A method for three-dimensional medical image segmentation fusing multi-view information

Abstract

The application discloses a three-dimensional medical image segmentation method fusing multi-view information, mainly comprising the following steps: (1) reconstructing two-dimensional slice images from sagittal, coronal and transverse three view directions for a three-dimensional medical image to be detected; (2) adopting a U-shaped 2D convolution network based on a hollow space pyramid convolution to segment the two-dimensional slices in different view directions; and (3) adopting a light 3D convolution network to fuse the segmentation results in different view directions to obtain accurate three-dimensional segmentation results. The application can effectively overcome the problems that a 2D network cannot extract three-dimensional space features and a 3D network has a large memory overhead, and can obtain accurate segmentation results in the case of a light network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a three-dimensional medical image segmentation method fusing multi-view information. BACKGROUND

[0002] In the field of medical image processing and analysis, the segmentation of medical images is an important and challenging task. Accurate segmentation of organs and tissues in medical images is an important prerequisite for radiotherapy, clinical evaluation, surgical navigation, computer-aided detection, and medical data visualization. With the development of imaging equipment and technology, three-dimensional medical images such as CT and MRI sequence images are widely used in the clinical diagnosis and treatment of various diseases. The number of slices used in three-dimensional imaging is huge, and manual segmentation is time-consuming and labor-intensive, and the segmentation quality is heavily dependent on the experience and professional level of the operator. Therefore, it is urgent to develop a method that can automatically and effectively segment three-dimensional medical images.

[0003] Due to the complex geometric shapes and positional relationships between different anatomical structures, the diverse morphologies between normal and diseased tissues, automatic segmentation of three-dimensional medical images is difficult, and traditional methods are usually ineffective. In recent years, with the breakthrough progress of deep learning in the field of medical image target recognition and segmentation, domestic and foreign scholars have also proposed a variety of deep learning-based automatic segmentation methods for three-dimensional medical images. Considering the space and time efficiency, most of the current deep learning methods are based on two-dimensional networks, which sequentially segment each two-dimensional slice image by using two-dimensional networks to achieve the purpose of segmenting three-dimensional images. Two-dimensional networks are difficult to extract three-dimensional spatial information, and the segmentation accuracy is usually low. In order to extract the three-dimensional spatial features of the image, some experts and scholars have also proposed segmentation methods based on three-dimensional networks. However, due to the limitation of computing resources, this type of method usually cannot directly process three-dimensional medical images, and needs to pre-process them by downsampling to reduce the size or cutting them into multiple small three-dimensional data sub-blocks, which will result in the loss of some image detail information, and thus affect the network segmentation accuracy. SUMMARY

[0004] In view of the shortcomings and deficiencies of the prior art, the present application provides a three-dimensional medical image segmentation method fusing multi-view information by combining 2D deep convolution and 3D lightweight network, which can effectively overcome the problems of 2D network unable to extract three-dimensional spatial features and 3D network large memory overhead.

[0005] A three-dimensional medical image segmentation method fusing multi-view information, comprising the following steps:

[0006] (1) Establishing original training data sets A and B containing original three-dimensional medical images and target region manual segmentation results;

[0007] (2) Construct a U-shaped 2D convolution network based on a cavity space pyramid convolution, denoted as ASPP-UNet, which specifically includes:

[0008] (2-a) Adopt a U-shaped network as the backbone network, which contains three encoding layers, two skip connections, a cavity space pyramid convolution layer, three decoding layers and a 1x1 convolution layer, wherein: the output of the first encoding layer is not only the input of the second encoding layer, but also connected to the second decoding layer through the first skip connection as the input of the decoding layer; the output of the second encoding layer is not only the input of the third encoding layer, but also connected to the first decoding layer through the second skip connection as the input of the decoding layer; the output of the third encoding layer is the input of the cavity space pyramid convolution layer, and the output of the cavity space pyramid convolution layer is the input of the first decoding layer; in addition, the output of the previous decoding layer is the input of the next decoding layer; in order to obtain the segmentation result, the last decoding layer is connected with a 1x1 convolution layer, wherein the output of the last decoding layer is the input of the 1x1 convolution layer, and the output of the 1x1 convolution layer is the probability of each pixel belonging to the target, and the segmentation result can be obtained by introducing a threshold value ε1; the ε1 is preferably a constant of 0.3-0.7;

