A two-stage 3D image segmentation method based on dimensionality decomposition attention

Abstract

This invention proposes a two-stage 3D image segmentation method based on dimensionality decomposition attention, effectively overcoming the segmentation performance degradation caused by the similar intensity of the target region and adjacent tissues. The first stage of this method uses a 3D U-shaped network to locate the region of interest, reducing interference from irrelevant tissues in the image and providing input for the second stage. The second stage adds dimensionality decomposition spatial attention and dimensionality decomposition channel attention to the 3D U-shaped network to achieve accurate target segmentation. Both dimensionality decomposition spatial attention and dimensionality decomposition channel attention decompose the features output by the encoder into three one-dimensional features (length, width, and depth), generating attention weights. The weights of the two attentions are multiplied by the features output by the encoder as input to the decoder, generating the final segmentation result. This invention can effectively utilize all the information in a 3D image, enhance the weights of the region of interest, and improve segmentation performance.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more particularly to a two-stage three-dimensional image segmentation method and system based on dimensionality decomposition attention. Background Technology

[0002] With the development of deep learning, convolutional neural networks (CNNs), with their ability to automatically learn deeper and more discriminative features from samples, have been widely applied to image segmentation tasks. Researchers have proposed a series of two-dimensional convolutional models for image segmentation, demonstrating excellent performance. However, three-dimensional images can store more information than two-dimensional images, making them more effective in real-world applications. Therefore, researchers have proposed three-dimensional convolutional models specifically for three-dimensional images, achieving better results than two-dimensional models.

[0003] Because target regions in 3D images often have similar intensity to adjacent tissues, 2D and 3D models are prone to misclassification during segmentation, severely impacting segmentation performance. In 3D images, target regions typically occupy only a small portion, meaning most of the data is useless. Excessive useless data places a significant burden on the computer during training. Therefore, this invention introduces a two-stage strategy: the first stage extracts the region of interest, removing irrelevant regions from the image, significantly reducing the computational load in the second stage.

[0004] To mitigate the performance degradation caused by the similarity in intensity between the target region and the background, researchers have introduced attention methods from natural language processing. This method allows the model to focus its attention on important information while ignoring less important information, thus improving segmentation performance. With the rapid development of deep learning, a series of attention methods have been proposed and have shown good results. However, most of these methods are designed for two-dimensional images and do not consider the situation of three-dimensional images, making them difficult to use directly in three-dimensional models. Therefore, this invention proposes a dimensionality decomposition attention module for three-dimensional image segmentation, improving segmentation accuracy.

[0005] To achieve accurate and efficient 3D image segmentation, this invention proposes a two-stage U-Net method based on dimensionality decomposition attention. First, a 3D U-Net coarse segmentation network is used for coarse localization to extract regions of interest (ROIs). In the second stage, the ROIs extracted in the first stage are used for segmentation, and dimensionality decomposition spatial attention and dimensionality decomposition channel attention are added to improve the final segmentation results. Summary of the Invention

[0006] The purpose of this invention is to provide a two-stage three-dimensional image segmentation method based on dimensionality decomposition attention, which utilizes all the information in the three-dimensional image to enhance the weight of the region of interest and solve the problem of decreased segmentation performance caused by the similar intensity of the target region and adjacent tissues.

[0007] To achieve the above objectives, this invention provides a two-stage 3D image segmentation method based on dimensionality decomposition attention, characterized by using a two-stage strategy and dimensionality decomposition attention to segment 3D images; the method includes the following steps:

[0008] Step S1: Preprocess and coarsely segment the input image to obtain the region of interest through cropping;

[0009] Step S2: Downsample the obtained region of interest to obtain depth features, and feed the depth features into the dimensionality decomposition spatial attention to obtain spatial weights;

[0010] Step S3: Feed the deep features into the dimensionality decomposition channel attention to obtain the channel weights;

[0011] Step S4: Multiply the depth features with the spatial weights and channel weights to obtain the final features, and then feed the final features into the upsampling network to obtain the segmentation result.

