Blood vessel segmentation device, blood vessel segmentation method, medical image processing apparatus, and deep learning model generation device

Abstract

To accurately perform blood vessel segmentation to improve connectivity of blood vessels in a segmentation result even when a contrast between blood vessels and tissues is not high in a medical image.SOLUTION: A blood vessel segmentation device according to the present embodiment comprises an image acquisition unit, a deep learning model acquisition unit, and a segmentation unit. The image acquisition unit acquires a plurality of medical images. The deep learning model acquisition unit acquires, on the basis of the medical images, a trained deep learning model obtained by performing deep learning model training using a center line emphasis loss function and a statistical shape restriction loss function. The segmentation unit performs segmentation of an acquired target medical image on the basis of the trained deep learning model to obtain a segmentation result related to blood vessels.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The embodiments disclosed in this specification and the drawings relate to a blood vessel segmentation device, a blood vessel segmentation method, a medical image processing device, and a deep learning model generation device. [Background technology]

[0002] In medical clinical research and applications, 3D reconstruction of blood vessels based on medical images acquired by a medical imaging diagnostic device can help doctors quickly locate the positions of organs, tissues, tumors, etc. For example, accurate 3D reconstruction of liver blood vessels can help doctors quickly identify liver tumors. Therefore, fast and accurate segmentation of blood vessels is important in medical image processing.

[0003] With the development of computer technology, algorithms used for vessel segmentation include those based on vessel enhancement filtering, deformable models, tracking, and deep learning.

[0004] The vessel enhancement filtering algorithm enhances the vessels in the original image, but contains noise.

[0005] In addition, the algorithm based on the deformable model requires seed point initialization. For example, in the algorithm based on the deformable model, a point with a high gray level value after blood vessel enhancement is used as an initial seed point, and then blood vessel segmentation is performed by combining algorithms such as the watershed algorithm and the region growing method. However, the initialization of the seed point affects the accuracy of blood vessel segmentation and involves complex calculations.

[0006] In addition, the tracking algorithm needs to use multiple pre-defined templates and combine the initial position and tracking direction to achieve vessel segmentation. The process of multiple pre-defined templates is complicated.

[0007] In addition, deep learning algorithms achieve vascular segmentation by classifying the original image pixel by pixel. However, deep learning algorithms create training data (e.g., Ground Truth; GT) by performing annotation to draw the segmentation target on the image, so they depend on the amount and quality of annotation data. In addition, for example, when a tumor that a doctor focuses on is the liver, it is difficult to obtain highly accurate training data by performing vascular annotation because the contrast between liver blood vessels and liver tissue is low in medical images. Therefore, it is difficult for deep learning algorithms to achieve high-precision vascular segmentation, and the connectivity of blood vessels in the segmentation results is poor.

[0008] Thus, algorithms used for vessel segmentation need to be improved in terms of vessel segmentation accuracy and to reduce poor vessel connectivity due to vessel and tissue contrast (e.g., a situation in which the vessels are actually connected but not connected in the segmentation results). [Prior art documents] [Non-patent literature]

[0009] [Non-Patent Document 1] Ruochen Gao, Zhihui Hou, Jun Li, Hu Han, Bin Lu, S. Kevin Zhou, "Joint Coronary Centerline Extraction And Lumen Segmentation From Ccta Using Cnntracker And Vascular Graph Convolutional Network" Published in: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), Page:1897-1901, 13 April 2021. Summary of the Invention [Problem to be solved by the invention]

[0010] One of the problems to be solved by the embodiments disclosed in this specification and the drawings is to perform blood vessel segmentation with high accuracy and improve the connectivity of blood vessels in the segmentation results even when the contrast between blood vessels and tissues is not high in medical images. However, the problems to be solved by the embodiments disclosed in this specification and the drawings are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described later can also be positioned as other problems. [Means for solving the problem]

[0011] The blood vessel segmentation device according to the present embodiment includes an image acquisition unit, a deep learning model acquisition unit, and a segmentation unit. The image acquisition unit acquires a plurality of medical images. The deep learning model acquisition unit acquires a trained deep learning model obtained by training a deep learning model using a centerline enhancement loss function and a statistical shape constraint loss function based on the medical images. The segmentation unit segments the acquired target medical image based on the trained deep learning model to obtain a segmentation result related to blood vessels. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a block diagram showing the configuration of a blood vessel segmentation apparatus according to the first embodiment. [Diagram 2] Figure 2 shows a schematic diagram of deep learning model training. [Diagram 3] FIG. 3 is a flowchart showing the procedure of processing (vessel segmentation method) performed by the vessel segmentation device according to the first embodiment. [Figure 4A] FIG. 4A is a diagram showing an example of an image used in the blood vessel segmentation device according to the first embodiment. [Figure 4B]FIG. 4B is a diagram showing an example of an image used in the blood vessel segmentation device according to the first embodiment. [Figure 4C] FIG. 4C is a diagram showing an example of an image used in the blood vessel segmentation device according to the first embodiment. [Diagram 5] FIG. 5 is a block diagram showing the configuration of a blood vessel segmentation apparatus according to the second embodiment. [Figure 6] FIG. 6 is a flowchart showing the procedure of processing (vessel segmentation method) performed by the vessel segmentation device according to the second embodiment. [Figure 7] FIG. 7 is a schematic diagram of a process of performing a region growing method in the vessel segmentation apparatus according to the second embodiment. [Figure 8] FIG. 8 is a diagram showing an example of an image used in the blood vessel segmentation apparatus according to the second embodiment. [Figure 9] FIG. 9 is a block diagram showing the configuration of a blood vessel segmentation apparatus according to the third embodiment. [Figure 10] FIG. 10 is a schematic diagram for explaining the process (connectivity restoration process) in which the image region enhancement unit restores the connectivity of the segmentation result. [Figure 11] FIG. 11 is a flowchart showing the procedure of processing (vessel segmentation method) performed by the vessel segmentation device according to the third embodiment. [Figure 12] FIG. 12 is a diagram showing an example of an image used in the blood vessel segmentation device according to the third embodiment. [Figure 13] FIG. 13 is a block diagram showing the configuration of a blood vessel segmentation apparatus according to the fourth embodiment. [Figure 14] FIG. 14 is a flowchart showing the procedure of processing (vessel segmentation method) performed by the vessel segmentation device according to the fourth embodiment. [Figure 15] FIG. 15 is a diagram showing an example of an image used in the blood vessel segmentation device according to the fourth embodiment. [Figure 16]FIG. 16 is a block diagram showing the configuration of a blood vessel segmentation apparatus according to the fifth embodiment. [Figure 17] FIG. 17 is a flowchart showing the procedure of processing (vessel segmentation method) performed by the vessel segmentation device according to the fifth embodiment. [Figure 18] FIG. 18 is a block diagram showing the configuration of a blood vessel segmentation apparatus according to the sixth embodiment. [Figure 19] FIG. 19 is a flowchart showing the procedure of processing (vessel segmentation method) performed by the vessel segmentation device according to the sixth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, the blood vessel segmentation device, blood vessel segmentation method, medical image processing device, and deep learning model generation device according to the present embodiment will be described with reference to the drawings.