[0009] (2-b) In the backbone network described in step (2-a), each encoding layer is composed of two 2D convolution module connections, i.e. 2D double convolution module, wherein each 2D convolution module contains a 3x3 convolution layer, a batch normalization layer and a Relu activation layer; in order to downsample the image, in the second and third encoding layers, a 2x2 maximum pooling layer is added at the end of the 2D double convolution module;

[0010] (2-c) In the backbone network described in step (2-a), the cavity space pyramid convolution layer specifically includes: using n 3x3 convolution kernels with different sampling radii {r v |v=1,2,...,n} to perform cavity convolution on the input feature map respectively, and the cavity convolution results are spliced as the output of the cavity space pyramid convolution layer, wherein n is a natural number greater than 1; in order to expand the receptive field of the convolution kernel and obtain multi-scale context information, the sampling radius is set as r v =k×v+1, wherein k is a natural number greater than 0; the n is preferably a natural number of 2-10, and the k is preferably a natural number of 1-8;

[0011] (2-d) In the backbone network described in step (2-a), the first and second decoding layers are both composed of a 2D double convolution module described in step (2-b), a 2x2 deconvolution layer and a concatenation operation, and the third decoding layer is only composed of a 2D double convolution module described in step (2-b), wherein: the input of the 2D double convolution module in the first decoding layer is the output of the empty spatial pyramid convolution layer, and thereafter, the input of the 2D double convolution module in the next decoding layer is the output of the previous decoding layer; the concatenation operation in the first decoding layer is used to concatenate the deconvolution result in the decoding layer and the output of the second encoding layer, and the concatenation result is taken as the output of the decoding layer; the concatenation operation in the second decoding layer is used to concatenate the deconvolution result in the decoding layer and the output of the first encoding layer, and the concatenation result is taken as the output of the decoding layer;

[0012] (3) Construct a lightweight 3D convolution network, denoted as LW-3DNet, which involves three inputs and one output. The specific structure of the network includes: first, three 3D double convolution modules are used to convolve the three inputs respectively, then a concatenation operation is used to concatenate the convolution results, and a 3D double convolution module is used to convolve the concatenation results to obtain a feature map F, finally, in order to obtain the segmentation result, a 1x1x1 convolution layer is used to convolve the feature map F, and the output of the 1x1x1 convolution layer is the probability of each voxel belonging to the target, and by introducing a threshold ε2, the segmentation result can be obtained; the 3D double convolution module described in step (3) is composed of two 3D convolution modules, wherein each 3D convolution module contains a convolution layer with a size of 3x3x3, a batch normalization layer and a Relu activation layer;

[0013] (4) Use ASPP-UNet to train multiple network models that can be used to segment two-dimensional slices in different view directions. The specific steps include: for each three-dimensional medical image in the training data set A, first, two-dimensional slices in the sagittal, coronal and transverse view directions are reconstructed respectively to obtain two-dimensional slices in different view directions, denoted as and Then, the two-dimensional slices in the sagittal view direction and the corresponding two-dimensional slice manual segmentation results in the training data set are input into the ASPP-UNet network for training to obtain a network model ASPP-UNet that can be used to segment two-dimensional slices in the sagittal view direction X The two-dimensional slices in the coronal view direction and the corresponding two-dimensional slice manual segmentation results in the training data set are input into the ASPP-UNet network for training to obtain a network model ASPP-UNet that can be used to segment two-dimensional slices in the coronal view direction Y The two-dimensional slices in the transverse view direction and its corresponding two-dimensional slice manual segmentation result input ASPP-UNet network for training, to obtain a network model ASPP-UNet that can be used for segmentation of transverse two-dimensional slices Z ; training to obtain network model ASPP-UNet X , ASPP-UNet Y and ASPP-UNet Z , the loss function is preferably a hybrid loss function based on cross-entropy and Dice, which is defined as follows:

[0014] l = l c + η·l d

[0015]

[0016]

[0017] wherein, l c and l d respectively represent the cross-entropy and Dice loss, η is a weight parameter, preferably a constant of 0.5-2, g w represents the expert manual segmentation result of the wth pixel in the image, where the background is marked as 0 and the target is marked as 1, p w represents the probability of the network model predicting that the wth pixel belongs to the target, and T is the number of pixels in the image.