[0012] Preferably, the preprocessing and coarse segmentation of the input image to obtain the region of interest through cropping includes:

[0013] The input image is cropped and resized to a uniform size to remove the influence of image edge information;

[0014] A preprocessed image of uniform size is fed into a coarse segmentation network to obtain a coarse segmentation result. Then, erosion and maximum connectivity location operations are performed on the coarse segmentation result. The original image and the ground truth label are cropped using rectangular blocks of fixed size to obtain the region of interest.

[0015] Preferably, after downsampling the obtained region of interest to obtain depth features, the method further includes feeding the obtained depth features into a dimensionality decomposition spatial attention module to generate spatial features.

[0016] Preferably, feeding the depth features into the dimensionality decomposition spatial attention to obtain spatial weights includes:

[0017] The dimensionality decomposition spatial attention first generates spatial features X from the input depth features through pooling operations; spatial features X are then decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1), and (1, 1, D). H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively;

[0018] The three obtained one-dimensional directional features are fused in the order of H, W, and then D to generate spatial weights, including: first fusion of the height directional feature OH With width direction feature O W The feature is unified to a size of (H, W, 1) through a dilation operation, and then a 3D convolution operation with a kernel size of 3×3×3 is performed to fuse the features in the height and width directions, resulting in the lateral intermediate feature f1, as shown in the formula f1=δ1(F1([t(O H ),t(O W As shown in the diagram, [,] represents connection along the channel, t represents expansion operation, and δ1 is the nonlinear activation function ReLU; then, the lateral intermediate feature f1 and the depth direction feature O are combined. D The size is unified to (H, W, D) through a dilation operation, and then fused using a 3×3×3 three-dimensional convolution with a kernel size of 3×3×3. The spatial weights F are obtained after passing through the sigmoid activation function. s The spatial weight F s =δ2(F2([t(f1),t(O) D )])), where δ2 is the nonlinear activation function Sigmoid.

[0019] Preferably, the step of feeding the depth features into the dimensionality decomposition channel attention to obtain channel weights includes: the dimensionality decomposition channel attention first generates spatial features X from the input depth features through pooling operations; the spatial features X are decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1) and (1, 1, D) respectively. H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively;

[0020] The input depth features are processed through pooling and convolution operations to generate channel features f. c Among them, the channel feature f c =δ3(F3(avg(X))), where X is the input depth feature, avg represents global average pooling, F3 is a 1×1×1 3D convolution with a kernel size of δ3, and ReLU is the non-linear activation function. c ∈R 3c×1×1×1 The resulting set of dimensional channel features;

[0021] The resulting dimensional channel feature set is split into three intermediate features and then combined with the one-dimensional directional feature O. H O W and O D Each channel is concatenated and fed into a 1×1×1 3D convolutional layer for fusion, yielding channel features in three dimensions: H, W, and D. The feature f in the H dimension is... H ∈R c×1×1×1 From formula f H =δH (F H ([O H ,s(f c The result is obtained as follows: [,] represents splicing along the channel dimension, s is the splitting operation, and F... H For a 1×1×1 convolution, δ H ReLU is a non-linear activation function.

[0022] The features of the W and D dimensions are given by formula f W =δ W (F W ([O W ,s(f c )])) and f D =δ D (F D ([O D ,s(f c )]))get;

[0023] The obtained three-dimensional channel features are concatenated by channel, and then processed using formula F. c =δ4(F4([f H ,f W ,f Z Generate the final channel weights F. c , where F4 is a 3D convolution with a kernel size of 1×1×1, and δ4 is a non-linear activation function Sigmoid.

[0024] Preferably, the step of multiplying the depth features with spatial weights and channel weights to obtain the final features, and then feeding the final features into an upsampling network to obtain the segmentation result, includes:

[0025] Depth features F and spatial weights F s and channel weight F c According to formula F f =F×F s ×F c The features are fused together to obtain the final features, and then fed into an upsampling network to obtain the segmentation result.

