A three-dimensional image segmentation method, system and storage medium

Abstract

The present invention provides a three-dimensional image segmentation method, system and storage medium, the method comprising: a compression step: performing multiple stages of compression processing on the input three-dimensional image in sequence to obtain a compressed image, each stage of compression processing includes extracting fine-grained features of the image; a decompression step: performing multiple stages of decompression processing on the compressed image in sequence to obtain a decompressed image, each stage of decompression processing includes performing resolution enhancement processing on the image to be decompressed in the previous stage and the fine-grained features to obtain a high-resolution image to be decompressed; after fusing the image to be decompressed with the high-resolution image to be decompressed; an image segmentation step: segmenting the decompressed image to obtain a foreground segmentation image. The solution provided by the present invention can improve the efficiency of three-dimensional image segmentation while ensuring that the fine-grained features corresponding to the foreground part in the decompressed image obtained in the decompression stage are retained to the greatest extent, so that the contour of the foreground part after the final image segmentation is clearer and more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of stereoscopic image processing, and in particular to a three-dimensional image segmentation method, system and storage medium. Background Art

[0002] With the development and progress of science and technology and people's visual needs, the application of three-dimensional images has been increasingly widely used in various fields. If a three-dimensional image is to be processed, the foreground part (such as the portrait area) of the three-dimensional image needs to be separated from the background part, that is, three-dimensional image segmentation processing.

[0003] However, since the number of voxels in the foreground part of a general three-dimensional image is less than the number of voxels in the background part (voxel: volume element, Volume Pixel), the number of voxels that need to be segmented in the input image is not very large, but the total number of voxels that need to be processed in the three-dimensional image segmentation process is very large. This unbalanced number of voxels may affect the extraction speed of the foreground voxels and reduce the image segmentation efficiency.

[0004] In order to speed up image segmentation, some solutions adopt the method of reducing the resolution of the image and directly segment the low-resolution image. Although this method can improve the segmentation efficiency, since the resolution of the foreground part is also reduced, the final segmentation effect will be poor, for example, the outline of the segmented foreground part is not clear. Summary of the invention

[0005] The present invention aims to provide a three-dimensional image segmentation method, system and storage medium to solve the technical problems in the prior art that the three-dimensional image segmentation process is low in efficiency or the segmentation effect is poor.

[0006] To this end, some embodiments of the present invention provide a three-dimensional image segmentation method, comprising the following steps:

[0007] Compression step: The input 3D image is compressed in N stages to obtain the compressed image Y N The compression process of the nth stage includes: obtaining the image Y to be compressed obtained in the (n-1)th stage n-1 , reduce the Y of the image to be compressed n-1 The resolution of the low-resolution image YL to be compressed is obtained n-1 ; The image to be compressed Y n-1 and the low-resolution image to be compressed YL n-1 After fusion, the image Y to be compressed in the nth stage is obtained n , extract the image Y to be compressed n The fine-grained features F n; Wherein, 1≤n≤N, and when n=1, the input three-dimensional image is used as the image to be compressed obtained in the previous stage;

[0008] Decompression step: Decompression of the compressed image Y N The decompression process of N stages is performed in sequence to obtain a decompressed image. The decompression process of the mth stage includes: obtaining the image to be decompressed J obtained in the (m-1)th stage m-1 , get the image Y to be compressed (N-m+1) The fine-grained features F (N-m+1) , treat the decompressed image J m-1 With fine-grained features F (N-m+1) The resolution enhancement process is performed together to obtain a high-resolution image to be decompressed JG m-1 ; The image to be decompressed J m-1 With high resolution image to be decompressed JG m-1 After fusion, we get the image J to be decompressed in the mth stage m ; Where 1≤m≤N, and when m=1, the compressed image Y obtained by the Nth stage compression process N As the image to be decompressed obtained in the previous stage;

[0009] Image segmentation step: segment the decompressed image to obtain a foreground segmentation image.

