Image segmentation apparatus, image segmentation method, and storage medium

Abstract

The invention discloses an image segmentation device, which can obtain a smooth boundary segmentation result, can provide visual mark information convenient for a doctor to identify, and is high in segmentation accuracy. The image segmentation device includes: an acquisition unit that acquires a block region with mark information and a tubular region with mark information; a coarse division unit that generates a region to be divided, which is a boundary region of each substructure and which does not have mark information, and a non-boundary region, which is a result of coarse division and which has mark information, by distance conversion using the mark information of the acquired block region and the tubular region; a classifier generation unit that generates a classifier that classifies spatial coordinates by using the mark information of the non-boundary region and the mark information of a specific position specified by the region to be divided and the tubular region; and the fine segmentation part is used for segmenting voxels of the region to be segmented by utilizing the classifier and determining a final segmentation result.

Description

Technical Field

[0001] The present invention relates to an image segmentation device, an image segmentation method and a storage medium, and in particular to an image segmentation device, an image segmentation method and a storage medium for segmenting lung segments according to a lung lobe image and a trachea or blood vessel image with marking information. Background Art

[0002] It is known that the localization of lesions is a key step in formulating treatment plans in the diagnosis of diseases. Taking the lungs as an example, the segmentation of lobes and segments in the lungs is the first step in locating the lesion.

[0003] In this regard, the prior art proposes a technique for segmenting lung segments based on the distance from the unlabeled area to the labeled area (trachea, blood vessels, etc.) in the lung lobe. However, in the segmentation result of this method, there are slight jagged edges at the boundaries, which is not conducive to doctor observation.

[0004] In addition, the prior art also proposes a technique for segmenting lung segments by performing surface fitting according to the feature points of the lung segment boundaries marked by the user. However, this marking method is not intuitive and cannot directly use information such as trachea or blood vessels, but can only be confirmed based on the generated surface. In addition, different doctors mark different locations, so the results may be quite different. Summary of the invention

[0005] The present invention provides an image segmentation device, an image segmentation method and a storage medium, which can obtain a smooth boundary segmentation result, and can also provide intuitive marking information that is easy for doctors to identify, and the segmentation accuracy is high.

[0006] Technical solution 1 is an image segmentation device, which comprises: an acquisition unit, which acquires a block area with labeling information and a tubular area with labeling information; a coarse segmentation unit, which uses the labeling information of the acquired block area and the tubular area to generate, through distance transformation, an area to be segmented without labeling information as the boundary area of ​​each substructure, and a non-border area with labeling information as the coarse segmentation result; a classifier generation unit, which uses the labeling information of the non-border area and the labeling information of a specific position determined by the area to be segmented and the tubular area to generate a classifier for classifying spatial coordinates; and a fine segmentation unit, which uses the classifier to segment the voxels of the area to be segmented to determine the final segmentation result.

[0007] Technical solution two is an image segmentation method, comprising the following steps: an acquisition step, obtaining a block area with marking information and a tubular area with marking information; a coarse segmentation step, using the marking information of the above-mentioned block area and the above-mentioned tubular area, through distance transformation, to generate a to-be-segmented area without marking information as the boundary area of ​​each substructure, and a non-boundary area with marking information as the coarse segmentation result; a classifier generation step, using the marking information of the above-mentioned non-boundary area and the marking information of the specific position determined by the above-mentioned to-be-segmented area and the above-mentioned tubular area, to generate a classifier for classifying spatial coordinates; and a fine segmentation step, using the above-mentioned classifier to segment the voxels of the above-mentioned to-be-segmented area to determine the final segmentation result.

[0008] Technical solution three is a storage medium storing a program that enables a computer to execute the following steps: an acquisition step, acquiring a block area with marking information and a tubular area with marking information; a coarse segmentation step, using the marking information of the obtained block area and the tubular area, through distance transformation, to generate a to-be-segmented area without marking information as the boundary area of ​​each substructure, and a non-boundary area with marking information as the coarse segmentation result; a classifier generation step, using the marking information of the non-boundary area and the marking information of the specific position determined by the to-be-segmented area and the tubular area to generate a classifier for classifying spatial coordinates; and a fine segmentation step, using the classifier to segment the voxels of the to-be-segmented area to determine the final segmentation result.