[0026] According to another aspect of the present invention, a two-stage 3D image segmentation system based on dimensionality decomposition attention is provided, characterized in that a two-stage strategy and dimensionality decomposition attention are used to segment 3D images; the system includes:

[0027] A cropping device is used to preprocess and coarsely segment the input image in order to obtain the region of interest through cropping;

[0028] A sampling device is used to downsample the obtained region of interest to obtain depth features, and to feed the depth features into the dimensionality decomposition spatial attention to obtain spatial weights.

[0029] A device for obtaining deep features by feeding them into a dimensionality-decomposition channel attention to obtain channel weights;

[0030] The segmentation device is used to multiply the depth features with spatial weights and channel weights to obtain the final features, and then feed the final features into the upsampling network to obtain the segmentation result.

[0031] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the disclosure of the present invention. Attached Figure Description

[0032] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a structural diagram of a two-stage method based on dimensionality decomposition attention provided in an embodiment of the present invention.

[0035] Figure 2 A dimensional decomposition spatial attention structure diagram is provided for the embodiments disclosed in this invention;

[0036] Figure 3 This invention provides a dimension decomposition channel attention structure diagram for embodiments thereof;

[0037] Figure 4 A flowchart of a two-stage 3D image segmentation method based on dimensionality decomposition attention provided in the embodiments of the present invention; and

[0038] Figure 5 This is a schematic diagram of the structure of a two-stage three-dimensional image segmentation system based on dimensionality decomposition attention, as provided in an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without innovative effort are within the scope of protection of the present invention.

[0040] The purpose of this invention is to provide a two-stage three-dimensional image segmentation method based on dimensionality decomposition attention, which utilizes all the information in the three-dimensional image to enhance the weight of the region of interest and solve the problem of decreased segmentation performance caused by the similar intensity of the target region and adjacent tissues.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Figure 1 This is a structural diagram of a two-stage method based on dimensionality decomposition attention provided in an embodiment of the present invention. Figure 2 This invention discloses a dimensional decomposition spatial attention structure diagram for its embodiments. Figure 3 This invention provides a dimensionality decomposition channel attention structure diagram for embodiments disclosed herein. See also... Figure 1 , Figure 2 , Figure 3 The specific steps of the two-stage 3D image segmentation method based on dimensionality decomposition attention provided by this invention are as follows:

[0043] Step S1: First, the input image is cropped at the edges and resized to a uniform size to remove the influence of image edge information. Then, the preprocessed image of uniform size is fed into a coarse segmentation network to obtain coarse segmentation results. Finally, erosion and maximum connected component localization operations are performed on the coarse segmentation results, using fixed-size rectangular blocks to crop the original image and the ground truth label to obtain the region of interest.

[0044] Step S2: First, the region of interest obtained in Step S1 is fed into a downsampling network to obtain depth features. Then, the obtained depth features are fed into the dimensionality decomposition spatial attention module to generate spatial features. The dimensionality decomposition spatial attention module includes dimensionality decomposition and dimensionality decomposition spatial attention generation parts.

[0045] The dimensionality decomposition part first follows the traditional spatial attention method, performing pooling operations along the channel directions to generate spatial features, compressing channel information for subsequent dimensionality decomposition. For the generated spatial features, three pooling kernels of sizes (H, 1, 1), (1, W, 1), and (1, 1, D) are used to perform feature decomposition along the three dimensions of the 3D image to obtain the height direction feature O. H Width direction feature O W and depth direction features O D Among them, the height-direction feature O H Output O at h h It is given by formula (1).

[0046]

[0047] Similarly, the width-direction feature O W The output O at point w w and the output O at depth d d The results are given by formulas (2) and (3) respectively.

[0048]

[0049]

[0050] The dimensionality decomposition spatial attention generation part first extracts the height direction features O H With width direction O W After being expanded to the same size, they are connected and then fed into a 3×3×3 convolution transformation F1 for feature fusion to obtain formula (4).

[0051] f1=δ1(F1([t(O H ),t(O W (4)

[0052] Where [,] represents concatenation along the channel dimension, t represents the expansion operation, δ1 is the non-linear activation function ReLU, and f1∈R 1×H×W×1 This is a lateral intermediate feature for feature fusion of spatial information in the height and depth directions. Then, the obtained lateral intermediate f1 feature is combined with the depth direction feature O. D After the features are expanded to the same size, they are reconnected and fed into a three-dimensional convolution F2 with a kernel size of 3×3×3 for feature fusion, and the spatial weights are output as shown in formula (5).

