hausdorff distance (HD) are used as evaluation criteria.

[0042] The slice is a cross-sectional slice of a 3D image.

[0043] The beneficial effects of this invention are: First, the network does not perform fully connected operations directly in the spatial dimension, so it can accept variable-scale inputs, and the trained U-shaped network model has a certain generalization ability for the size of the input image;

[0044] Second, the network consists entirely of fully connected operations along the channel dimension, resulting in linear computational complexity with respect to the input image size.

[0045] Third, the network uses the input feature map to calculate the offset, causing ChannelFC to deform in the spatial dimension toward the organ location, effectively reducing the interference caused by a large amount of background information and achieving higher accuracy. Attached Figure Description

[0046] Figure 1 This is an overall architecture diagram of the MLP-like medical image segmentation method applicable to multiple modalities of the present invention;

[0047] Figure 2 This is a schematic diagram of the image embedding layer in this invention;

[0048] Figure 3(a) is a schematic diagram of a standard ChannelFC; Figure 3(b) is a schematic diagram of an improved ChannelFC.

[0049] Figure 4 This is a schematic diagram of the first stage of the basic block proposed in this invention.

[0050] Figure 5(a) is the input image, Figure 5(b) is the ground truth label, and Figure 5(c) is the segmentation result of the method of the present invention. Figures 5(d)-5(i) These are the segmentation results of existing advanced methods, in order: CycleUnet, SwinUnet, TransUnet, AtteUnet, ResUnet, and Unet. Detailed Implementation

[0051] The present invention will be further described in detail below with reference to the accompanying drawings.

[0052] This invention provides a multi-modal MLP-like medical image segmentation method, comprising the following steps:

[0053] Step 1: Collect a medical image segmentation dataset and select a standardization method based on the modality of the medical image segmentation dataset. In this embodiment, abdominal CT images and MR images of the heart of a cardiac patient are used as examples. For CT images, the image pixel values ​​are first cropped to the range of [-125, 275], and then standardized to [0, 1]. For MRI, no cropping is required, and the entire 3D image is directly standardized to [0, 1].

[0054] Step 2: Perform data preprocessing. This step first divides the training set and validation set into an 8:2 ratio. For the training set, the entire 3D image needs to be cut into cross-sectional slices along the vertical direction. The image is then scaled to 224×224 pixels using bilinear interpolation. Finally, some data augmentation methods, such as random flipping, are used to improve the model's generalization ability. The test set, on the other hand, is not processed in much way.

[0055] Step 3: Construct a fully symmetrical U-shaped network for image segmentation. The overall structure includes an image embedding layer, an encoder, a decoder, and a segmentation head.

[0056] The image embedding layer transforms slices into a sequence of embedded feature vectors. This process involves two steps: first, a sliding window divides the slice into several image patches of equal size; second, a linear embedding layer embeds each image patch into a feature vector; and finally, the feature vectors are rearranged according to the relative positions of the image patches to form a feature map. The entire process is illustrated in the following formula.

[0057] IEL(img)=[patch1;…;(patch) n ]·

[0058] in Let be the weight matrix of the linear layer. If the sliding window size is k and the stride is s, then this process is equivalent to a convolution with a kernel size of k, a stride of s, input channels equal to the number of channels in the patch, and output channels of D. Therefore, this example uses a convolution with a kernel size of 7 and a stride of 4 to implement this process. The final output feature map size is...

[0059] Design an improved ChannelFC operator, to This represents the input feature vector sequence. This represents the offset corresponding to each feature vector calculated from the feature map. and These are the weight matrix and biases of the fully connected layer. The improved FC operator formula is as follows:

[0060]

[0061] As can be seen from the formula, the ordinary Channel FC directly samples along the channel dimension of the feature vector and multiplies it with the weight matrix, while the improved FC requires an additional set of offsets to change the sampling position of the Channel FC, where p i (c,O i,c ) represents the sampling position on channel c when calculating the i-th feature vector. To compute p at a lower cost... iWe divide the feature vector into g groups alternately along the channel dimension, with each group sharing a position. This allows the offset O to be shifted from... Compress to p i The specific algorithm formula is as follows:

[0062]

[0063] p i It consists of two parts, Start i (c) indicates the initial position, which is set manually.