[0014] (First embodiment) <Configuration of blood vessel segmentation device 100> First, the configuration of a blood vessel segmentation device 100 according to the first embodiment will be described with reference to Fig. 1. Note that, although the blood vessel segmentation device includes various processing functions in Fig. 1, only the processing functions related to this embodiment are shown in Fig. 1, and other processing functions are omitted.

[0015] As shown in FIG. 1, the blood vessel segmentation device 100 of this embodiment includes an image acquisition unit 101, a deep learning model acquisition unit 102, and a segmentation unit 103. For example, the blood vessel segmentation device 100 is a computer, and includes a processing circuit that executes each processing function, and a storage circuit that stores each processing function in the form of a program executable by a computer. For example, the processing circuit of the blood vessel segmentation device 100 includes an image acquisition unit 101, a deep learning model acquisition unit 102, and a segmentation unit 103. Each processing function executed by the components of the processing circuit is recorded in the storage circuit, for example, in the form of a program executable by a computer. The processing circuit is a processor that reads each program from the storage circuit and executes it to realize a processing function corresponding to each program.

[0016] The term "processor" used in the above description means a circuit such as a central processing unit (CPU), a micro controller unit (MCU), a graphics processing unit (GPU), an application specific integrated circuit (ASIC), a programmable logic device (e.g., a simple programmable logic device (SPLD), a complex programmable logic device (CPLD), and a field programmable gate array (FPGA)). When the processor is a CPU, for example, the processor realizes a function by reading and executing a program stored in a memory circuit. On the other hand, when the processor is an ASIC, for example, instead of storing a program in a memory circuit, a program is directly incorporated in the circuit of the processor. Note that each processor in this embodiment is not limited to being configured as a single circuit for each processor, but may be configured as one processor by combining multiple independent circuits to realize its function. Furthermore, multiple components may be integrated into one processor to realize its function.

[0017] The image acquisition unit 101 acquires a plurality of medical images. For example, the image acquisition unit 101 acquires a plurality of medical images from a medical image diagnostic device through a communication interface. Alternatively, the plurality of medical images may be stored in a storage device in advance, and the image acquisition function 101 may read and acquire the plurality of medical images from the storage device. As an example of the storage device, the medical images may be managed in a medical image management system (PACS: Picture Archiving and Communication System) or an image storage device. Furthermore, the image acquisition function 101 may read a plurality of medical images processed by another device.

[0018] Here, the medical image is generated by a medical image diagnostic device. Examples of the medical image diagnostic device include an ultrasound diagnostic device, which is an ultrasound system, and an X-ray CT (Computed Tomography) device and an MRI (Magnetic Resonance Imaging) device, which are modalities other than the ultrasound system. That is, the medical image may be a three-dimensional image or a two-dimensional image.

[0019] The deep learning model acquisition unit 102 acquires a trained deep learning model obtained by training the deep learning model based on a plurality of medical images. The deep learning model acquisition unit 102 acquires a trained deep learning model obtained by training the deep learning model using at least a centerline enhancement loss function and a statistical shape constraint loss function. Specifically, the deep learning model acquisition unit 102 acquires a trained deep learning model obtained by training the deep learning model using an integrated loss function obtained by weighting the Dice coefficient loss function, the cross entropy loss function, the centerline enhancement loss function, and the statistical shape constraint loss function, for example.

[0020] FIG. 2 is a schematic diagram of deep learning model training. As shown in FIG. 2, deep learning model training is performed based on teacher data and multiple medical images acquired by an image acquisition unit 101. The teacher data is annotated in advance and acquired by the image acquisition function 101. Hereinafter, in this embodiment, the teacher data will be referred to as GT (Ground Truth), annotation GT, etc. In deep learning model training, a medical image is input to a deep learning network as an input image (input volume), and a prediction result (Predication) is acquired. The value L of the loss function (LOSS) total is used to measure the degree of deviation between the prediction result (Predication) and GT (Ground Truth).

[0021] In this embodiment, a trained deep learning model obtained by training a deep learning model by introducing a centerline enhancement loss function and a statistical shape constraint loss function is used.

[0022] The centerline enhancement loss function is a function for evaluating the similarity between two elements (the centerline structure of the blood vessel in the prediction result and the centerline structure of the blood vessel in the annotation GT) while taking into account the topology. In this embodiment, the segmentation result of a thin blood vessel due to branching can be improved by segmenting a tubular object (blood vessel) while maintaining the topological structure.

[0023] The statistical shape constraint loss function is a function that constrains the length and volume of the blood vessel contour as the shape of the blood vessel. In this embodiment, the geometric topology structure can be enhanced by segmenting the tubular object (blood vessel) while taking into account geometric information.

[0024] In this manner, in this embodiment, by using a trained deep learning model obtained by training a deep learning model by introducing a centerline enhancement loss function and a statistical shape constraint loss function, it is possible to enhance the connectivity and geometric topology of blood vessels in the segmentation results.

[0025] In this embodiment, for example, a trained deep learning model obtained by training a deep learning model using an integrated loss function obtained by weighting the Dice coefficient loss function, the cross entropy loss function, the centerline enhancement loss function, and the statistical shape constraint loss function is used. The integrated loss function L total is obtained by the following equation 1.

[0026]

number

[0027] In formula 1, L total represents the loss function (LOSS), and L DICE represents the Dice coefficient loss function, and L CE represents the cross-entropy loss function, and L CL represents the centerline enhancement loss function, and L AC represents the statistical shape constraint loss function.

[0028] Here, in this embodiment, the Dice coefficient loss function L DICE , the cross-entropy loss function L CE , the centerline enhancement loss function L CL , the statistical shape constraint loss function L AC The weights of are set to 0.3:0.3:0.3:1, respectively.

[0029] The centerline enhancement loss function L in Eq. CL is obtained by the following equation 2.