[0053] F s =δ2(F2([t(f1),t(O) D (5)

[0054] Where δ2 is the nonlinear activation function Sigmoid.

[0055] Step S3: The deep features are simultaneously fed into the dimensionality decomposition channel attention to obtain channel weights. The dimensionality decomposition channel attention consists of two parts: dimensionality decomposition and dimensionality decomposition channel attention generation. The dimensionality decomposition part is the same as the dimensionality decomposition process described in step S2.

[0056] The dimensionality decomposition channel attention generation part first uses global average pooling to compress spatial information and generate a channel feature, which is used for subsequent dimensional channel feature generation.

[0057] After the channel features are generated, the set of dimensional channel features is obtained by formula (6).

[0058] f c =δ3(F3(avg(X))) (6)

[0059] Where X represents the input depth features, avg represents the global average pooling operation, F3 is a 3D convolution with a kernel size of 1×1×1, δ3 is the ReLU non-linear activation function, and f c ∈R 3c×1×1×1 The resulting dimensional channel feature set is obtained. After partitioning, the resulting dimensional channel feature set is concatenated with the generated directional features by channel and then fed into a 3D convolutional layer with a kernel size of 1×1×1 for fusion. The output f in the H dimension... H ∈R c×1×1×1 It is given by formula (7).

[0060] f H =δ H (F H ([O H ,s(f c (7)

[0061] Where [,] represents splicing along the channel dimension, s is the split operation, and F H For a 3D convolution with a kernel size of 1×1×1, δ H The ReLU is a non-linear activation function. Similarly, the outputs in the W and Z dimensions are given by equations (8) and (9).

[0062] f W =δ W (F W ([O W ,s(f c (8)

[0063] f D =δ D (F D ([O D ,s(f c (9)

[0064] After connecting the three-dimensional channel features, the final channel weights are generated using formula (10).

[0065] F c =δ4(F4([f H ,f W ,f D ])) (10)

[0066] Where F3 is a 3D convolution with a kernel size of 1×1×1, and δ3 is a non-linear activation function, Sigmoid.

[0067] Step S4: The depth features F and spatial weights F obtained in steps S3 and S2 s and channel weight Fc According to formula F f =F×F s ×F c The features are then fused together to generate the final features, which are then fed into an upsampling network to obtain the final segmentation result.

[0068] Figure 4 This is a flowchart illustrating a two-stage 3D image segmentation method based on dimensionality decomposition attention, as provided in an embodiment of the present invention. Figure 4 As shown, the two-stage 3D image segmentation method based on dimensionality decomposition attention includes:

[0069] Step 401 involves preprocessing and coarse segmenting the input image to obtain the region of interest through cropping. This includes, for example, cropping the input image and resizing it to a uniform size to remove the influence of image edge information; feeding the preprocessed image of uniform size into a coarse segmentation network to obtain the coarse segmentation result; and performing erosion and maximum connected component localization operations on the coarse segmentation result, using fixed-size rectangular blocks to crop the original image and the ground truth label to obtain the region of interest.

[0070] Step 402 involves downsampling the obtained region of interest to obtain depth features, and then feeding these depth features into a dimensionality decomposition spatial attention module to obtain spatial weights. After downsampling the obtained region of interest to obtain depth features, the step further includes feeding the obtained depth features into a dimensionality decomposition spatial attention module to generate spatial features.