[0064] A basic block for feature extraction is constructed, using an improved Channel FC as the basic operator and an MLP-like architecture as the basic block. The construction method includes two stages, as follows:

[0065] The first stage extracts spatial and channel feature information through an improved Channel FC, and the second stage introduces nonlinear transformation to further enhance feature representation capabilities. Each stage includes a LayerNorm and residual connections.

[0066] The first stage involves performing a linear operation on the input feature map to obtain a set of offsets, and then applying these offsets to the feature map. Figure 1 The input is fed into the improved ChannelFC for execution. Since FC operations only accept two-dimensional tensors, the feature map needs to be flattened along the spatial dimensions to convert it into a sequence of feature vectors. This approach employs an asymmetric parallel design, expanding F along both the x and y axes for parallel full-function (FC) operations. Additionally, to fuse channel features, a separate Channel FC branch is added. Finally, the results from all three are summed for output. For offset extraction, this example uses a single linear operation—a convolution with a kernel size of 1. The number of output channels for this operation is determined by the number of groups, g. Due to the presence of two improved Channel FC operations, the number of channels in the output O should be 2g. Figure 4 As shown, odd-numbered channels represent the offset in the x-direction, and even-numbered channels represent the offset in the y-direction. Finally, the offsets with mismatched directions are masked out and input to FC. The number of groups in this instance is set to 3.

[0067] The second stage introduces nonlinear operations using a ChannelFC→GeLU→Channel FC approach to improve feature representation capabilities. Since this second-stage module is similar to the Transformer method, processing the embedded feature vector sequence of image patches, the activation function chosen is GeLU, which performs well in the Transformer. Its approximate expression is as follows:

[0068]

[0069] A fully symmetrical U-shaped network is constructed using basic blocks, including an encoding part, a decoding part, and a segmentation head. The construction method of this example is as follows:

[0070] A downsampling layer is introduced to build a hierarchical encoder. This example uses feature vector fusion for downsampling, which involves concatenating spatially adjacent 2×2 vectors end-to-end and then transforming their channel count through a linear layer. The entire process is equivalent to a convolution with a kernel size of 2 and a stride of 2, which is how this example is implemented. The encoder is constructed by alternating stacks of basic blocks and downsampling layers. During the encoding process, only the downsampling layers change the size and number of channels of the feature map. The encoder structure of this example is: Basic Block × 2 → Downsampling Layer → Basic Block × 2 → Downsampling Layer → Basic Block × 6 → Downsampling Layer → Basic Block × 2. The spatial resolution and number of channels of the feature map between downsampling layers are...

[0071] An upsampling layer is introduced to build a decoder that is completely symmetrical to the encoder. Skip connections are also introduced to fuse feature maps of different scales step by step and restore them to the original image size. This example uses bilinear interpolation for upsampling. The decoder structure is: Upsampling layer → Skip connection → 6 basic blocks → Upsampling layer → Skip connection → 2 basic blocks → Upsampling layer → Skip connection → 2 basic blocks → Upsampling layer.

[0072] The segmentation head is constructed, which is implemented by a convolution and softmax with a kernel size of 1, an input channel number equal to the embedded feature vector dimension D, and an output channel number equal to the number of classes. The convolution operation compresses the number of features of each pixel to be consistent with the number of classes. After the softmax operation, the probability of each class is obtained. The class with the largest value is the class to which the pixel belongs. In this example, the CT dataset has a total of 8 classes and the MR dataset has a total of 4 classes.