[0030]

number

number

[0031] In Equation 2, V L represents the hepatic blood vessels in the annotation GT (teaching data), and V p represents the liver blood vessels in the predicted results, and S L represents the centerline structure of the blood vessel in the annotation GT (teaching data), and S P represents the centerline structure of the blood vessel in the predicted result. prec (S P ,V L ) is the centerline structure of the blood vessel in the predicted result S P However, the liver vascular V in annotation GT L It also represents the percentage of T sens (S L ,V P ) is the centerline structure of the blood vessel in the annotation GT. L However, the predicted liver vascular V p This represents the percentage contained in

[0032] In Equation 2, the centerline structure S of the blood vessel in the predicted result is P and the centerline structure of the blood vessel in annotation GT, S LAs described above, the center line emphasis loss function L obtained by Equation 2 is calculated. CL takes into account topology and uses the centerline structure of blood vessels in the training data for blood vessel segmentation (blood vessel annotation GT). Therefore, in this embodiment, by segmenting tubular objects (blood vessels) while maintaining the topology structure, it is possible to improve the segmentation results of thin blood vessels due to branching.

[0033] The statistical shape constraint loss function L in Eq. AC is obtained by the following equation 4.

[0034]

number

number

[0035] In Equation 4, c1 and c2 respectively represent the average values ​​inside and outside the segmentation target contour C, where f(x) represents the pixel intensity of the image at position x, Length(C) represents the length of the segmentation contour C, and λ represents a parameter for balancing each weight.

[0036] In Equation 4, the length and volume of the blood vessel contour are calculated as the blood vessel shape. The statistical shape constraint loss function L obtained by Equation 4 ACregards segmentation as an energy minimization problem, where the energy is a function related to the contour. Here, the statistical shape constrained loss function minimizes the energy function by using a method based on partial differentiation to constantly update the contour toward the direction of the target boundary. In this case, the curve position when the energy is minimized is the target contour, and the relationship between the contour and the volume of the segmentation is taken into account. Thus, in this embodiment, by introducing the statistical shape constrained loss function, a priori knowledge of the shape can be added to the normalized part to make it fit better to the shape features of the segmentation target, thereby improving the image segmentation result. That is, in this embodiment, the geometric topological structure can be emphasized by segmenting tubular objects (blood vessels) while taking into account geometric information as the shape of the blood vessel.

[0037] The Dice coefficient loss function L in Eq. DICE is a function that uses the Dice coefficient to evaluate the similarity between two elements, and calculates, for example, how well the prediction result can be detected for the training data (vascular annotation GT) of blood vessel segmentation.

[0038] The cross-entropy loss function L in Eq. CE is an index that indicates how far apart two probability distributions are, and for example, it is used to calculate how far the predicted result is from the training data for blood vessel segmentation (blood vessel annotation GT).

[0039] In addition, the Dice coefficient loss function L in Eq. DICE and the cross-entropy loss function L CE Since this is an existing technique, detailed description thereof will be omitted in this embodiment.

[0040] In this embodiment, as described above, the integrated loss function L obtained by Equation 1 is totalThe acquired target medical image is segmented based on the trained deep learning model obtained by training the deep learning model using the Dice coefficient loss function L DICE and the cross-entropy loss function L CE In the case where only the centerline enhancement loss function L is used, only information related to the gradation or texture is taken into account, and therefore geometric information is not taken into account, and the connectivity and geometric topology of blood vessels in the segmentation result are not emphasized. On the other hand, in this embodiment, a centerline enhancement loss function L CL and the statistical shape constraint loss function L AC By using this method, we can make up for the deficiencies in existing techniques by taking into account geometric information, and can emphasize the connectivity and geometric topology of blood vessels in the segmentation results.

[0041] Returning to Fig. 1, the segmentation unit 103 segments the acquired target medical image based on the trained deep learning model acquired by the deep learning model acquisition unit 102, thereby obtaining a segmentation result related to blood vessels.

[0042] Here, the target medical image may be acquired by, for example, the image acquisition unit 101, or may be acquired by another device.

[0043] <Operation of the Blood Vessel Segmentation Apparatus 100> Next, a description will be given of the processing (blood vessel segmentation method) performed by the blood vessel segmentation device 100 according to the first embodiment. Fig. 3 is a flowchart showing the procedure of the processing (blood vessel segmentation method) performed by the blood vessel segmentation device 100 according to the first embodiment.

[0044] 3, the image acquisition unit 101 acquires a plurality of medical images. The plurality of medical images are, for example, images of the liver.

[0045] Next, annotation GT (teaching data) is acquired. In step S200 of Fig. 3, the deep learning model acquisition unit 102 acquires a trained deep learning model obtained by training the deep learning model based on the multiple medical images acquired by the image acquisition unit 101 in step S100 and the teaching data. Here, in the deep learning model training, a centerline enhancement loss function and a statistical shape constraint loss function are introduced.

[0046] Next, a target medical image as shown in FIG. 4A is acquired. In step S300 of FIG. 3, the segmentation unit 103 segments the acquired target medical image based on the trained deep learning model acquired by the deep learning model acquisition unit 102, thereby obtaining a segmentation result shown in FIG. 4C as data related to blood vessels. Here, it is assumed that the acquired target medical image is similar to a part of the multiple medical images acquired by the image acquisition unit 101. For example, in the segmentation result of the existing technology shown in FIG. 4B, the part hatched with a mesh or diagonal lines indicates a blood vessel region. Here, in the segmentation result shown in FIG. 4C, the part hatched with a mesh or diagonal lines is a blood vessel segmented by the segmentation unit 103, and the part further segmented with respect to the segmentation result of the existing technology shown in FIG. 4B is indicated by a solid line frame.

[0047] A comparison is made between the target medical image shown in Fig. 4A, the segmentation result of the existing technology shown in Fig. 4B, and the segmentation result shown in Fig. 4C. Here, in the segmentation result shown in Fig. 4C, blood vessels in the target medical image shown in Fig. 4A are segmented, and more blood vessels are segmented than in the segmentation result of the existing technology shown in Fig. 4B.

[0048] As described above, the blood vessel segmentation device 100 according to this embodiment uses a trained deep learning model obtained by training a deep learning model by introducing a centerline enhancement loss function and a statistical shape constraint loss function. As a result, the blood vessel segmentation device 100 according to this embodiment can segment more blood vessels, perform blood vessel segmentation with high accuracy, and improve the connectivity of blood vessels in the segmentation result, even when the contrast between blood vessels and tissues is not high in a medical image.