[0009] Technical Effects

[0010] The image segmentation device, image segmentation method and storage medium of the present invention can obtain a smooth boundary segmentation result, and can also provide intuitive marking information that is easy for doctors to identify, and the segmentation accuracy is high. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a block diagram of the image segmentation device according to the first embodiment.

[0012] Figure 2A An image of the lung lobe region with labeled information is shown in FIG. Figure 2B An image of the trachea region with labeled information is shown in FIG.

[0013] Figure 3 This is a flowchart of the image segmentation method according to the first embodiment.

[0014] Figure 4 The method of generating the boundary area and the non-boundary area (rough segmentation result) is shown.

[0015] Figure 5 A detailed flowchart for generating a classifier using label information is shown.

[0016] Figure 6 Represents the process of uniformly sampling the coarse segmentation results.

[0017] Figure 7 Indicates how a specific location with labeled information is generated.

[0018] Figure 8 Represents the process of generating a classifier using a first input and a second input.

[0019] Fig. 9 The final segmentation result obtained by the first embodiment and the segmentation result of the comparative example for comparison are shown.

[0020] Fig.10 A block diagram showing an image segmentation device according to the second embodiment.

[0021] Fig.11 This is a flowchart of the image segmentation method according to the second embodiment.

[0022] Fig.12 This is a comparison chart before and after the result of fine segmentation is empirically labeled and corrected based on the empirical labeling. DETAILED DESCRIPTION

[0023] Hereinafter, the embodiments will be described with reference to the accompanying drawings. The embodiments described below do not limit the present invention. In addition, in the following description, the same reference numerals are used for components having substantially the same function and configuration, and repeated descriptions are made only when necessary. In addition, even the same components may be represented in different proportions in each of the accompanying drawings.

[0024] It is known that the human lungs, trachea and blood vessels have obvious physical boundaries, so segmentation is not difficult and can be manually marked by doctors or automatically marked by machine learning. However, sometimes in order to locate the lesion, the lungs need to be further divided into multiple substructures (segments) based on the supply information of the trachea, etc. The lung lobe is an outer structure that covers the trachea and blood vessels, and its interior does not have obvious physical boundaries, so it is difficult to divide the lung lobe into multiple substructures (segments).

[0025] In this regard, the present invention provides an image segmentation device and an image segmentation method. By utilizing the image segmentation device, a smooth boundary segmentation result can be obtained, and intuitive marking information that is easy for doctors to identify can be provided, and the segmentation accuracy is high.

[0026] (First Embodiment)

[0027] The image segmentation device 10 of this embodiment is as follows: Figure 1As shown, the system includes an acquisition unit 11 , a rough segmentation unit 12 , a classifier generation unit 13 , a fine segmentation unit 14 , and an output unit 15 .

[0028] The acquisition unit 11 acquires a block region with label information and a tubular region with label information. The block region is an area that can anatomically cover the tubular region. Here, as the block region, for example, a lung lobe can be listed, and as the tubular region, for example, a trachea or a blood vessel can be listed. In this embodiment, the trachea is used as an example for description, but a blood vessel can also be used as the tubular region.

[0029] In this embodiment, the lung lobe region is segmented into a first number of segments and each segment is labeled with different label information as an independent labeled region, and the trachea region is segmented into a second number of segments and each segment is labeled with different label information as an independent labeled region.

[0030] The labeled region mentioned above can be understood as a collection of voxels with the same labeling information. For example, if all voxels in a lobe region are labeled RA, the region can be called an RA labeled region, and if all voxels in a segment are labeled r1, the segment can be called an r1 labeled region.

[0031] Specifically, the lungs generally include a left lung and a right lung, and the left lung and the right lung are divided into five lobes by a fissure with a clear physical boundary, wherein the left lung has two lobes and the right lung has three lobes. The block region with label information in this embodiment refers to five label regions obtained by assigning different label information to the five lobes. For example, Figure 2A As shown, the voxels belonging to the right 1st segment of the lung are marked with RA, the voxels belonging to the right 2nd segment of the lung are marked with RB, the voxels belonging to the right 3rd segment of the lung are marked with RC, the voxels belonging to the left 1st segment are marked with LD, and the voxels belonging to the left 2nd segment are marked with LE. In addition, for the convenience of observation, the same color label is given to each voxel with the same label, so that different lung segments can be intuitively distinguished by color.