[0073] Step 4: Select a suitable loss function and training strategy to optimize the model weight parameters, and use the trained model for inference. The training strategy and inference method for this example are as follows:

[0074] 4.1: After building the U-shaped network model, select an appropriate loss function to optimize the model weight parameters. This example uses the weighted sum of Dice loss and modified cross-entropy as the loss function, and its calculation formula is as follows:

[0075]

[0076] The Adam optimizer is used for backward gradient propagation to optimize the weight parameters, and the learning rate decay strategy is Poly decay, as shown in the following formula.

[0077]

[0078] The hyperparameter β of this example was set to 0.7, the power was set to 0.9, and it was trained for a total of 200 epochs with an initial learning rate of 0.01. The model's weight parameters were initialized using a Gaussian distribution, and the results of pre-training on a large-scale dataset were not transferred.

[0079] 4.2: Validate the trained U-shaped network model on the validation set. During instance inference validation, the U-shaped network model is fed one standardized 3D instance at a time. The U-shaped network model infers layer by layer along the z-axis. Before being fed into the U-shaped network model, slices need to be scaled to the same size as the training set to ensure maximum segmentation accuracy. Finally, the results of each slice are reassembled into a 3D segmentation result for output. The Dice coefficient (DSC) is used as the evaluation criterion, and its calculation method is as follows:

[0080]

[0081] The overall average DSC for this instance is 80.00% on the CT dataset and 89.95% on the MRI dataset.

Claims

1. A multi-modal MLP-like medical image segmentation method, characterized in that, Includes the following steps: Step 1: Collect medical image segmentation datasets and select a standardization method based on the modality of the medical image segmentation datasets; Step 2: Perform data preprocessing on the standardized medical image segmentation dataset to obtain slices; Step 3: Construct a fully symmetrical U-shaped network for image segmentation. The overall structure includes an image embedding layer, an encoder, a decoder, and a segmentation head. The image embedding layer embeds pre-processed slices into feature maps. The encoder consists of basic blocks and downsampling layers, which downsample the feature maps output by the image embedding layer step by step to output a set of feature maps at different scales. The decoder consists of basic blocks and upsampling layers. It upsamples the feature map step by step and makes skip connections with the feature map output at the corresponding position in the encoder in the channel dimension, finally obtaining a feature map with the same output scale as the image embedding layer. The segmentation head reshapes the feature map output by the decoder into slice size and compresses the number of channels to be the same as the number of classes. SoftMax operation is performed on the channel dimension to obtain the class to which each pixel belongs. The basic blocks in the encoder and decoder use the improved ChannelFC as the basic operator. Step 4: Select the loss function and training strategy to optimize the weight parameters of the U-shaped network model, and use the trained U-shaped network for inference. The basic operator of the basic block is an improved Channel FC, with both input and output being a sequence of feature vectors of the same dimension; by This represents the input feature vector sequence. This represents the offset corresponding to each feature vector calculated from the feature map output by the image embedding layer. It is the weight matrix of the improved Channel FC. This is the bias of the improved Channel FC; the formula for the improved Channel FC operator is as follows: ； This represents the number of channels in the input feature vector; c represents the number of channels. This indicates the sampling position on channel c when the Channel FC operator calculates the i-th eigenvector; Indicates the initial position; In order to calculate at a lower cost The feature vector is alternately divided along the channel dimension. Groups, each sharing a location, will have an offset. from Compress to ; It consists of two parts, and the specific algorithm formula is shown below; ； The basic block is constructed using the improved Channel FC as the basic operator. The input and output are feature maps of the same size. The basic block consists of two stages. The first stage extracts spatial and channel feature information through the improved Channel FC. The second stage introduces nonlinear transformation to further enhance the feature representation capability. Each stage includes a LayerNorm and a residual connection. (1): In the first stage, a linear operation is performed on the feature map output by the image embedding layer to obtain a set of offsets. These offsets, along with the feature map, are then input into the improved Channel FC for execution. Since the improved Channel FC operation only accepts two-dimensional tensors, the feature map output by the image embedding layer is flattened along the spatial dimension to convert it into a sequence of feature vectors. A parallel design is adopted to... Full-joint (FC) operations are performed in parallel along both the x and y axes. To further fuse features from the feature map's channel dimension, an additional Channel FC branch is added, and finally, it is combined with... The results of the two-way FC operations are summed and output. (2): In the second stage, nonlinear operations are introduced using the Channel FC → GeLU → Channel FC method to improve feature representation capability; The activation function is GeLU, and its expression is as follows: This represents the value at a specific location in the feature map.