[0049] Second embodiment <Configuration of blood vessel segmentation device 100A> First, the configuration of a blood vessel segmentation device 100A according to the second embodiment will be described with reference to Fig. 5. Fig. 5 is a block diagram showing the configuration of the blood vessel segmentation device according to the second embodiment.

[0050] As shown in FIG. 5, the blood vessel segmentation device 100A according to this embodiment includes an image acquisition unit 101, a deep learning model acquisition unit 102, and a segmentation unit 103, and further includes a feature region enhancement unit 104.

[0051] Here, the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the second embodiment are the same as the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the first embodiment, respectively. Therefore, in this embodiment, overlapping descriptions will be omitted.

[0052] In the second embodiment, the feature region enhancement unit 104 extracts tubular object features from the segmentation result by the segmentation unit 103 using vesselness filtering, obtains feature regions of the enhanced tubular object features, and connects the tubular object features using a region growing method to obtain the segmentation result after the feature region enhancement.

[0053] Specifically, first, the feature region highlighting unit 104 extracts tubular object features from the segmentation result by, for example, using a Hessian matrix as tubular object filtering. This is because the feature value of the Hessian matrix can measure information about a tubular structure. However, the method of extracting tubular object features is not limited to the method using the Hessian matrix, and various methods in the publicly known art can be applied, and detailed description thereof will be omitted in this embodiment.

[0054] Next, the characteristic region highlighting unit 104 connects the blood vessel cut points that are actually connected but are determined to be cut in the segmentation result by the segmentation unit 103, by using a region growing method.

[0055] For example, the feature region enhancing unit 104 uses a tubular object metric to describe the probability of whether a voxel belonging to an image is tubular or not. By using the tubular object metric, the feature region enhancing unit 104 can identify blood vessels that are difficult to distinguish from tissues (e.g., solid parts other than blood vessels in an organ) in an image due to a relatively low contrast with the tissue. For example, the feature region enhancing unit 104 assumes that the diameter of the target blood vessel is approximately 0.5 mm to 6 mm, selects 0.5 mm as the step size, and uses a Hessian matrix to analyze and extract second-order gradient information of the image that is the segmentation result of the segmentation unit 103. That is, when a calculation based on the Hessian matrix is ​​performed for each voxel, and there are only two relatively large feature values ​​in the calculation result based on the Hessian matrix, the voxel is considered to belong to a tubular structure. Here, the relatively large feature value can be appropriately set according to, for example, an organ, an image, etc. Note that the above-mentioned settings of the diameter of the target blood vessel and the measurement value as the step size are merely examples, and are not limited to these, and can be arbitrarily set as necessary.

[0056] In the region expanding method, the feature region highlighting unit 104 uses a point with a large probability value in the probability map (mask) included in the segmentation result of the segmentation unit 103 as a seed point, and performs region expanding in the tubular object feature region using, for example, 60 as a lower threshold, thereby repairing the severed blood vessels (the locations where the blood vessels were severed) and replenishing the microvessels.

[0057] <Operation of blood vessel segmentation device 100A> The following describes the process (blood vessel segmentation method) performed by the blood vessel segmentation device 100A according to the second embodiment. Fig. 6 is a flowchart showing the procedure of the process (blood vessel segmentation method) performed by the blood vessel segmentation device 100A according to the second embodiment.

[0058] First, the processes of steps S100 to S300 in Fig. 6 are executed. The processes of steps S100 to S300 in Fig. 6 are the same as the processes of steps S100 to S300 in Fig. 3, so detailed description will be omitted here. In addition, in this embodiment, the segmentation result further includes a probability map in addition to data related to blood vessels.

[0059] Next, in step S301 of FIG. 6, the feature region enhancement unit 104 extracts tubular object features from the segmentation result of the segmentation unit 103 in step S300 using tubular object filtering, obtains feature regions of the enhanced tubular object features, and connects the tubular object features using a region growing method to obtain the segmentation result after the feature region has been enhanced.

[0060] 7, in step S301, the feature region enhancing unit 104 extracts tubular object features using tubular object filtering with a Hessian matrix based on the medical image acquired by the image acquiring unit 101 in step S100 and the probability map included in the segmentation result of the segmentation unit 103 in step S300, and acquires a feature region of the enhanced tubular object features. Then, the feature region enhancing unit 104 performs a region growing method in the tubular object feature region using points with large probability values ​​in the probability map as seed points, and connects the tubular object features to repair the cut blood vessel (cut portion of the blood vessel) and replenish microvessels.

[0061] Here, the specific example of the first embodiment is also applied to the second embodiment. The target image shown in FIG. 4A, the segmentation result of the existing technology shown in FIG. 4B, the segmentation result of the first embodiment shown in FIG. 4C, and the segmentation result of the second embodiment shown in FIG. 8 are compared. Here, in the segmentation result of the second embodiment shown in FIG. 8, the parts hatched with mesh or diagonal lines are blood vessels segmented by the segmentation unit 103, and the parts further segmented with respect to the segmentation result of the existing technology shown in FIG. 4B are shown in solid line frames. In the segmentation result of the second embodiment shown in FIG. 8, the blood vessels of the target medical image shown in FIG. 4A are segmented, and more blood vessels are segmented than in the segmentation result of the existing technology shown in FIG. 4B. Furthermore, in the segmentation result of the second embodiment shown in FIG. 8, there are more parts shown in solid line frames than in the segmentation result of the first embodiment shown in FIG. 4C, and even fine blood vessels are segmented.

[0062] As described above, in the blood vessel segmentation device 100A according to the second embodiment, as in the first embodiment, more blood vessels are segmented as a segmentation result, so that blood vessel segmentation can be performed with high accuracy and blood vessel connectivity can be improved in the segmentation result. In addition, since the blood vessel segmentation device 100A according to the second embodiment includes the feature region highlighting unit 104, blood vessel cut points in the segmentation result by the deep learning model are repaired in the tubular object feature region based on the tubular object filtering and region growing method, and further, the segmentation quality of microvessels can be improved and microvessels can be replenished.

[0063] Third embodiment <Configuration of blood vessel segmentation device 100B> First, the configuration of a blood vessel segmentation device 100B according to the third embodiment will be described with reference to Fig. 9. Fig. 9 is a block diagram showing the configuration of the blood vessel segmentation device according to the third embodiment.

[0064] As shown in FIG. 9, the blood vessel segmentation device 100B of this embodiment includes an image acquisition unit 101, a deep learning model acquisition unit 102, and a segmentation unit 103, and further includes an image region enhancement unit 105.