[0032] The trachea can be divided into 18 sections according to the supply relationship. The tubular region with label information in this embodiment refers to 18 label regions obtained by assigning different label information to the 18 sections. Specifically, the bronchi belonging to the 8 sections on the left are assigned label information of l1 to l8, and the bronchi belonging to the 10 sections on the right are assigned label information of r1 to r10. Figure 2B The characters l1 to l8 and r1 to r10 are omitted and only the different marking information is indicated by color.

[0033] The rough segmentation unit 12 uses the obtained Figure 2A , Figure 2BThe marking information of the lung lobes and the marking information of the trachea shown are used to generate a to-be-segmented region without marking information and a rough segmentation result with marking information through distance transformation.

[0034] Distance transform is an algorithm for measuring distance in the prior art. The algorithm can be used to calculate the voxels whose distance difference between two adjacent and different marked areas is less than a threshold, and the set of voxels at these positions is determined as the boundary area. The boundary area is then used as the subsequent area to be segmented without retaining the original marking information.

[0035] In addition, the remaining lung area after removing the boundary area, i.e., the non-boundary area, is segmented into a plurality of labeled areas corresponding to the branching structure of the trachea by distance transformation. The plurality of labeled areas corresponding to the non-boundary area obtained by distance transformation is equivalent to the coarse segmentation result in this embodiment. Through the above distance transformation, the remaining lung area is changed from the original 5 labeled areas to 18 labeled areas (l1 to l8, r1 to r10) corresponding to the branching structure of the trachea.

[0036] The classifier generation unit 13 generates a classifier using the label information in the rough segmentation result and the label information at the specific position. The classifier is, for example, a classifier obtained by using the SVM algorithm in the prior art. The principle is that the classifier can be visualized as a hyperplane, and the label information in the rough segmentation result and the label information at the specific position are used as the first input and the second input, respectively, so that the voxels of the label information of a certain category (l1~l8, r1~r10) in the rough segmentation result and the voxels of the label information of the same category in the specific position are located on the same side of the hyperplane, and the label information on both sides of the hyperplane belongs to different categories, respectively, to generate a hyperplane.

[0037] As the first input, the voxels obtained by uniformly sampling the voxels in the rough segmentation result are used. The sampled voxels can be named and stored by adding the label information after the three-dimensional coordinates in space. For example, the point set of the voxels in the right segment 1 is The point set of the voxels in the right 3 segments is

[0038] As the second input, all voxels at a specific position are used without sampling. All voxels at the specific position can also be named and stored by adding the corresponding tag information after the three-dimensional coordinates in space. For example, the point set of the voxels in the right 1 segment is The point set of the voxels in the right 3 segments is The specific position is the position where the trachea overlaps with the region to be segmented as the boundary region. The portion of the trachea that overlaps with the region to be segmented is taken as the specific position, and the marking information of the trachea at this position is taken as the marking information of the specific position. The region to be segmented is a set of voxels whose distance difference from two adjacent and different marked regions is less than a threshold value. The specific position is a part of the region to be segmented, so the voxel at the specific position is the voxel whose distance difference from the adjacent and different marked regions is less than a threshold value, which is important for distinguishing two adjacent marked regions. Therefore, when used as an input sample, it is not sampled and all voxels are used as they are.

[0039] As the input sample used when generating the classifier (hyperplane), the features used in the first input and the second input are spatial three-dimensional coordinate information and label information, and the spatial three-dimensional coordinate information and label information in the first input and the second input are related to each other. Therefore, when using such a classifier to classify spatial coordinates, the spatial three-dimensional coordinate information can be used as input to infer the classification (label information) to which it belongs. In other words, the generated classifier can use the voxels of the area to be segmented that only have spatial three-dimensional coordinate information to infer the label information to which it belongs.

[0040] The fine segmentation unit 14 uses the classifier generated by the classifier generation unit 13 to segment the voxels in the area to be segmented and determine the final segmentation result. The classifier, that is, the hyperplane, can divide the area to be segmented without label information into multiple segments. Through the spatial three-dimensional coordinate information of each segment, the label information of each segment can be inferred, so that all voxels in the area to be segmented are given different label information due to the hyperplane, and the label information of the area to be segmented and the label information of the non-boundary area (coarse segmentation result) together constitute the final segmentation result.