2. The MLP-like medical image segmentation method applicable to multiple modalities according to claim 1, characterized in that, The image embedding layer takes preprocessed slices as input and outputs an embedding feature map. It divides the input slices into equal-sized image blocks using a sliding window, then converts these image blocks into embedding feature vectors. Finally, it rearranges these embedding feature vectors according to the relative positions of the image blocks to form the embedding feature map. The entire process is illustrated by the following formula. ； in, This is the weight matrix of the image embedding layer; when the size of the sliding window is taken as... Step size The process of image embedding layer is equivalent to kernel size of Step size The input channel is The number of channels, the number of output channels is The convolution.

3. The MLP-like medical image segmentation method applicable to multiple modalities according to claim 1, characterized in that, The standardization method selected based on the modality of the medical image segmentation dataset in step 1 is as follows: (1) When the modality of the medical image segmentation dataset is CT image, the pixel values ​​of the CT image are first cropped to the range of [-125, 275], and then standardized to [0, 1]. (2) When the modality of the medical image segmentation dataset is MRI, no cropping is performed, and the entire MRI image is directly standardized to [0,1].

4. The MLP-like medical image segmentation method applicable to multiple modalities according to claim 1, characterized in that, The specific method for preprocessing in step 2 is as follows: The standardized medical image segmentation dataset was divided into training and validation sets in an 8:2 ratio. In the training set, the entire 3D image was cut into cross-sectional slices along the vertical direction. The cross-sectional slices were scaled to 224×224 pixels using bilinear interpolation. Finally, data augmentation methods were used to improve the generalization ability of the U-shaped network model.

5. The MLP-like medical image segmentation method applicable to multiple modalities according to claim 1, characterized in that, In step 3, a fully symmetrical U-shaped network is constructed, with the following structure: image embedding layer → encoder → decoder → segmentation head; the encoder has the following structure: basic block × 2 → downsampling layer → basic block × 2 → downsampling layer → basic block × 6 → downsampling layer → basic block × 2; the decoder has the following structure: upsampling layer → skip connection → basic block × 6 → upsampling layer → skip connection → basic block × 2 → upsampling layer → skip connection → basic block × 2 → upsampling layer.

6. The MLP-like medical image segmentation method applicable to multiple modalities according to claim 1, characterized in that, The specific method for step 4 is as follows: (1) The weighted sum of Dice loss and modified cross-entropy is used as the loss function, and its calculation formula is as follows: ； Labels that indicate authenticity Labels indicating predictions Represents the sign function, when the sample The true category equals hour, Select 1 if the value is 1, otherwise select 0. express weights, Indicates the number of samples; The Adam optimizer is used for backward gradient propagation to optimize the weight parameters, and the learning rate decay strategy is Poly decay, as shown in the following formula. ； Indicates learning rate, Indicates the initial learning rate, Indicates the current iteration number, Indicates the attenuation rate; The weight parameters of the U-shaped model are initialized using a Gaussian distribution, eliminating the need for pre-training on a large-scale dataset. (2) During verification, a standardized 3D instance is input into the U-shaped network model at one time. The U-shaped network model reasones layer by layer along the z-axis. Before the slices are fed into the U-shaped network model, they are scaled to the same size as the training set to ensure the maximum segmentation accuracy. Finally, the results of each slice are reassembled into the 3D segmentation results and output. The Dice coefficient (DSC) and Hausdorff distance (HD) are used as evaluation criteria.

Images