[0065] Here, the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the third embodiment are the same as the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the first embodiment, respectively. Therefore, in this embodiment, overlapping descriptions will be omitted.

[0066] In the third embodiment, the image region enhancement unit 105 performs a connectivity restoration process on isolated blood vessel segments in the segmentation result in the image region of the segmentation result of the segmentation unit 103 to obtain the segmentation result after the image region enhancement. For example, the image region enhancement unit 105 performs connectivity restoration on the segmentation result of the segmentation unit 103 by applying a super voxel method.

[0067] Here, the process in which the image region enhancement unit 105 performs connectivity restoration on the segmentation result by applying the super voxel method will be described with reference to Fig. 10. Fig. 10 is a schematic diagram for explaining the process in which the image region enhancement unit 105 performs connectivity restoration on the segmentation result (connectivity restoration process). For convenience of explanation, only a part of the segmentation result is shown.

[0068] First, in the connectivity restoration process, the image region enhancement unit 105 performs isolated branch analysis on the segmentation result of the segmentation unit 103 to distinguish the vascular main trunk and isolated branches that are not connected in the segmentation result. For example, as shown in FIG. 10, the image region enhancement unit 105 determines the connected region with the largest volume in the segmentation result as the vascular main trunk M, and determines the vascular segments other than the vascular main trunk M that are not connected to the vascular main trunk and have a relatively small volume as isolated branches D. In FIG. 10, for convenience of explanation, only the vascular main trunk M and one isolated branch D are shown, but there may be multiple isolated branches. Since the process in which the image region enhancement unit 105 performs connectivity restoration for each isolated branch is the same, only the connectivity restoration process of the isolated branch D shown in FIG. 10 will be described. In addition, when there are multiple isolated branches, the image region enhancement unit 105 can perform connectivity restoration for each isolated branch sequentially or in parallel.

[0069] Next, in the connectivity restoration process, the image region enhancement unit 105 extracts a center line (thick line shown in FIG. 10) for the free branch D, and identifies two end points P1 and P2 (star marks shown in FIG. 10) of the center line of the free branch D. Then, the image region enhancement unit 105 calculates the distance between the two end points P1 and P2 and the vascular main trunk M, and sets the end point with the shortest distance as the near end point of the free branch D. Here, the closest distance (i.e., the shortest distance) from the two end points P1, P2 to the vascular main trunk M is defined as the distance between the end points and the vascular main trunk M. In the example shown in FIG. 10, the distance between the end point P2 and the vascular main trunk M is shorter than the distance between the end point P1 and the vascular main trunk M, so the end point P2 is set as the near end point of the free branch D.

[0070] Next, in the connectivity restoration process, the image region enhancement unit 105 performs supervoxel analysis to analyze supervoxels S from the free branch D to the main trunk blood vessel M. Then, based on the supervoxel analysis, the image region enhancement unit 105 connects the near end point P2 side of the free branch D to the main trunk blood vessel M. Note that the method of connecting the near end point P2 side of the free branch D to the main trunk blood vessel M based on the supervoxel analysis is an existing technology, so a detailed description thereof will be omitted in this embodiment.

[0071] <Operation of blood vessel segmentation device 100B> Next, a process (blood vessel segmentation method) performed by the blood vessel segmentation device 100B according to the third embodiment will be described. Fig. 11 is a flowchart showing the procedure of the process (blood vessel segmentation method) performed by the blood vessel segmentation device 100B according to the third embodiment.

[0072] First, the process of steps S100 to S300 in Fig. 11 is executed. Since the process of steps S100 to S300 in Fig. 11 is the same as the process of steps S100 to S300 in Fig. 3, a detailed description thereof will be omitted here.

[0073] Next, in step S302 of FIG. 11, the image region enhancement unit 105 performs connectivity restoration of isolated vascular segments in the image region as described with reference to FIG. 10 on the segmentation result of the segmentation unit 103 in step S300 to obtain a segmentation result after image region enhancement.

[0074] Here, the specific example of the first embodiment is also applied to the third embodiment. The target image shown in FIG. 4A, the segmentation result of the existing technology shown in FIG. 4B, the segmentation result of the first embodiment shown in FIG. 4C, and the segmentation result of the third embodiment shown in FIG. 12 are compared. Here, in the segmentation result of the third embodiment shown in FIG. 12, the parts hatched with mesh or diagonal lines are blood vessels segmented by the segmentation unit 103, and the parts further segmented with respect to the segmentation result of the existing technology shown in FIG. 4B are shown in solid line frames. In the segmentation result of the third embodiment shown in FIG. 12, the blood vessels of the target medical image shown in FIG. 4A are segmented, and more blood vessels are segmented than in the segmentation result of the existing technology shown in FIG. 4B. Furthermore, in the segmentation result of the third embodiment shown in FIG. 12, there are more parts shown in solid line frames than in the segmentation result of the first embodiment shown in FIG. 4C, and even fine blood vessels are segmented.

[0075] As described above, in the blood vessel segmentation device 100B according to the third embodiment, as in the first embodiment, more blood vessels are segmented as a segmentation result, so that blood vessel segmentation can be performed with high accuracy and the connectivity of blood vessels in the segmentation result can be improved. In addition, since the blood vessel segmentation device 100B according to the third embodiment includes an image region enhancement unit 105, an isolated segment analysis (above-mentioned isolated branch analysis) is performed on the blood vessel segmentation result based on the supervoxel method, and the connectivity of the cut blood vessel (blood vessel cut point) is restored in the image region according to the magnitude of the distance from the end point of the center line of the isolated segment to the main trunk. This makes it possible to solve the connectivity problem of an isolated blood vessel that does not have a protruding tubular structure, which cannot be processed by the feature region enhancement unit 104 in the second embodiment.

[0076] (Fourth embodiment) <Configuration of blood vessel segmentation device 100C> First, the configuration of a blood vessel segmentation device 100C according to the fourth embodiment will be described with reference to Fig. 13. Fig. 13 is a block diagram showing the configuration of the blood vessel segmentation device according to the fourth embodiment.

[0077] As shown in FIG. 13, the vascular segmentation device 100C of this embodiment includes an image acquisition unit 101, a deep learning model acquisition unit 102, and a segmentation unit 103, as well as a feature region enhancement unit 104 and an image region enhancement unit 105a.