[0041] The output unit 15 can output and display on a display, etc. the lung lobes and trachea with labeled information acquired by the acquisition unit 10, the areas to be segmented without labeled information and the coarse segmentation results with labeled information generated by the coarse segmentation unit 12, the classifiers (hyperplanes) generated by the classifier generation unit 13, and the final segmentation results determined by the fine segmentation unit 14.

[0042] Figure 3 A flowchart of the image segmentation method according to Embodiment 1 is shown.

[0043] First, in step S101 , the lung lobes and trachea with label information are obtained.

[0044] Next, in step S102, the label information of the lung lobe and the trachea is used to generate a boundary area without label information through distance transformation, and the remaining area is used as a non-boundary area with label information.

[0045] Next, in step S103, a classifier is generated using the label information of the non-boundary region and the label information of the specific position determined by the boundary region and the trachea.

[0046] Next, in step S104, the voxels in the boundary area are segmented using a classifier to determine a final segmentation result.

[0047] Finally, in step S105, the final segmentation result is output.

[0048] Below, refer to Figures 4 to 9 The image segmentation method according to the present embodiment will be described in detail.

[0049] In the images of the block region (lung lobe) and the tubular region (trachea) obtained in step S101, the lung lobe is segmented by a first number and each segment is labeled with different label information as an independent label region, and the trachea is segmented by a second number and each segment is labeled with different label information as an independent label region. The first number is 5, and the second number is 18. In other words, the lung lobe image with 5 different label information and the trachea image with 18 different label information obtained in step S101 correspond to Figure 4 The two images on the upper side.

[0050] Next, in step S102, the images of the lung lobes with 5 kinds of label information and the trachea with 18 kinds of label information obtained in step S101 are used to generate the region to be segmented as the boundary region through distance transformation. The principle is that if the difference between the distance between a voxel and two different substructures is less than a certain threshold, it means that it is close to the boundary of the substructure. For voxels close to the boundary part, obvious jagged edges will appear when segmenting directly using distance transformation, which is not conducive to the doctor's observation. Therefore, the boundary part is retained as the region to be segmented, and the region to be segmented is segmented in the fine segmentation step described later.

[0051] Next, through distance transformation, the remaining area after removing the area to be segmented is segmented into multiple segments corresponding to the branch structure of the trachea that is segmented into multiple segments. Specifically, in each lobe, the lobe is divided into multiple segments according to the supply relationship of the trachea, and the 5 lobe areas are divided into 18 segments in total, that is, segments with 18 different labeling information. The remaining area after removing the area to be segmented is called a non-boundary area, and the non-boundary area with labeling information can be understood as the coarse segmentation result of this embodiment.

[0052] Next, proceed to step S103. Step S103 can also be divided into steps S1031 to S1034. Figure 5 A flowchart of steps S1031 to S1034 is shown in FIG.

[0053] In step S1031 , label information of the rough segmentation result is uniformly sampled as the first input.

[0054] In step S1032, the specific position with marked information is determined by using the boundary area without marked information and the trachea with marked information.

[0055] In step S1033, the label information of the specific position is used as the second input without sampling.

[0056] In step S1034 , a classifier is generated using the first input and the second input.

[0057] Use below Figure 6 to Figure 8 Let's explain steps S1031 to S1034.

[0058] like Figure 6 As shown, in step S1031, the coarse segmentation results (non-boundary areas) with label information are uniformly sampled, and the sampled voxels are divided into point sets according to the lung segments to which they belong. Although the voxels belonging to the right 1 lung segment have different spatial three-dimensional coordinates, they all have the labeling information of the right 1 segment. Although the voxels belonging to the right 3 lung segments have different spatial three-dimensional coordinates, they all have the labeling information of the right 3 segments. The above-mentioned sampled voxels with spatial three-dimensional position information and labeling information are the first input samples.