[0010] A three-dimensional image segmentation method provided by some embodiments of the present invention:

[0011] In the compression step, in the compression process of the nth stage, the image Y to be compressed in the (n-1)th stage is compressed by the residual network. n-1 The resolution is reduced to obtain a low-resolution image YL to be compressed. n-1 ; The number of feature channels of the residual network used in the compression process of the nth stage is twice the number of feature channels of the residual network used in the compression process of the (n-1)th stage;

[0012] In the decompression step, in the decompression process of the mth stage, the decompressed image J is treated by a deconvolution network corresponding to the residual network in the compression step. m-1 With fine-grained features F (N-m+1) Perform the resolution enhancement process together to obtain the high-resolution image to be decompressed JG m-1 ; Among them, the number of feature channels of the deconvolution network used in the decompression processing of the mth stage is half the number of feature channels of the deconvolution network used in the decompression processing of the (m-1)th stage; the number of feature channels of the deconvolution network when m=1 is the same as that when m=2.

[0013] A three-dimensional image segmentation method provided by some embodiments of the present invention:

[0014] In the compression step, in the compression process of the nth stage, the residual network uses the i×i×i voxel volume kernel to compress the image Y to be compressed in the (n-1)th stage n-1 Perform convolution processing;

[0015] In the decompression step, in the mth stage of decompression processing, the deconvolution network uses the i×i×i voxel volume kernel to treat the decompressed image J m-1 With fine-grained features F (N-m+1) Perform convolution processing;

[0016] In the above, i is an integer greater than 3.

[0017] A three-dimensional image segmentation method provided by some embodiments of the present invention:

[0018] In the compression step, in the nth stage of compression processing, the fine-grained features F are extracted by non-overlapping j×j×j volume blocks in the convolutional neural network. n .

[0019] A three-dimensional image segmentation method provided by some embodiments of the present invention:

[0020] In the decompression step, the decompression process of the mth stage further includes: extracting the image J to be decompressed in the mth stage through the non-overlapping j×j×j volume blocks in the deconvolution neural network m The fine-grained features F m .

[0021] A three-dimensional image segmentation method provided by some embodiments of the present invention:

[0022] In the image segmentation step, the decompressed image is processed by a convolution layer with a kernel size of 1×1×1 and a normalization processing function to segment the decompressed image into a foreground segmentation image and a background segmentation image.

[0023] The three-dimensional image segmentation method provided by some embodiments of the present invention further includes the following steps:

[0024] The similarity between the decompressed image and the input three-dimensional image is determined according to the objective function, and the variable parameters in the compression process and the decompression process are adjusted according to the similarity result until the similarity result reaches a maximum value.

[0025] Some embodiments of the present invention further provide a three-dimensional image segmentation system, including:

[0026] The image compression module performs N stages of compression processing on the input 3D image to obtain the compressed image Y N The compression process of the nth stage includes: obtaining the image Y to be compressed obtained in the (n-1)th stagen-1 , reduce the Y of the image to be compressed n-1 The resolution of the low-resolution image YL to be compressed is obtained n-1 ; The image to be compressed Y n-1 and the low-resolution image to be compressed YL n-1 After fusion, the image Y to be compressed in the nth stage is obtained n , extract the image Y to be compressed n The fine-grained features F n ; Wherein, 1≤n≤N, and when n=1, the input three-dimensional image is used as the image to be compressed obtained in the previous stage;

[0027] The image decompression module decompresses the compressed image Y N The decompression process of N stages is performed in sequence to obtain a decompressed image. The decompression process of the mth stage includes: obtaining the image to be decompressed J obtained in the (m-1)th stage m-1 , get the image Y to be compressed (N-m+1) The fine-grained features F (N-m+1) , treat the decompressed image J m-1 With fine-grained features F (N-m+1) The resolution enhancement process is performed together to obtain a high-resolution image to be decompressed JG m-1 ; The image to be decompressed J m-1 With high resolution image to be decompressed JG m-1 After fusion, we get the image J to be decompressed in the mth stage m ; Where 1≤m≤N, and when m=1, the compressed image Y obtained by the Nth stage compression processing in the image compression module N As the image to be decompressed obtained in the previous stage;

[0028] The image segmentation module segments the decompressed image to obtain a foreground segmentation image.

[0029] In a three-dimensional image segmentation system provided by some embodiments of the present invention, the image segmentation module further includes:

[0030] The target prediction submodule determines the similarity between the decompressed image and the input three-dimensional image according to the target function, and adjusts the variable parameters in the compression process and the decompression process according to the similarity result until the similarity result reaches a maximum value.

[0031] Some embodiments of the present invention further provide a storage medium, wherein the storage medium stores program information for a computer to read. After the computer reads the program information, it executes any one of the above three-dimensional image segmentation methods.