[0078] Here, the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the fourth embodiment are the same as the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the first embodiment, respectively. Also, the characteristic region enhancement unit 104 in the fourth embodiment is the same as the characteristic region enhancement unit 104 in the second embodiment. Also, the difference between the image region enhancement unit 105a in the fourth embodiment and the image region enhancement unit 105 in the second embodiment is that the target processed by the image region enhancement unit 105a is not the segmentation result of the segmentation unit 103, but the segmentation result after the characteristic region is enhanced by the characteristic region enhancement unit 104. Here, overlapping explanations are omitted.

[0079] <Operation of blood vessel segmentation device 100C> Next, a process (blood vessel segmentation method) performed by the blood vessel segmentation device 100C according to the fourth embodiment will be described. Fig. 14 is a flowchart showing the procedure of the process (blood vessel segmentation method) performed by the blood vessel segmentation device 100C according to the fourth embodiment.

[0080] First, the processes of steps S100 to S301 in Fig. 14 are executed. The processes of steps S100 to S301 in Fig. 14 are the same as the processes of steps S100 to S301 in Fig. 6. Therefore, in this embodiment, a duplicated description will be omitted.

[0081] Next, in step S302 of FIG. 14, the image region enhancement unit 105 performs connectivity restoration of isolated vascular segments, as described with reference to FIG. 10, in the image region for the segmentation result processed by the feature region enhancement unit 104 in step S301, to obtain the segmentation result after image region enhancement.

[0082] Here, the specific example of the first embodiment is also applied to the fourth embodiment. The target image shown in FIG. 4A, the segmentation result of the existing technology shown in FIG. 4B, the segmentation result of the first embodiment shown in FIG. 4C, and the segmentation result of the fourth embodiment shown in FIG. 15 are compared. Here, in the segmentation result of the fourth embodiment shown in FIG. 15, the parts hatched with mesh or diagonal lines are blood vessels segmented by the segmentation unit 103, and the parts further segmented with respect to the segmentation result of the existing technology shown in FIG. 4B are shown in solid line frames. In the segmentation result of the fourth embodiment shown in FIG. 15, the blood vessels of the target medical image shown in FIG. 4A are segmented, and more blood vessels are segmented than in the segmentation result of the existing technology shown in FIG. 4B. Furthermore, in the segmentation result of the fourth embodiment shown in FIG. 15, there are more parts shown in solid line frames than in the segmentation result of the first embodiment shown in FIG. 4C, and even fine blood vessels are segmented.

[0083] As described above, in the blood vessel segmentation device 100C according to the fourth embodiment, as in the first embodiment, a larger number of blood vessels are segmented as a segmentation result, so that blood vessel segmentation can be performed with high accuracy and the connectivity of blood vessels in the segmentation result can be improved. In addition, since the blood vessel segmentation device 100C according to the fourth embodiment includes the feature region enhancement unit 104, it is possible to repair the cut part of blood vessels in the segmentation result and replenish microvessels. In addition, since the blood vessel segmentation device 100C according to the fourth embodiment includes the image region enhancement unit 105a, it is possible to solve the problem of connectivity of isolated blood vessels that do not have a protruding tubular structure and cannot be processed by the feature region enhancement unit 104.

[0084] Fifth embodiment <Configuration of blood vessel segmentation device 100D> First, the configuration of a blood vessel segmentation device 100D according to the fifth embodiment will be described with reference to Fig. 16. Fig. 16 is a block diagram showing the configuration of the blood vessel segmentation device according to the fifth embodiment.

[0085] As shown in FIG. 16, the blood vessel segmentation device 100D of this embodiment includes an image acquisition unit 101, a deep learning model acquisition unit 102, and a segmentation unit 103, and further includes a pre-processing unit 106.

[0086] Here, the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the fifth embodiment are the same as the image acquisition unit 101, the deep learning model acquisition unit 102, and the segmentation unit 103 in the first embodiment, respectively. Therefore, in this embodiment, overlapping descriptions will be omitted.

[0087] In the fifth embodiment, the preprocessing unit 106 performs preprocessing on a plurality of medical images acquired by the image acquisition unit 101 to obtain a standardized medical image.

[0088] Furthermore, in the fifth embodiment, the deep learning model acquisition unit 102 acquires a trained deep learning model obtained by training a deep learning model using a centerline enhancement loss function and a statistical shape constraint loss function based on the standardized medical image preprocessed by the preprocessing unit 106 and the annotation GT (teacher data).

[0089] The pre-processing performed by the pre-processing unit 106 on the multiple medical images includes, for example, at least one of data format conversion, cropping, resampling, and normalization.

[0090] For preprocessing by data format conversion, commonly used medical image formats include DICOM (with the extension ".dcm"), MHD (with the extension ".mhd" or ".raw"), and NIFTY (with the extension ".nii" or ".nii.gz"). These formats are not directly suitable for deep learning operations and require regular data format conversion.

[0091] In addition, crop preprocessing is to crop a 3D medical image to its non-zero region. Compared with the pre-cropping, the cropped medical image can reduce the image size, avoid unnecessary calculations, and improve computation efficiency without affecting the final segmentation result.

[0092] In addition, the resampling preprocessing aims to solve the problem that in some 3D medical image data sets, the actual spatial spacing represented by one voxel in different medical images does not match. To avoid such differences, it is necessary to perform a dimension reset in voxel space for different medical image data to ensure that the actual physical space represented by each voxel in the different image data is consistent.

[0093] In addition, the purpose of preprocessing by normalization is to ensure that the grayscale values ​​of each medical image in the training set have the same distribution.

[0094] Here, in this embodiment, the preprocessing unit 106 performs, for example, at least preprocessing by cropping and preprocessing by standardization.

[0095] <Operation of the blood vessel segmentation device 100D> Next, a process (vessel segmentation method) performed by the blood vessel segmentation device 100D according to the fifth embodiment will be described. Fig. 17 is a flowchart showing the procedure of the process (vessel segmentation method) performed by the blood vessel segmentation device 100D according to the fifth embodiment.

[0096] First, the process of step S100 in FIG. 17 is executed.

[0097] Next, in step S101 of FIG. 17, the preprocessing unit 106 performs, for example, at least cropping preprocessing and standardization preprocessing on the acquired medical image.

[0098] Next, the processes of steps S200 and S300 in FIG. 17 are the same as the processes of steps S200 and S300 in FIG. 3, so detailed description thereof will be omitted here.

[0099] Here, the blood vessel segmentation device 100D according to the fifth embodiment has the same effects as those of the first embodiment. In addition, since the blood vessel segmentation device 100D according to the fifth embodiment includes the pre-processing unit 106, blood vessel segmentation can be performed quickly and efficiently.