[0059] like Figure 7 As shown, in step S1032, the specific position and the marking information of the specific position are determined using the area to be segmented and the trachea with marking information. The specific position is the position where the trachea overlaps with the area to be segmented as the boundary area. The part of the trachea that overlaps with the area to be segmented is taken as the specific position, and the marking information of the trachea at this position is taken as the marking information of the specific position. Since the voxel at the specific position is the voxel at the position where the distance difference from the adjacent marking area is less than the threshold, it is important for distinguishing two adjacent marking areas. Therefore, when used as an input sample, it is not sampled and all voxels are used as they are. In step S1033, all voxels at the specific position with spatial position information and marking information are used as the second input.

[0060] like Figure 8 As shown, in step S1034, a classifier is generated using the first input and the second input obtained in step S1031 and steps S1032 and S1033. As the generated classifier, Figure 8The effect of these multiple hyperplanes is that the label information belonging to the same category is located on the same side of the hyperplane, and the label information on both sides of the hyperplane belongs to different categories. Since the hyperplane is determined by input with spatial three-dimensional position information and label information, each space divided by the hyperplane corresponds to a unique and determined category. In addition, here Figure 8 The figure on the right also shows the trachea segmented by the hyperplane, but the trachea is only illustrated to illustrate the effect of the hyperplane and is not part of the generated classifier.

[0061] Then, using the generated classifier, Figure 7 The region to be segmented shown on the upper side is divided into 18 lung segments with different label information. The segmentation result of the region to be segmented and the segmentation result of the non-boundary region are taken as the final segmentation result.

[0062] Fig. 9 The left side of shows the final segmentation result obtained by using the image segmentation method of embodiment 1. It can be clearly seen that the boundary of the above segmentation result is smooth. In contrast, Fig. 9 The right side shows the final segmentation result when only distance transform is used to segment the image. It can be seen that the boundaries of the segmented substructures are jagged and the doctor's observation is not utilized.

[0063] By adopting the image segmentation device and the image segmentation method of the present embodiment, a segmentation result with smooth boundaries can be obtained. In addition, compared with a method in which a user marks feature points of boundaries and performs surface fitting, the marking method of the present invention is more intuitive.

[0064] (Second Embodiment)

[0065] The difference between the second embodiment and the first embodiment is that the image segmentation device 20 further includes an input unit 25 for the user to input empirical annotations.

[0066] Specifically, the image segmentation device 20 is as follows: Fig.10 As shown, the system includes an acquisition unit 21 , a rough segmentation unit 22 , a classifier generation unit 23 , a fine segmentation unit 24 , an input unit 25 , and an output unit 26 .

[0067] The functions of the acquisition unit 21, the rough segmentation unit 22, the classifier generation unit 23, the fine segmentation unit 24, and the output unit 26 are the same as those of the acquisition unit 11, the rough segmentation unit 12, the classifier generation unit 13, the fine segmentation unit 14, and the output unit 15 of the first embodiment, and therefore, repeated descriptions are omitted. In the following, only the input unit 25 will be described.

[0068] like Fig.12As shown in the top figure of , sometimes in the final segmentation result obtained by the image segmentation method of embodiment 1, there is a situation where the segmented boundary does not fit the vein. However, in practice, it is desired to obtain a segmentation result in which the segmented boundary fits the vein. Therefore, it is necessary to correct the final segmentation result.

[0069] The input unit 25 may be an input interface for the user to input the empirical mark. By using the input unit 25, the user can artificially add the empirical mark. The user can use the input unit 25 at different stages of image segmentation. For example, after the final segmented image is generated, the user can add the empirical mark at the boundary position that does not meet the expectation. Fig.12 By replacing the original labels with the added empirical labels, a better segmentation result can be obtained. Alternatively, the empirical labels can be input through the input unit 25 while obtaining the lung lobe image with label information and the trachea image with label information.

[0070] By using the second embodiment, human intervention can be applied to the automatic segmentation process of the image, and more accurate segmentation can be achieved by adjusting a small local position.

[0071] Below, refer to Fig.11 The second embodiment is described with reference to the flowchart shown.

[0072] In step S201, at least the lung lobes and trachea with label information are acquired. In addition to the lung lobes and trachea with label information, empirical labels may be inputted via the input unit 25 in step S201.

[0073] In step S202, at least the label information of the lung lobe and the trachea is used to generate a boundary region without label information through distance transformation, and the remaining region is used as a non-boundary region with label information. The empirical label obtained in step S201 can also be used in step S202.