[0071] In one embodiment, feeding the depth features into a dimensionality decomposition spatial attention to obtain spatial weights includes: the dimensionality decomposition spatial attention first generates spatial features X from the input depth features through pooling operations; the spatial features X are decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1), and (1, 1, D). H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively;

[0072] The three obtained one-dimensional directional features are fused in the order of H, W, and then D to generate spatial weights, including: first fusion of the height directional feature O H With width direction feature O W The feature is unified to a size of (H, W, 1) through a dilation operation, and then a 3D convolution operation with a kernel size of 3×3×3 is performed to fuse the features in the height and width directions, resulting in the lateral intermediate feature f1, as shown in the formula f1=δ1(F1([t(O H ),t(O WAs shown in the diagram, [,] represents connection along the channel, t represents expansion operation, and δ1 is the nonlinear activation function ReLU; then, the lateral intermediate feature f1 and the depth direction feature O are combined. D The size is unified to (H, W, D) through a dilation operation, and then fused using a 3×3×3 three-dimensional convolution with a kernel size of 3×3×3. The spatial weights F are obtained after passing through the sigmoid activation function. s The spatial weight F s =δ2(F2([t(f1),t(O) D )])), where δ2 is the nonlinear activation function Sigmoid.

[0073] Step 403: The depth features are fed into the dimensionality decomposition channel attention to obtain channel weights. For example, the dimensionality decomposition channel attention first generates spatial features X from the input depth features through pooling operations; the spatial features X are decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1) and (1, 1, D). H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively;

[0074] The input depth features are processed through pooling and convolution operations to generate channel features f. c Among them, the channel feature f c =δ3(F3(avg(X))), where X is the input depth feature, avg represents global average pooling, F3 is a 1×1×1 3D convolution with a kernel size of δ3, and ReLU is the non-linear activation function. c ∈R 3c×1×1×1 The resulting set of dimensional channel features;

[0075] The resulting dimensional channel feature set is split into three intermediate features and then combined with the one-dimensional directional feature O. H O W and O D Each channel is concatenated and fed into a 1×1×1 3D convolutional layer for fusion, yielding channel features in three dimensions: H, W, and D. The feature f in the H dimension is... H ∈R c×1×1×1 From formula f H =δ H (F H ([O H ,s(f c The result is obtained as follows: [,] represents splicing along the channel dimension, s is the splitting operation, and F... H For a 1×1×1 convolution, δ H ReLU is a non-linear activation function.

[0076] The features of the W and D dimensions are given by formula f W =δ W (F W ([O W ,s(f c )])) and f D =δ D (F D ([O D ,s(f c )]))get;

[0077] The obtained three-dimensional channel features are concatenated by channel, and then processed using formula F. c =δ4(F4([f H ,f W ,f Z Generate the final channel weights F. c , where F4 is a 3D convolution with a kernel size of 1×1×1, and δ4 is a non-linear activation function Sigmoid.

[0078] Step S4: Multiply the depth features with the spatial weights and channel weights to obtain the final features, and then feed the final features into the upsampling network to obtain the segmentation result, for example, including:

[0079] Depth features F and spatial weights F s and channel weight F c According to formula F f =F×F s ×F c The features are fused together to obtain the final features, and then fed into an upsampling network to obtain the segmentation result.

[0080] Figure 5 This is a schematic diagram of the structure of a two-stage 3D image segmentation system based on dimensionality decomposition attention, provided in an embodiment of the present invention. The system includes:

[0081] The cropping device 501 is used to preprocess and coarsely segment the input image in order to obtain the region of interest through cropping.

[0082] The sampling device 502 is used to downsample the obtained region of interest to obtain depth features, and to send the depth features into the dimensionality decomposition spatial attention to obtain spatial weights.

[0083] The acquisition device 503 is used to feed the depth features into the dimensionality decomposition channel attention to obtain the channel weights.

[0084] The segmentation device 504 is used to multiply the depth features with the spatial weights and channel weights to obtain the final features, and then feed the final features into the upsampling network to obtain the segmentation result.

[0085] This invention proposes a two-stage 3D image segmentation method based on lightweight attention using dimensionality decomposition, overcoming the performance degradation caused by the similar intensity of the target region and adjacent tissues. It offers advantages such as low computational cost, high accuracy, and strong robustness.