[0100] Sixth embodiment <Configuration of blood vessel segmentation device 100E> First, the configuration of a blood vessel segmentation device 100E according to the sixth embodiment will be described with reference to Fig. 18. Fig. 18 is a block diagram showing the configuration of the blood vessel segmentation device according to the sixth embodiment.

[0101] As shown in Figure 18, the vascular segmentation device 100E of this embodiment includes an image acquisition unit 101, a deep learning model acquisition unit 102, a segmentation unit 103, and a pre-processing unit 106, as well as a morphological processing unit 107.

[0102] Here, the image acquisition unit 101, the deep learning model acquisition unit 102, the segmentation unit 103, and the pre-processing unit 106 in the sixth embodiment are the same as the image acquisition unit 101, the deep learning model acquisition unit 102, the segmentation unit 103, and the pre-processing unit 106 in the fifth embodiment, respectively. Therefore, in this embodiment, overlapping descriptions will be omitted.

[0103] In the sixth embodiment, the morphological processing unit 107 performs morphological image processing on the segmentation result to obtain a final segmentation result. Here, the segmentation result may be the segmentation result of the segmentation unit 103, the segmentation result after the feature region enhancement in the second embodiment, the segmentation result after the image region enhancement in the third embodiment, or the segmentation result after the feature region enhancement and the image region enhancement in the fourth embodiment.

[0104] Image morphological processing includes, for example, erosion, dilation, open operation, close operation, skeleton extraction, polar erosion, hit-miss transformation, top-hat transformation, particle analysis, watershed transformation, morphological gradient, etc. Among them, the two most basic morphological operations are dilation and erosion, and the main functions of erosion and dilation are to remove noise, segment independent image elements and connect adjacent elements in the image, etc.

[0105] <Operation of the blood vessel segmentation device 100E> Next, a process (blood vessel segmentation method) performed by the blood vessel segmentation device 100E according to the sixth embodiment will be described. Fig. 19 is a flowchart showing the procedure of the process (blood vessel segmentation method) performed by the blood vessel segmentation device 100E according to the sixth embodiment.

[0106] First, the processes of steps S100, S101, S200, and S300 in Fig. 19 are executed. Here, the processes of steps S100, S101, S200, and S300 in Fig. 19 are the same as the processes of steps 100, S101, S200, and S300 in Fig. 17, so detailed explanations will be omitted here.

[0107] Next, in step S400 of FIG. 19, morphological image processing is performed on the segmentation result to obtain the final segmentation result.

[0108] Here, the blood vessel segmentation device 100E according to the sixth embodiment has the same effects as those of the fifth embodiment. In addition, the blood vessel segmentation device 100E according to the sixth embodiment includes a morphological processing unit 107, and thus can further remove small isolated blood vessels due to noise or the like in the segmentation result, thereby improving the accuracy of blood vessel segmentation.

[0109] The first to sixth embodiments have been described above. The above-described embodiments can be omitted, substituted, combined, and modified within the scope of the gist of the invention.

[0110] For example, in the sixth embodiment, the characteristic region enhancing section 104 and the image region enhancing section 105a in the fourth embodiment can be applied to configure the seventh embodiment.

[0111] Here, a comparative example will be described.

[0112] When the same specific example of hepatic vascular segmentation is applied to an existing technology based on deep learning (existing technology) and a technology based on deep learning using the integrated loss function in this embodiment (e.g., the fifth embodiment), the segmentation results are as shown in Table 1.

[0113] [Table 1]

[0114] In the example shown in Table 1, the centerline conformance of the hepatic vein and the portal vein is "0.795" and "0.868", respectively, when hepatic blood vessel segmentation is performed using the existing technology. Also, the centerline conformance of the hepatic vein and the portal vein is "0.814" and "0.875", respectively, when hepatic blood vessel segmentation is performed using the configuration of this embodiment (the fifth embodiment).

[0115] As is clear from Table 1, in the hepatic vessel segmentation in this embodiment (the fifth embodiment), both the hepatic vein and the portal vein have a higher centerline conformance than the existing technology. The centerline conformance is an index for characterizing the accuracy of the vessel segmentation, and the higher the centerline conformance, the higher the accuracy of the vessel segmentation.

[0116] Moreover, when the same specific example of hepatic blood vessel segmentation is applied to the fifth and seventh embodiments, the segmentation results are as shown in Table 2.

[0117] [Table 2]

[0118] In the example shown in Table 2, as a result of performing hepatic vascular segmentation using the configuration of the fifth embodiment, the numbers of unconnected regions other than the main vascular trees of the hepatic vein and portal vein are "1.4" and "0.7", respectively. Also, as a result of performing hepatic vascular segmentation using the configuration of the seventh embodiment, the numbers of unconnected regions other than the main vascular trees of the hepatic vein and portal vein are "0" and "0", respectively.

[0119] As can be seen from Table 2, when hepatic vascular segmentation is performed using the configuration of the fifth embodiment, the segmentation results for the hepatic vein and portal vein both contain unconnected regions other than the main vascular tree. Here, the unconnected regions other than the main vascular tree refer to regions that are actually connected but are unconnected in the segmentation results.

[0120] Furthermore, for example, when the seventh embodiment is configured by applying the feature region enhancing unit 104 and image region enhancing unit 105a of the fourth embodiment to the configuration of the sixth embodiment, and hepatic blood vessel segmentation is performed with the configuration of the seventh embodiment, no unconnected regions other than the main vascular tree exist in either the segmentation results of the hepatic vein or the portal vein. In other words, it has been realized that the feature region enhancing unit 104 and the image region enhancing unit 105a can perform blood vessel segmentation with high accuracy and improve the connectivity of blood vessels in the segmentation results, even when the contrast between blood vessels and tissues is not high in medical images.

[0121] In addition, the same specific example of hepatic vascular segmentation is applied to the fifth and seventh embodiments, and the statistics of the hepatic vein branching and portal vein branching related to the segmentation results are as shown in Table 3.

[0122] [Table 3]

[0123] In the example shown in Table 3, as a result of performing hepatic vascular segmentation using the configuration of the fifth embodiment, the annotation GT (teaching data) is "55" as a statistic of hepatic vein vascular branching, whereas the predicted result is "45", so the detection rate is "82%". Also, as a result of performing hepatic vascular segmentation using the configuration of the seventh embodiment, the annotation GT is "55" as a statistic of hepatic vein vascular branching, whereas the predicted result is "52", so the detection rate is "95%".