[0018] (5) Using LW-3DNet, a network model that can be used to fuse segmentation results of different view directions is trained, which specifically includes:

[0019] (5-a) Constructing the training data set C of LW-3DNet network, specifically including: first, for each three-dimensional medical image in the original training data set B, reconstruct two-dimensional slices from the sagittal, coronal and transverse three view directions, respectively, to obtain two-dimensional slices in different view directions, respectively denoted as and Then, respectively input and to the trained network model ASPP-UNet X , ASPP-UNet Y and ASPP-UNet Z for testing, to obtain two-dimensional slice segmentation results S X , S Y and S Z ; finally, input S X , S Y and S ZThe manual segmentation results of 3D medical images in training dataset B are used as labels to construct the training dataset C of the LW-3DNet network as input for training the network.

[0020] (5-b) Input the training dataset C into the LW-3Dnet network for training. The Dice loss function is preferred to obtain the trained network model LW-3DNet. F ;

[0021] (6) For the three-dimensional medical image to be detected, two-dimensional slice reconstruction is first performed from the sagittal, coronal and transverse views to obtain two-dimensional slices of the three-dimensional medical image in different view directions, denoted as T. X T Y and T Z Then, T respectively X T Y and T Z Input into the pre-trained network model ASPP-UNet X ASPP-UNet Y and ASPP-UNet Z Tests were conducted to obtain 2D slice segmentation results F from different view directions. X F Y and F Z Finally, F X F Y and F Z Input LW-3DNet F The network model is tested to obtain the final segmentation result. Attached Figure Description

[0022] Figure 1 ASPP-UNet network structure diagram

[0023] Figure 2 Schematic diagram of the convolutional layer structure of the hollow space pyramid

[0024] Figure 3 LW-3DNet network structure diagram

[0025] Figure 4 Overall segmentation process diagram of the embodiment of the present invention

[0026] Figure 5 Examples of spleen segmentation results according to embodiments of the present invention, wherein 5(a) to (c) are examples of two-dimensional slice results of the three-dimensional segmentation results in the sagittal, coronal and transverse view directions respectively. Detailed Implementation

[0027] Example 1

[0028] A three-dimensional medical image segmentation method of fusing multi-view information, the specific implementation steps are as follows:

[0029] (1) Obtain the original three-dimensional medical image of the patient and the manual segmentation result of the target region in the image from the medical image public data set to construct training data sets A and B;

[0030] (2) Construct a U-shaped 2D convolution network based on a hollow space pyramid convolution, denoted as ASPP-UNet, the specific structure is as shown in Figure 1 , which specifically comprises:

[0031] (2-a) A U-shaped network is used as the backbone network, which contains three encoding layers, two jump connections, a hollow space pyramid convolution layer, three decoding layers and a 1x1 convolution layer, wherein: the output of the first encoding layer is not only used as the input of the second encoding layer, but also connected through the first jump connection with the second decoding layer as the input of the decoding layer; the output of the second encoding layer is not only used as the input of the third encoding layer, but also connected through the second jump connection with the first decoding layer as the input of the decoding layer; the output of the third encoding layer is used as the input of the hollow space pyramid convolution layer, and the output of the hollow space pyramid convolution layer is used as the input of the first decoding layer; in addition, the output of the previous decoding layer is used as the input of the next decoding layer; in order to obtain the segmentation result, the last decoding layer is connected with a 1x1 convolution layer, wherein the output of the last decoding layer is used as the input of the 1x1 convolution layer, and the output of the 1x1 convolution layer is the probability of each pixel belonging to the target, and the threshold value ε1 is introduced to obtain the segmentation result; the embodiment preferably ε1=0.5;

[0032] (2-b) In the backbone network described in step (2-a), each encoding layer is composed of two 2D convolution module connections, i.e. 2D double convolution module, wherein each 2D convolution module contains a 3x3 convolution layer, a batch normalization layer and a Relu activation layer; in order to downsample the image, a 2x2 maximum pooling layer is added at the end of the 2D double convolution module in the second and third encoding layers;

[0033] (2-c) In the backbone network described in step (2-a), the hollow space pyramid convolution layer has the structure as shown in Figure 2 , which specifically comprises: n 3x3 convolution kernels with different sampling radii {r v |v=1,2,...,n} are used to perform hollow convolution on the input feature map respectively, and the hollow space convolution results are spliced as the output of the hollow pyramid convolution layer, wherein n is a natural number greater than 1; in order to expand the receptive field of the convolution kernel and obtain multi-scale context information, the sampling radius is set to r v= k x v + 1, wherein k is a natural number greater than 0; preferably, n = 5 and k = 2 in this embodiment;