[0074] In step S203, a classifier is generated using the label information of the non-boundary area and the label information of the specific position. As mentioned above, the specific position is the voxel at the position where the distance difference from the adjacent label area is less than the threshold, which is important for distinguishing two adjacent label areas, and the empirical label is the voxel of the artificially added important area, and the two are essentially the same. Therefore, the label information of the specific position in step S203 includes the artificially added empirical label in addition to the label information of the specific position determined by the boundary area and the trachea.

[0075] Step S204 and step S205 are the same as those in the first embodiment, and their description is omitted here.

[0076] The second embodiment further includes step S206, where after the final segmentation result is output and displayed on a display, the user can input an experience tag via the input unit 25. Furthermore, the input experience tag can be reused to generate a classifier, the generated classifier can be updated, and the final segmentation result can be re-determined and output using the updated classifier (hyperplane).

[0077] Thus we get Fig.12 The boundary shown in the lower section fits the segmentation result of the vein.

[0078] By using the image segmentation device and the image segmentation method of the second embodiment, it is possible to easily correct the unexpected local segmentation results, which not only saves the user's time for overall correction, but also improves the accuracy of image segmentation.

[0079] (Other embodiments)

[0080] In the above embodiments, the image segmentation device and the image segmentation method are described, but it can be understood that as long as the functions of the image segmentation method can be realized, the above functions can also be realized by hardware or by a combination of hardware and software.

[0081] For example, when the image segmentation method is implemented by a processing circuit, i.e., a processor, the various processing functions of the processor are stored in the storage circuit in the form of programs that can be executed by a computer. The processor implements the functions corresponding to the various programs by reading out and executing the various programs from the storage circuit. In addition, instead of storing the programs in the storage circuit, it is also possible to configure the programs to be directly loaded into the circuit of the processor. In this case, the processor implements the functions by reading out and executing the programs loaded into the circuit. In addition, the processor of the above-mentioned embodiment is not limited to being configured as a single circuit, and multiple independent circuits can also be combined to form a processor to implement its functions.

[0082] The term "processor" used in the above description means, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or an application-specific integrated circuit (ASIC), a programmable logic device (for example, a simple programmable logic device (SPLD), a complex programmable logic device (CPLD), and a field programmable gate array (FPGA)) and other circuits.

[0083] The program executed by the processor may be provided by pre-loading it into a ROM (Read Only Memory) or a storage circuit, etc. Alternatively, the program may be provided by recording it in a computer-readable storage medium such as a CD (Compact Disc)-ROM, FD (Flexible Disk), CD-R (Recordable), DVD (Digital Versatile Disc), etc., in a form that can be installed in these devices or in a form that can be executed.

[0084] Alternatively, the program may be stored in a computer connected to a network such as the Internet, and provided or distributed by being downloaded via the network.

[0085] According to at least one of the above-described embodiments, compared with the case where only distance transformation is used to segment an image, a smooth boundary segmentation result can be obtained. In addition, compared with the method in which the user marks the feature points of the boundary and performs surface fitting, the operator can obtain a more intuitive experience. In addition, since empirical markings can be added at different stages, a segmentation result with higher accuracy can be easily obtained.

[0086] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other ways, and various omissions, substitutions, and changes can be made without departing from the scope of the subject matter of the invention. These embodiments and their variations are included in the scope and subject matter of the invention, and are also included in the invention described in the claims and their equivalents.

Claims

1. An image segmentation device, comprising: An acquisition unit, which acquires a block region with label information and a tubular region with label information; The rough segmentation unit uses the obtained label information of the block region and the tubular region to generate a to-be-segmented region as a boundary region of each substructure without label information and a non-boundary region as a rough segmentation result with label information through distance transformation; A classifier generating unit generates a classifier for classifying spatial coordinates by using the label information of the non-boundary region and the label information of the specific position determined by the region to be segmented and the tubular region; as well as The fine segmentation unit uses the above classifier to segment the voxels of the above area to be segmented and determines the final segmentation result.