[0032] Compared with the prior art, the technical solution provided by the embodiment of the present invention has at least the following beneficial effects: the solution provided by the present invention separately extracts the complete fine-grained features of the current stage in each compression stage, avoiding the loss of fine-grained features in the subsequent compression path, thereby ensuring that the fine-grained features corresponding to the foreground part in the decompressed graphics obtained in the decompression stage are retained to the greatest extent, and the contour of the foreground part after the final image segmentation is clearer and more accurate.

[0033] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings are used to provide a further understanding of the scheme of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0035] Figure 1 A flowchart of a three-dimensional image segmentation method according to an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of a network framework of a three-dimensional image segmentation method according to an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of the principle of performing convolution and deconvolution operations by a convolutional neural network with a stride of 2 and a 2×2×2 volume block as a voxel kernel according to an embodiment of the present invention;

[0038] Figure 4 A flowchart of a three-dimensional image segmentation system according to an embodiment of the present invention;

[0039] Figure 5 A schematic diagram of the hardware connection relationship of a device for executing a three-dimensional image segmentation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention. The terms "first", "second" and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, the various technical solutions in the following embodiments provided in the present invention can be combined with each other, and the technical features therein can be replaced with each other unless they contradict each other.

[0041] The three-dimensional image segmentation method and system provided in this embodiment can be applied to a control system that needs to segment a three-dimensional image. Figure 1 As shown, the method comprises the following steps:

[0042] S101: Compression step: Perform N stages of compression processing on the input 3D image in sequence to obtain a compressed image Y N The compression process of the nth stage includes: obtaining the image Y to be compressed obtained in the (n-1)th stage n-1 , reduce the Y of the image to be compressed n-1 The resolution of the low-resolution image YL to be compressed is obtained n-1 ; The image to be compressed Y n-1 and the low-resolution image to be compressed YL n-1 After fusion, the image Y to be compressed in the nth stage is obtained n , extract the image Y to be compressed n The fine-grained features F n ; Wherein, 1≤n≤N, and when n=1, the input image is used as the image to be compressed obtained in the previous stage. Wherein, the fusion of images can be achieved by superimposing pixels at the same position.

[0043] S102: Decompression step: Decompressing the compressed image Y N The decompression process of N stages is performed in sequence to obtain a decompressed image. The decompression process of the mth stage includes: obtaining the image to be decompressed J obtained in the (m-1)th stage m-1 , get the image Y to be compressed (N-m+1) The fine-grained features F (N-m+1) , treat the decompressed image J m-1 With fine-grained features F (N-m+1) The resolution enhancement process is performed together to obtain a high-resolution image to be decompressed JG m-1 ; The image to be decompressed J m-1 With high resolution image to be decompressed JG m-1 After fusion, we get the image J to be decompressed in the mth stage m ; Where 1≤m≤N, and when m=1, the compressed image Y obtained by the Nth stage compression process N As the image to be decompressed obtained in the previous stage.

[0044] S103: Image segmentation step: segment the decompressed image to obtain a foreground segmentation image.

[0045] In the above scheme, N compression stages correspond to N decompression stages, so that the image obtained after the final decompression has a size and resolution that is suitable for the input image. Figure 2 In the example given in Figure 2In this example, N=5, therefore n=1,2,3,4,5; m=1,2,3,4,5, where n=5 and m=1 correspond to the case where N=5. Figure 2 The network shown in FIG. 1 corresponds to the compression step in step S101 on the left, and corresponds to the decompression step and the segmentation step in step S102 on the right. As can be seen from the figure, in the 3D image compression process on the left, when executing each stage of the compression step, the compressed image Y is treated as n-1 and the low-resolution image to be compressed YL n-1 The fused image to be compressed Y n , a fine-grained feature F is implemented n The extracted fine-grained features are forwarded to the decompression process on the right side. The fine-grained features correspond to the feature data of the foreground part of the input image, that is, the human-shaped area in the input image shown in the figure.

[0046] The above scheme extracts the complete fine-grained features of the current stage separately in each compression stage to avoid the loss of fine-grained features in the subsequent compression path, thereby ensuring that the fine-grained features corresponding to the foreground part in the decompressed image obtained in the decompression stage are retained to the greatest extent. The contour of the foreground part of the final image after segmentation is clearer and more accurate.