[0034] (2-d) In the backbone network described in step (2-a), the first and second decoding layers are each composed of a 2D double convolution module described in step (2-b), a 2x2 deconvolution layer and a concatenation operation, and the third decoding layer is composed of only a 2D double convolution module described in step (2-b), wherein: the input of the 2D double convolution module in the first decoding layer is the output of the empty spatial pyramid convolution layer, and thereafter, the input of the 2D double convolution module in the next decoding layer is the output of the previous decoding layer; the concatenation operation in the first decoding layer is used to concatenate the deconvolution result in the decoding layer and the output of the second encoding layer, and the concatenation result is taken as the output of the decoding layer; the concatenation operation in the second decoding layer is used to concatenate the deconvolution result in the decoding layer and the output of the first encoding layer, and the concatenation result is taken as the output of the decoding layer;

[0035] (3) Construct a lightweight 3D convolution network, denoted as LW-3DNet, and the network structure is as shown in Figure 3 The network involves three inputs and one output, and the specific structure of the network includes: first, three 3D double convolution modules are used to convolve the three inputs, then a concatenation operation is used to concatenate the convolution results, and a 3D double convolution module is used to convolve the concatenation results to obtain a feature map F, finally, in order to obtain the segmentation result, a 1x1x1 convolution layer is used to convolve the feature map F, and the output of the 1x1x1 convolution layer is the probability of each voxel belonging to the target, and the threshold ε2 is introduced to obtain the segmentation result; the 3D double convolution module described in step (3) is composed of two 3D convolution modules, wherein each 3D convolution module includes a convolution layer with a size of 3x3x3, a batch normalization layer and a Relu activation layer; preferably, ε2 = 0.5 in this embodiment

[0036] (4) Use ASPP-UNet to train multiple network models that can be used to segment two-dimensional slices in different view directions, and the specific steps include: for each three-dimensional image in the training data set A, first, two-dimensional slices in different view directions are reconstructed from the sagittal plane, the coronal plane and the transverse plane, and the two-dimensional slices in different view directions are denoted as and Then, the two-dimensional slices in different view directions obtained from the training data set and the corresponding two-dimensional slice manual segmentation results are input into the ASPP-UNet network for training, and three network models that can be used to segment two-dimensional slices in different view directions are obtained, which are denoted as ASPP-UNet X , ASPP-UNet Y and ASPP-UNet ZDuring the training process, the loss function is preferably a hybrid loss function based on cross-entropy and Dice, which is defined as follows:

[0037] l = l c + η·l d

[0038]

[0039]

[0040] wherein, l c and l d respectively represent the cross-entropy and Dice loss, η is a weight parameter, and the embodiment preferably has η = 1; g w represents the expert manual segmentation result of the wthpixel in the image, wherein the background is marked as 0 and the target is marked as 1, p w represents the probability of the network model predicting that the wthpixel belongs to the target, and T is the number of pixels in the image.

[0041] (5) Using the LW-3DNet, a network model that can be used to fuse segmentation results of different view directions is trained, which specifically includes:

[0042] (5-a) Constructing the training data set C of the LW-3DNet network, which specifically includes: first, for each three-dimensional image in the original training data set B, reconstructing two-dimensional slices from the sagittal plane, coronal plane and transverse plane, respectively, to obtain two-dimensional slices in different view directions, which are respectively denoted as and Then, respectively input and to the trained network models ASPP-UNet X , ASPP-UNet Y and ASPP-UNet Z to test, to obtain two-dimensional slice segmentation results S X , S Y and S Z ; finally, input the network prediction S X , S Y and S Z to the training of the LW-3DNet network, and the three-dimensional manual segmentation result in the training data set B is used as the label to construct the training data set C of the LW-3DNet network.

[0043] (5-b) Input the training data set C to the LW-3Dnet network for training, and the loss function is preferably a Dice loss function, to obtain the trained network model LW-3DNet F ;

[0044] (6) For the three-dimensional image to be detected, first, two-dimensional slice reconstruction is performed from the sagittal plane, the coronal plane and the transverse plane three view directions respectively, to obtain two-dimensional slices of the three-dimensional image in different view directions, denoted as T X , T Y and T Z respectively; then, T X , T Y and T Z are input into the trained network models ASPP-UNet X , ASPP-UNet Y and ASPP-UNet Z respectively for testing, to obtain two-dimensional slice segmentation results F X , F Y and F Z in different view directions; finally, F X , F Y and F Z are input into the LW-3DNet F network model for testing, to obtain the final segmentation result. Figure 4 The overall segmentation process schematic diagram of the embodiment of the application is given.