2. The image segmentation device according to claim 1, The block-shaped region is a region that can anatomically cover the tubular region, the block-shaped region obtained by the acquisition unit is segmented into a first number of segments, and each segment is marked with different labeling information as an independent labeling region, and the tubular region obtained by the acquisition unit is segmented into a second number of segments, and each segment is marked with different labeling information as an independent labeling region, The non-boundary area generated by the rough segmentation unit is divided into a second number of segments, and each segment is labeled with different label information as an independent label area.

3. The image segmentation device according to claim 1 or 2, The above-mentioned classifier generation unit uniformly samples the above-mentioned non-boundary area according to different labeling information, takes each sampled voxel as the first input, and takes the labeling information of the specific position determined by the above-mentioned area to be segmented and the above-mentioned tubular area with labeling information as the second input without sampling, to generate a classifier.

4. The image segmentation device according to claim 1, The block-shaped regions are lung lobes, and the tubular regions are trachea or blood vessels.

5. The image segmentation device according to claim 4, The non-boundary region is divided into a plurality of label regions corresponding to the branching structure of the tubular region by distance transformation.

6. The image segmentation device according to claim 5, The tubular region and the non-border region each have corresponding marking information in their marking regions.

7. The image segmentation device according to claim 1, An output unit is further provided to output the final segmentation result.

8. The image segmentation device according to claim 3, The rough segmentation unit determines a voxel whose distance difference from two adjacent and different labeling regions is less than a threshold value as the region to be segmented.

9. The image segmentation device according to claim 1, The specific positions are all voxels in the tubular region that overlap with the region to be segmented.

10. The image segmentation device according to claim 1, The above classifier is a hyperplane that divides voxels based on spatial three-dimensional position information and label information.

11. The image segmentation device according to claim 10, There are multiple hyperplanes mentioned above. For each hyperplane, voxels on the same side have the same label information, and voxels on different sides have different label information.

12. The image segmentation device according to claim 11, The above-mentioned fine segmentation unit uses the above-mentioned hyperplane to determine the labeling information of each voxel in the above-mentioned area to be segmented based on the spatial three-dimensional position information of each voxel in the above-mentioned area to be segmented, and uses the above-mentioned labeling information of each voxel in the above-mentioned area to be segmented together with the above-mentioned rough segmentation result as the above-mentioned final segmentation result.

13. The image segmentation device according to claim 1, The device also has an input unit, which is an input interface for users to add experience marks.

14. The image segmentation device according to claim 13, The empirical label input via the input unit is also used as label information of the specific position to generate a classifier.

15. An image segmentation method comprising the following steps: An acquisition step of acquiring a block region with marked information and a tubular region with marked information; A rough segmentation step, using the obtained label information of the block region and the tubular region, through distance transformation, generates a to-be-segmented region without label information as a boundary region of each substructure, and a non-boundary region with label information as a rough segmentation result; a classifier generating step, using the label information of the non-boundary area and the label information of the specific position determined by the area to be segmented and the tubular area to generate a classifier for classifying the spatial coordinates; as well as In the fine segmentation step, the voxels in the above-mentioned area to be segmented are segmented using the above-mentioned classifier to determine the final segmentation result.

16. The image segmentation method according to claim 15, In the above rough segmentation step, through distance transformation, the voxels whose distance difference between two adjacent and different marked areas is less than the threshold are determined as the above-mentioned area to be segmented and no marking information is assigned. Then, the remaining area is determined as the above-mentioned non-boundary area and the original marking information is retained.

17. A storage medium storing a program for causing a computer to execute the following steps: An acquisition step of acquiring a block region with marked information and a tubular region with marked information; A rough segmentation step, using the obtained label information of the block region and the tubular region, through distance transformation, generates a to-be-segmented region without label information as a boundary region of each substructure, and a non-boundary region with label information as a rough segmentation result; a classifier generating step, using the label information of the non-boundary area and the label information of the specific position determined by the area to be segmented and the tubular area to generate a classifier for classifying the spatial coordinates; as well as In the fine segmentation step, the voxels in the above-mentioned area to be segmented are segmented using the above-mentioned classifier to determine the final segmentation result.

18. The storage medium according to claim 17, In the above rough segmentation step, through distance transformation, the voxels whose distance difference between two adjacent and different marked areas is less than the threshold are determined as the above-mentioned area to be segmented and no marking information is assigned. Then, the remaining area is determined as the above-mentioned non-boundary area and the original marking information is retained.

Images