[0047] Because in the above scheme of the present invention, under the big framework of data processing, the compression and decompression process does not need to analyze all voxels in the input image, the data processing amount is reduced, so that the segmentation efficiency of the input three-dimensional image can be ensured. In fact, after testing, this scheme has greatly reduced the processing time compared with the existing image segmentation processing method. At the same time, because the fine-grained features are collected in the compression stage, they can be directly applied in the subsequent decompression process, thereby ensuring the integrity of the fine-grained features and ensuring that the outline of the foreground part is clearer.

[0048] In the above scheme, in step S101, in the compression process of the nth stage, the residual network is used to compress the image Y of the (n-1)th stage. n-1 The resolution is reduced to obtain a low-resolution image YL to be compressed. n-1 ; wherein the number of feature channels of the residual network used in the compression process of the nth stage is twice the number of feature channels of the residual network used in the compression process of the (n-1)th stage; in the step S102, in the decompression process of the mth stage, the decompressed image J is treated by the deconvolution network corresponding to the residual network in the compression step m-1 With fine-grained features F (N-m+1) Perform the resolution enhancement process together to obtain the high-resolution image to be decompressed JG m-1; Among them, the number of feature channels of the deconvolution network used in the decompression processing of the mth stage is half the number of feature channels of the deconvolution network used in the decompression processing of the (m-1)th stage, and the number of feature channels of the deconvolution network when m=1 is the same as that when m=2.

[0049] like Figure 2 As shown in the figure, the input 3D image resolution is set to 128×128×64. During the compression process, when n=1, the number of channels selected in the residual network is 16, and the image resolution is 128×128×64; when n=2, the number of channels selected in the residual network is 32, and the image resolution is 64×64×32; when n=3, the number of channels selected in the residual network is 64, and the image resolution is 32×32×16; when n=4, the number of channels selected in the residual network is 128, and the image resolution is 16×16×8; when n=5 (N=5), the number of channels selected in the residual network is 256, and the image resolution is 8×8×4. It can be determined from the figure that for each stage of the compression process, when the image resolution is different, the number of residual network channels allocated to the image is also different, and each stage of the compression process performs one to three convolutional layers. The residual network is a convolutional neural network commonly used in the prior art. Its advantage is that in each compression stage, the input data includes two data, one of which is the output data of the previous layer, and the other is the data after processing in this stage. The residual of these two data is learned to optimize the processing result of this stage. In the above scheme, since the number of feature channels is doubled at each stage of the compression path, and since the model for processing the image is a residual network, the above convolution operation can double the number of feature maps obtained in each stage compared to the number of feature maps obtained in the previous stage, while reducing the resolution of each feature map. As shown in Table 1, it provides Figure 2 The receptive fields of the network layers at each stage in the left part (in the convolutional neural network, the receptive field refers to the size of the area mapped on the original image by the pixels on the feature map output by each layer of the convolutional neural network). The data recorded in Table 1 can confirm that each network layer in the compression path in this scheme has captured the content of the entire input information of the network layer.

[0050] Table 1 Theoretical receptive fields of the convolutional layer during the compression and decompression process of this embodiment.

[0051] compression Input size Receptive Field Unzip Input size Receptive Field n＝1 128 5×5×5 m=1(N=5) 8 372×372×372 n＝2 64 22×22×22 m＝2 16 476×476×476 n＝3 32 72×72×72 m＝3 32 528×528×528 n＝4 16 172×172×172 m＝4 64 546×546×546 N＝5 8 372×372×372 m＝5 128 551×551×551

[0052] Based on a similar principle, in the decompression process, the deconvolution process corresponding to the residual network in the compression process is used to reverse the effect of convolution on image data processing in the compression stage. Figure 2As shown, when m=1 (N=5), the number of channels in the deconvolution network is 256, and the image resolution is 8×8×4; when m=2, the number of channels selected in the deconvolution network is 256, and the image resolution is 16×16×8; when m=3, the number of channels selected in the deconvolution network is 128, and the image resolution is 32×32×16; when m=4, the number of channels selected in the deconvolution network is 64, and the image resolution is 64×64×32; when m=5, the number of channels selected in the deconvolution network is 32, and the image resolution is 128×128×64. The deconvolution operation is a common way to implement the upsampling process in the prior art, and has been widely used in the field of image processing, and will not be repeated in this embodiment.