[0045] Example 2

[0046] The method in Example 1 was used to perform spleen organ segmentation on the AbdomenCT-1K public data set.

[0047] In this example, 150 abdominal CT original sequence images and their corresponding manually segmented spleen region results were randomly selected from the AbdomenCT-1K public database, 50 cases, i.e. 50 abdominal CT original sequence images and their corresponding manually segmented spleen region results, were selected as the training data set A, 50 cases were selected as the training data set B, and the remaining 50 cases were selected as the test data set. The average Dice coefficient was used as the evaluation standard in the experiment, and the Dice coefficient of the segmentation method of the application reached 95.3% in the spleen segmentation task.

[0048] Figure 5 (a)-(c) give examples of two-dimensional slice results of the three-dimensional spleen segmentation result in the sagittal plane, the coronal plane and the transverse plane three view directions respectively, wherein the region shown by the white closed curve is the spleen segmentation result obtained in this example, and it can be seen that the spleen region in the CT image in different view directions is effectively segmented.

[0049] The above only describes the preferred embodiments of the application and should not be used to limit the application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application should be included in the protection scope of the application.

Claims

1. A three-dimensional medical image segmentation method integrating multi-view information, characterized in that, Includes the following steps: (1) Establish original training datasets A and B containing original 3D medical images and manual segmentation results of target regions; (2) Construct a U-shaped 2D convolutional network based on dilated spatial pyramid convolution, denoted as ASPP-UNet, which specifically includes: (2-a) A U-shaped network is used as the backbone network, which includes three coding layers, two skip connections, one dilated spatial pyramid convolutional layer, three decoding layers, and one 1×1 convolutional layer. Specifically: the output of the first coding layer serves as the input to the second coding layer and is also connected to the second decoding layer via the first skip connection; the output of the second coding layer serves as the input to the third coding layer and is also connected to the first decoding layer via the second skip connection; the output of the third coding layer serves as the input to the dilated spatial pyramid convolutional layer, and the output of the dilated spatial pyramid convolutional layer serves as the input to the first decoding layer; furthermore, the output of each previous decoding layer serves as the input to the next decoding layer. To obtain the segmentation result, the last decoding layer is connected to a 1×1 convolutional layer, where the output of the last decoding layer serves as the input to the 1×1 convolutional layer. The output of the 1×1 convolutional layer represents the probability that each pixel belongs to the target. By introducing a threshold ε1, the segmentation result can be obtained. (2-b) In the backbone network described in step (2-a), each coding layer consists of two 2D convolutional modules connected together, namely a 2D dual convolutional module. Each 2D convolutional module contains a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer. In order to downsample the image, a 2×2 max pooling layer is added at the end of the 2D dual convolutional module in the second and third coding layers. (2-c) In the backbone network described in step (2-a), the hollow spatial pyramid convolutional layer specifically includes: using n layers with different sampling radii {r v A 3×3 convolutional kernel of |v=1,2,...,n} performs dilated convolution on the input feature map, and the results of the dilated convolution are concatenated as the output of the dilated spatial pyramid convolutional layer, where n is a natural number greater than 1; in order to expand the receptive field of the convolutional kernel and obtain multi-scale contextual information, the sampling radius is set to r. v = k × v + 1, where k is a natural number greater than 0; (2-d) In the backbone network described in step (2-a), the first and second decoding layers are each composed of a 2D dual convolutional module described in step (2-b), a 2×2 deconvolutional layer, and a concatenation operation. The third decoding layer consists of only one 2D dual convolutional module described in step (2-b). Specifically: the input of the 2D dual convolutional module in the first decoding layer is the output of the dilated spatial pyramid convolutional layer; thereafter, the input of the 2D dual convolutional module in the next decoding layer is the output of the previous decoding layer; the concatenation operation in the first decoding layer is used to concatenate the deconvolution result of that decoding layer with the output of the second encoding layer, and the concatenation result serves as the output of that decoding layer; the concatenation operation in the second decoding layer is used to concatenate the deconvolution result of that decoding layer with the output of the first encoding layer, and the concatenation result serves as the output of that decoding layer. (3) Construct a lightweight 3D convolutional