[0086] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

Claims

1. A two-stage 3D image segmentation method based on dimensionality decomposition attention, characterized in that, A two-stage strategy and dimensionality decomposition attention are used to segment 3D images; the method includes the following steps: Step S1: Preprocess and coarsely segment the input image to obtain the region of interest through cropping; Step S2: Downsample the obtained region of interest to obtain depth features, and feed the depth features into the dimensionality decomposition spatial attention to obtain spatial weights; Step S3: Feed the deep features into the dimensionality decomposition channel attention to obtain the channel weights; Step S4: Multiply the depth features with the spatial weights and channel weights to obtain the final features, and then feed the final features into the upsampling network to obtain the segmentation result; The step of feeding the depth features into the dimensionality decomposition spatial attention to obtain spatial weights includes: The dimensionality decomposition spatial attention first generates spatial features X from the input depth features through pooling operations; spatial features X are then decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1), and (1, 1, D). H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively; The three obtained one-dimensional directional features are fused in the order of H, W, and then D to generate spatial weights, including: first fusion of the height directional feature O H With width direction feature O W The feature is unified to a size of (H, W, 1) through a dilation operation, and then a 3D convolution operation with a kernel size of 3×3×3 is performed to fuse the features in the height and width directions, resulting in the lateral intermediate feature f1, as shown in the formula f1=δ1(F1([t(O H ),t(O W As shown in the diagram, [,] represents connection along the channel, t represents expansion operation, and δ1 is the nonlinear activation function ReLU; then, the lateral intermediate feature f1 and the depth direction feature O are combined. D The size is unified to (H, W, D) through a dilation operation, and then fused using a 3×3×3 three-dimensional convolution with a kernel size of 3×3×3. The spatial weights F are obtained after passing through the sigmoid activation function. s The spatial weight F s =δ2(F2([t(f1),t(O) D )])), where δ2 is the nonlinear activation function Sigmoid; The step of feeding the depth features into the dimensionality decomposition channel attention to obtain channel weights includes: the dimensionality decomposition channel attention first generates spatial features X from the input depth features through pooling operations; the spatial features X are decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1) and (1, 1, D) respectively. H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively; The input depth features are processed through pooling and convolution operations to generate channel features f. c Among them, the channel feature f c =δ3(F3(avg(X))), where X is the input depth feature, avg represents global average pooling, F3 is a 1×1×1 3D convolution with a kernel size of δ3, and ReLU is the non-linear activation function. c ∈R 3c×1×1×1 The resulting set of dimensional channel features; The resulting dimensional channel feature set is split into three intermediate features and then combined with the one-dimensional directional feature O. H O W and O D Each channel is concatenated and fed into a 1×1×1 3D convolutional layer for fusion, yielding channel features in three dimensions: H, W, and D. The feature f in the H dimension is... H ∈R c×1×1×1 From formula f H =δ H (F H ([O H ,s(f c The result is obtained as follows: [,] represents splicing along the channel dimension, s is the splitting operation, and F... H For a 1×1×1 convolution, δ H ReLU is a non-linear activation function. The features of the W and D dimensions are given by formula f W =δ W (F W ([O W ,s(f c )])) and f D =δ D (F D ([O D ,s(f c )]))get; The obtained three-dimensional channel features are concatenated by channel, and then processed using formula F. c =δ4(F4([f H ,f W ,f Z Generate the final channel weights F. c , where F4 is a 3D convolution with a kernel size of 1×1×1, and δ4 is a non-linear activation function Sigmoid.

2. The two-stage three-dimensional image segmentation method according to claim 1, characterized in that, The preprocessing and coarse segmentation of the input image to obtain the region of interest through cropping includes: The input image is cropped and resized to a uniform size to remove the influence of image edge information; A preprocessed image of uniform size is fed into a coarse segmentation network to obtain a coarse segmentation result. Then, erosion and maximum connectivity location operations are performed on the coarse segmentation result. The original image and the ground truth label are cropped using rectangular blocks of fixed size to obtain the region of interest.

3. The two-stage three-dimensional image segmentation method according to claim 1, characterized in that, After downsampling the obtained region of interest to obtain depth features, the process also includes feeding the obtained depth features into the dimensionality decomposition spatial attention module to generate spatial features.