[0053] In some embodiments, in step S101, in the compression process of the nth stage, the residual network uses the volume kernel of i×i×i voxels to compress the image Y to be compressed in the (n-1)th stage n-1 Convolution processing is performed; in the step S102, in the decompression processing of the mth stage, the deconvolution network uses the volume kernel of i×i×i voxels to treat the decompressed image J m-1 With fine-grained features F (N-m+1) Convolution processing is performed; in the above, i is an integer greater than 3, and i can be selected as 3, 5, 7, etc. Preferably, i=5, that is, a volume kernel of size 5×5×5 voxels is used to perform a convolution operation on the input data in each stage of compression and decompression, which can ensure both the speed of image processing and the contour quality of the image segmentation result.

[0054] Further preferably, in the step S101, in the nth stage of compression processing, the fine-grained features F are extracted by non-overlapping j×j×j volume blocks in the convolutional neural network. n , j<i, preferably j=2. Specifically, a 2×2×2 voxel wide kernel convolution with a stride of 2 is used in the convolutional neural network. By using non-overlapping 2×2×2 volume blocks to extract features, the size of the generated feature map is halved, thereby reducing the amount of data processing for extracting fine-grained features. Accordingly, in step S102, the decompression process of the mth stage also includes: extracting the image J to be decompressed in the mth stage through the non-overlapping j×j×j volume blocks in the deconvolution neural network m The fine-grained features F m , preferably j = 2. Figure 3As shown in FIG. 1 , the compression process and the decompression process perform convolution and deconvolution operations. In the compression process, a convolution operation with an appropriate step size is selected to reduce the size of the data. The deconvolution operation in the decompression process projects each input voxel to a larger area through the kernel to increase the data size. As mentioned above, after determining the model used for the convolution operation, the corresponding deconvolution algorithm model can be selected according to the existing upsampling process.

[0055] Figure 2 In the network structure shown, the number of features calculated at each stage of the left part is twice that of the previous layer, and the right part extracts features and expands the spatial support of the lower resolution feature map in order to collect and assemble the necessary information to output a two-channel volume segmentation. In the final image segmentation step, the decompressed image is processed by a convolution layer with a kernel size of 1×1×1 and a normalization processing function to segment the decompressed image into a foreground segmentation image and a background segmentation image. As shown in the figure, two feature maps can be obtained by using a convolution layer in the segmentation step. The two feature maps have the same volume size. Finally, the voxels are converted into probabilistic segmentation of the foreground and background regions by applying Softmax (normalization processing function). The so-called probabilistic segmentation means that for a voxel, if the probability of it belonging to the foreground part is large, it is considered that it should be segmented into the segmented image of the foreground part, and if the probability of it belonging to the background part is large, it is considered that it should be segmented into the segmented image of the background part.

[0056] Preferably, the three-dimensional image segmentation method described in the above scheme further includes: determining the similarity between the decompressed image and the input three-dimensional image according to the objective function, and adjusting the variable parameters in the compression process and the decompression process according to the similarity result until the similarity result reaches a maximum value. Wherein, the objective function can be implemented by selecting the Dice function, and the input three-dimensional image of the Dice function and the decompressed image with the same resolution as the original input three-dimensional image, the two images are input into the Dice function in the form of binary volume to obtain the overlap of the two images. The Dice function originates from binary classification, and is essentially a measure of the overlapping part of two samples. The index range of its output is from 0 to 1, where "1" indicates complete overlap and "0" indicates no overlap at all. In specific implementation, the adjustable weight parameters in the aforementioned convolution, deconvolution and other algorithm models can be adjusted to observe the final output of the Dice function, and the adjusted result with the maximum output result of the Dice function is used as the optimal three-dimensional image segmentation network, thereby ensuring that the segmentation result has the maximum restoration effect with the original input three-dimensional image.