network, denoted as LW-3DNet. This network involves three inputs and one output. The specific structure of the network includes: First, three 3D dual convolutional modules are used to convolve the three inputs respectively. Then, the convolution results are concatenated by a concatenation operation. A 3D dual convolutional module is used to convolve the concatenated results to obtain the feature map F. Finally, in order to obtain the segmentation result, a 1×1×1 convolutional layer is used to convolve the feature map F. The output of the 1×1×1 convolutional layer is the probability that each voxel belongs to the target. By introducing a threshold ε2, the segmentation result can be obtained. The 3D dual convolutional modules mentioned in step (3) are all composed of two 3D convolutional modules connected together. Each 3D convolutional module contains a 3×3×3 convolutional layer, a batch normalization layer and a ReLU activation layer. (4) Using ASPP-UNet, train multiple network models that can be used to segment two-dimensional slices in different view directions. The specific steps include: for each three-dimensional medical image in the training dataset A, first reconstruct two-dimensional slices from the three view directions of sagittal, coronal and transverse planes respectively, to obtain two-dimensional slices in different view directions, denoted as . and Then, the two-dimensional slices of the sagittal view direction obtained from the training dataset are... The manual segmentation results of the corresponding 2D slices are input into the ASPP-UNet network for training, resulting in the ASPP-UNet network model that can be used to segment sagittal 2D slices. X Two-dimensional slices of the coronal view direction obtained from the training dataset The manual segmentation results of the corresponding 2D slices are input into the ASPP-UNet network for training, resulting in the ASPP-UNet network model that can be used to segment coronal 2D slices. Y Two-dimensional slices obtained from the cross-sectional view direction in the training dataset The manual segmentation results of the corresponding 2D slices are input into the ASPP-UNet network for training, resulting in the ASPP-UNet network model that can be used to segment cross-sectional 2D slices. Z ; (5) Using LW-3DNet, train a network model that can be used to fuse segmentation results from different view orientations, specifically including: (5-a) Constructing the training dataset C for the LW-3DNet network specifically includes: First, for each three-dimensional medical image in the original training dataset B, reconstructing two-dimensional slices from the sagittal, coronal, and transverse view directions respectively, obtaining two-dimensional slices in different view directions, denoted as... and Then, respectively and Input into the pre-trained network model ASPP-UNet X ASPP-UNet Y and ASPP-UNet Z The test was conducted to obtain the 2D slice segmentation results S from different view directions. X S Y and S Z Finally, the network predicts S X S Y and S Z The manual segmentation results of 3D medical images in training dataset B are used as labels to construct the training dataset C of the LW-3DNet network as input for training the network. (5-b) Input the training dataset C into the LW-3Dnet network for training to obtain the trained network model LW-3DNet. F ; (6) For the three-dimensional medical image to be detected, two-dimensional slice reconstruction is first performed from the sagittal, coronal and transverse views to obtain two-dimensional slices of the three-dimensional medical image in different view directions, denoted as T. X T Y and T Z Then, T respectively X T Y and T Z Input into the pre-trained network model ASPP-UNet X ASPP-UNet Y and ASPP-UNet Z Tests were conducted to obtain 2D slice segmentation results F from different view directions. X F Y and F Z Finally, F X F Y and F Z Input LW-3DNet F The network model is tested to obtain the final segmentation result.

2. The three-dimensional medical image segmentation method integrating multi-view information as described in claim 1, characterized in that: In step (4), the ASPP-UNet network model is trained and obtained. X ASPP-UNet Y and ASPP-UNet Z At this time, the loss function is a hybrid loss function based on cross-entropy and Dice, specifically defined as follows: l=l c +η·l d Among them, l c and l d Let g represent the cross-entropy and Dice loss, respectively, where η is the weight parameter and g is the weight parameter. w This represents the expert-manual segmentation result for the w-th pixel in the image, where the background is labeled 0 and the target is labeled 1. w This represents the probability that the network model predicts the w-th pixel belongs to the target, where T is the number of pixels in the image.

3. The three-dimensional medical image segmentation method integrating multi-view information as described in claim 1, characterized in that: In step (5-b), the network model LW-3DNet is trained and obtained. F When the loss function is , the loss function is the Dice loss function.

4. The three-dimensional medical image segmentation method integrating multi-view information as described in claim 1, characterized in that: ε1 is a constant of 0.3 to 0.7, ε2 is a constant of 0.3 to 0.7, n is a natural number of 2 to 10, k is a natural number of 1 to 8, and η is a constant of 0.5 to 2.

Images