4. The two-stage three-dimensional image segmentation method according to claim 1, characterized in that, The step of multiplying the depth features with spatial weights and channel weights to obtain the final features, and then feeding the final features into the upsampling network to obtain the segmentation result includes: Depth features F and spatial weights F s and channel weight F c According to formula F f =F×F s ×F c The features are fused together to obtain the final features, and then fed into an upsampling network to obtain the segmentation result.

5. A two-stage 3D image segmentation system based on dimensionality decomposition attention, characterized in that, The system employs a two-stage strategy and dimensionality decomposition attention to segment 3D images; the system includes: A cropping device is used to preprocess and coarsely segment the input image in order to obtain the region of interest through cropping; A sampling device is used to downsample the obtained region of interest to obtain depth features, and to feed the depth features into the dimensionality decomposition spatial attention to obtain spatial weights. A device for obtaining deep features by feeding them into a dimensionality-decomposition channel attention to obtain channel weights; The segmentation device is used to multiply the depth features with spatial weights and channel weights to obtain the final features, and then feed the final features into the upsampling network to obtain the segmentation result; Specifically, feeding the depth features into a dimensionality decomposition spatial attention layer to obtain spatial weights includes: The dimensionality decomposition spatial attention first generates spatial features X from the input depth features through pooling operations; spatial features X are then decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1), and (1, 1, D). H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively; The three obtained one-dimensional directional features are fused in the order of H, W, and then D to generate spatial weights, including: first fusion of the height directional feature O H With width direction feature O W The feature is unified to a size of (H, W, 1) through a dilation operation, and then a 3D convolution operation with a kernel size of 3×3×3 is performed to fuse the features in the height and width directions, resulting in the lateral intermediate feature f1, as shown in the formula f1=δ1(F1([t(O H ),t(O W As shown in the diagram, [,] represents connection along the channel, t represents expansion operation, and δ1 is the nonlinear activation function ReLU; then, the lateral intermediate feature f1 and the depth direction feature O are combined. D The size is unified to (H, W, D) through a dilation operation, and then fused using a 3×3×3 three-dimensional convolution with a kernel size of 3×3×3. The spatial weights F are obtained after passing through the sigmoid activation function. s The spatial weight F s =δ2(F2([t(f1),t(O) D )])), where δ2 is the nonlinear activation function Sigmoid; The step of feeding the depth features into the dimensionality decomposition channel attention to obtain channel weights includes: the dimensionality decomposition channel attention first generates spatial features X from the input depth features through pooling operations; the spatial features X are decomposed into three one-dimensional directional features O by three pooling kernels of sizes (H, 1, 1), (1, W, 1) and (1, 1, D) respectively. H O W and O D , where H, W and D are the height, width and depth of the three-dimensional image, respectively; The input depth features are processed through pooling and convolution operations to generate channel features f. c Among them, the channel feature f c =δ3(F3(avg(X))), where X is the input depth feature, avg represents global average pooling, F3 is a 1×1×1 3D convolution with a kernel size of δ3, and ReLU is the non-linear activation function. c ∈R 3c×1×1×1 The resulting set of dimensional channel features; The resulting dimensional channel feature set is split into three intermediate features and then combined with the one-dimensional directional feature O. H O W and O D Each channel is concatenated and fed into a 1×1×1 3D convolutional layer for fusion, yielding channel features in three dimensions: H, W, and D. The feature f in the H dimension is... H ∈R c×1×1×1 From formula f H =δ H (F H ([O H ,s(f c The result is obtained as follows: [,] represents splicing along the channel dimension, s is the splitting operation, and F... H For a 1×1×1 convolution, δ H ReLU is a non-linear activation function. The features of the W and D dimensions are given by formula f W =δ W (F W ([O W ,s(f c )])) and f D =δ D (F D ([O D ,s(f c )]))get; The obtained three-dimensional channel features are concatenated by channel, and then processed using formula F. c =δ4(F4([f H ,f W ,f Z Generate the final channel weights F. c , where F4 is a 3D convolution with a kernel size of 1×1×1, and δ4 is a non-linear activation function Sigmoid.

Images