[0057] like Figure 4As shown, the embodiment of the present invention further provides a three-dimensional image segmentation system, including: an image compression module 401, which performs N stages of compression processing on the input three-dimensional image in sequence to obtain a compressed image Y N The compression process of the nth stage includes: obtaining the image Y to be compressed obtained in the (n-1)th stage n-1 , reduce the Y of the image to be compressed n-1 The resolution of the low-resolution image YL to be compressed is obtained n-1 ; The image to be compressed Y n-1 and the low-resolution image to be compressed YL n-1 After fusion, the image Y to be compressed in the nth stage is obtained n , extract the image Y to be compressed n The fine-grained features F n ; Wherein, 1≤n≤N, and when n=1, the input 3D image is used as the image to be compressed obtained in the previous stage. The image decompression module 402 decompresses the compressed image Y N The decompression process of N stages is performed in sequence to obtain a decompressed image. The decompression process of the mth stage includes: obtaining the image to be decompressed J obtained in the (m-1)th stage m-1 , get the image Y to be compressed (N-m+1) The fine-grained features F (N-m+1) , treat the decompressed image J m-1 With fine-grained features F (N-m+1) The resolution enhancement process is performed together to obtain a high-resolution image to be decompressed JG m-1 ; The image to be decompressed J m-1 With high resolution image to be decompressed JG m-1 After fusion, we get the image J to be decompressed in the mth stage m ; Where 1≤m≤N, and when m=1, the compressed image Y obtained by the Nth stage compression processing in the image compression module N As the image to be decompressed obtained in the previous stage, the image segmentation module 403 segments the decompressed image to obtain a foreground segmentation image.

[0058] In the above scheme of the present invention, under the framework of data processing, the compression and decompression process does not need to analyze all voxels in the input image, which reduces the amount of data processing, thereby ensuring the efficiency of segmentation of the input three-dimensional image. In fact, after testing, this scheme has greatly reduced the processing time compared with the existing image segmentation processing method. At the same time, since the fine-grained features are collected in the compression stage, they can be directly applied in the subsequent decompression process, thereby ensuring the integrity of the fine-grained features and ensuring that the outline of the foreground part is clearer.

[0059] Preferably, the image segmentation module 403 further includes: a target prediction submodule, which determines the similarity between the decompressed image and the input three-dimensional image according to the target function, and adjusts the variable parameters in the compression process and the decompression process according to the similarity result until the similarity result reaches a maximum value. This solution can ensure that the segmentation result has the maximum restoration effect with the original input three-dimensional image.

[0060] Some embodiments of the present invention provide a storage medium, wherein program instructions are stored in the storage medium. After a computer reads the program instructions, the three-dimensional image segmentation method described in any one of the method embodiments above is executed.

[0061] Figure 5 5 is a schematic diagram of the hardware structure of a device for performing a three-dimensional image segmentation method provided in this embodiment. The device includes one or more processors 501 and a memory 502. Figure 5 A processor 501 is taken as an example. The device may further include: an input device 503 and an output device 504. The processor 501, the memory 502, the input device 503 and the output device 504 may be connected via a bus or other means.

[0062] The memory 502 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The processor 501 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 502, that is, the three-dimensional image segmentation method of the above method embodiment is implemented. The above device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0063] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0064] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0065] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. If these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A three-dimensional image segmentation method, characterized in that: The steps include: Compression step: The input 3D image is compressed in N stages to obtain the compressed image Y N The compression process of the nth stage includes: obtaining the image Y to be compressed obtained in the (n-1)th stage n-1 , reduce the Y of the image to be compressed n-1 The resolution of the low-resolution image YL to be compressed is obtained n-1 ; The image to be compressed Y n-1 and the low-resolution image to be compressed YL n-1 After fusion, the image Y to be compressed in the nth stage is obtained n , extract the image Y to be compressed n The fine-grained features F n ; Wherein, 1≤n≤N, and when n=1, the input three-dimensional image is used as the image to be compressed obtained in the previous stage; Decompression step: Decompression of the compressed image Y N The decompression process of N stages is performed in sequence to obtain a decompressed image. The decompression process of the mth stage includes: obtaining the image to be decompressed J obtained in the (m-1)th stage m-1 , get the image Y to be compressed (N-m+1) The fine-grained features F (N-m+1) , treat the decompressed image J n-1 With fine-grained features F (N-m+1) The resolution enhancement process is performed together to obtain a high-resolution image to be decompressed JG m-1 ; The image to be decompressed J m-1 With high resolution image to be decompressed JG m-1 After fusion, we get the image J to be decompressed in the mth stage m ; Where 1≤m≤N, and when m=1, the compressed image Y obtained by the Nth stage compression process N As the image to be decompressed obtained in the previous stage; Image segmentation step: segment the decompressed image to obtain a foreground segmentation image.

2. The three-dimensional image segmentation method according to claim 1, characterized in that: In the compression step, in the compression process of the nth stage, the image Y to be compressed in the (n-1)th stage is compressed by the residual network. n-1 The resolution is reduced to obtain a low-resolution image YL to be compressed. n-1 ; The number of feature channels of the residual network used in the compression process of the nth stage is twice the number of feature channels of the residual network used in the compression process of the (n-1)th stage; the number of feature channels of the residual network when n=1 is the same as that when n=2; In the decompression step, in the decompression process of the mth stage, the decompressed image J is treated by a deconvolution network corresponding to the residual network in the compression step. n-1 With fine-grained features F (N-m+1) Perform the resolution enhancement process together to obtain the high-resolution image to be decompressed JG m-1 ; Among them, the number of feature channels of the deconvolution network used in the decompression processing of the mth stage is half the number of feature channels of the deconvolution network used in the decompression processing of the (m-1)th stage; the number of feature channels of the deconvolution network when m=1 is the same as that when m=2.

3. The three-dimensional image segmentation method according to claim 1, characterized in that: In the compression step, in the compression process of the nth stage, the residual network uses the i×i×i voxel volume kernel to compress the image Y to be compressed in the (n-1)th stage n-1 Perform convolution processing; In the decompression step, in the mth stage of decompression processing, the deconvolution network uses the i×i×i voxel volume kernel to treat the decompressed image J n-1 With fine-grained features F (N-m+1) Perform convolution processing; In the above, i is an integer greater than 3.

4. The three-dimensional image segmentation method according to claim 2, characterized in that: In the compression step, in the nth stage of compression processing, the fine-grained features F are extracted by non-overlapping j×j×j volume blocks in the convolutional neural network. n , j<i.

5. The three-dimensional image segmentation method according to claim 4, characterized in that: In the decompression step, the decompression process of the mth stage further includes: extracting the image J to be decompressed in the mth stage through the non-overlapping j×j×j volume blocks in the deconvolution neural network m The fine-grained features F m .

6. The three-dimensional image segmentation method according to any one of claims 1 to 5, characterized in that: In the image segmentation step, the decompressed image is processed by a convolution layer with a kernel size of 1×1×1 and a normalization processing function to segment the decompressed image into a foreground segmentation image and a background segmentation image.

7. The three-dimensional image segmentation method according to claim 6, characterized in that: The following steps are also included: The similarity between the decompressed image and the input three-dimensional image is determined according to the objective function, and the variable parameters in the compression step and the decompression step are adjusted according to the similarity result until the similarity result reaches a maximum value.

8. A three-dimensional image segmentation system, characterized in that: include: The image compression module performs N stages of compression processing on the input 3D image to obtain the compressed image Y N The compression process of the nth stage includes: obtaining the image Y to be compressed obtained in the (n-1)th stage n-1 , reduce the Y of the image to be compressed n-1 The resolution of the low-resolution image YL to be compressed is obtained n-1 ; The image to be compressed Y n-1 and the low-resolution image to be compressed YL n-1 After fusion, the image Y to be compressed in the nth stage is obtained n , extract the image Y to be compressed n The fine-grained features F n ; Wherein, 1≤n≤N, and when n=1, the input three-dimensional image is used as the image to be compressed obtained in the previous stage; The image decompression module decompresses the compressed image Y N The decompression process of N stages is performed in sequence to obtain a decompressed image. The decompression process of the mth stage includes: obtaining the image to be decompressed J obtained in the (m-1)th stage m-1 , get the image Y to be compressed (N-m+1) The fine-grained features F (N-m+1) , treat the decompressed image J n-1 With fine-grained features F (N-m+1) The resolution enhancement process is performed together to obtain a high-resolution image to be decompressed JG m-1 ; The image to be decompressed J m-1 With high resolution image to be decompressed JG m-1 After fusion, we get the image J to be decompressed in the mth stage m ; Where 1≤m≤N, and when m=1, the compressed image Y obtained by the Nth stage compression processing in the image compression module N As the image to be decompressed obtained in the previous stage; The image segmentation module segments the decompressed image to obtain a foreground segmentation image.

9. The three-dimensional image segmentation system according to claim 8, characterized in that: The image segmentation module also includes: The target prediction submodule determines the similarity between the decompressed image and the input three-dimensional image according to the target function, and adjusts the variable parameters in the image compression module and the image decompression module according to the similarity result until the similarity result reaches a maximum value.

10. A storage medium, characterized in that: The storage medium stores program information for a computer to read, and the computer executes the three-dimensional image segmentation method according to any one of claims 1 to 7 after reading the program information.

Images