[0124] In the example shown in Table 3, as a result of performing hepatic vascular segmentation with the configuration of the fifth embodiment, the annotation GT (teaching data) is "37" as a statistic of the portal vein vascular branching, whereas the predicted result is "34", so the detection rate is "92%". Also, as a result of performing hepatic vascular segmentation with the configuration of the seventh embodiment, the annotation GT is "34" as a statistic of the portal vein vascular branching, whereas the predicted result is "36", so the detection rate is "97%".

[0125] As can be seen from Table 3, in the configuration of the seventh embodiment including the feature region enhancing unit 104 and the image region enhancing unit 105a, the detection rates of the prediction results for the hepatic vein and the portal vein for the annotation GT are higher than those of the configuration of the fifth embodiment, being 95% and 97%, respectively. That is, in the configuration of the seventh embodiment, the accuracy of blood vessel segmentation is higher than that of the configuration of the fifth embodiment.

[0126] In this way, in the configuration of the seventh embodiment, the feature region enhancement unit 104 and the image region enhancement unit 105a make it possible to perform vascular segmentation with high accuracy and improve vascular connectivity in the segmentation results even when the contrast between blood vessels and tissues in a medical image is not high.

[0127] In addition, according to the present embodiment, when training and segmenting based on a deep learning model, blood vessels that are annotated in GT and blood vessels that are not annotated in GT can be predicted, because it is practically difficult to achieve completely leak-free blood vessel annotation.

[0128] (Modification) Although the embodiment has been described above, the present invention may be embodied in various different forms other than the above-described embodiment.

[0129] Although the present embodiment has been described mainly with respect to hepatic vascular segmentation, it is possible to apply the present embodiment to vascular segmentation of various organs, i.e., the present embodiment can be effectively applied only when the contrast between blood vessels and tissues is not high in medical images.

[0130] The numerical ranges according to this embodiment are merely examples, and can be appropriately set or changed as necessary.

[0131] In addition, in this embodiment, a blood vessel segmentation device and a blood vessel segmentation method have been specifically described, but this embodiment is not limited to this and may be embodied in various forms including a medical image processing device, an integrated circuit, a program, and a medium on which a program is recorded, etc.

[0132] For example, a medical image processing apparatus according to the present invention includes a target image acquisition unit that acquires a target medical image, the above-mentioned blood vessel segmentation device, and an output unit that outputs a segmentation result by the blood vessel segmentation device.

[0133] For example, the medical image processing device is a computer and includes a processing circuit for executing each processing function and a storage circuit for storing each processing function in the form of a program executable by the computer. For example, the processing circuit of the medical image processing device includes a target image acquisition unit, an image acquisition unit 101, a deep learning model acquisition unit 102, a segmentation unit 103, and an output unit.

[0134] In the medical image processing device, the target image acquisition unit acquires the target medical image from the medical image diagnostic device, for example, through a communication interface. Alternatively, the target medical image may be stored in advance in a storage device, and the target image acquisition unit may read and acquire the target medical image from the storage device. As an example of the storage device, for example, the target medical image may be stored in a storage circuit of the medical image processing device, or the target medical image may be managed in a PACS or an image storage device. Furthermore, the target image acquisition unit may read the target medical image processed by a device other than the medical image diagnostic device.

[0135] Here, the target medical image is generated by a medical image diagnostic device, that is, the target medical image may be a three-dimensional image or a two-dimensional image.

[0136] In addition, in the medical image processing device, the image acquisition unit 101 acquires annotations GT (teaching data) and a plurality of medical images. The deep learning model acquisition unit 102 acquires a trained deep learning model obtained by training the deep learning model using at least a centerline enhancement loss function and a statistical shape constraint loss function based on the teaching data and the plurality of medical images acquired by the image acquisition unit 101. Then, the segmentation unit 103 obtains a segmentation result related to blood vessels by segmenting the target medical image acquired by the target image acquisition unit based on the trained deep learning model obtained by the deep learning model acquisition unit 102. The output unit outputs the segmentation result obtained by the segmentation unit 103.

[0137] Each component of the above-mentioned blood vessel segmentation device may be distributed. As an example of the distributed arrangement, for example, the trained deep learning model may be generated by a deep learning model generation device. For example, the deep learning model generation device is a computer and includes a processing circuit that executes each processing function and a storage circuit that stores the processing function in the form of a program executable by the computer. For example, the processing circuit of the deep learning model generation device includes an image acquisition unit 101 and a deep learning model acquisition unit 102.

[0138] In the deep learning model generation device, an image acquisition unit 101 acquires annotations GT (teacher data) and multiple medical images. A deep learning model acquisition unit 102 acquires a trained deep learning model obtained by training a deep learning model using at least a centerline enhancement loss function and a statistical shape constraint loss function based on the teacher data and multiple medical images acquired by the image acquisition unit 101. The acquired trained deep learning model is a model that outputs a segmentation result related to blood vessels in a target medical image by inputting the target medical image.

[0139] Furthermore, the processing circuit of the deep learning model generation device may include a preprocessing unit 106. In this case, in the deep learning model generation device, the preprocessing unit 106 performs preprocessing on the multiple medical images acquired by the image acquisition unit 101 to obtain a standardized medical image. The deep learning model acquisition unit 102 acquires a trained deep learning model based on the teacher data acquired by the image acquisition unit 101 and the standardized medical image preprocessed by the preprocessing unit 106.

[0140] Also, in the medical image processing device, the segmentation unit 103 acquires a trained deep learning model from the deep learning model generation device via a communication interface, for example, and obtains a segmentation result based on the trained deep learning model. Alternatively, the segmentation unit 103 accesses a storage device that stores the trained deep learning model acquired by the deep learning model generation device, acquires the trained deep learning model, and obtains a segmentation result based on the trained deep learning model. As an example of the storage device, for example, the trained deep learning model may be stored in a storage circuit of the deep learning model generation device, the trained deep learning model may be stored in a storage circuit of the medical image processing device, or the trained deep learning model may be stored in another storage device.

[0141] In addition, each component of each device illustrated in the present embodiment is a functional concept, and does not necessarily have to be physically configured as illustrated. In other words, the specific form of distribution and integration of each device is not limited to that illustrated, and all or a part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions. Furthermore, each processing function performed by each device can be realized in whole or in any part by a CPU and a program analyzed and executed by the CPU, or can be realized as hardware using wired logic.

[0142] The method described in this embodiment can be realized by executing a prepared program on a computer such as a personal computer or a workstation. This program can be distributed via a network such as the Internet. This program can also be recorded on a non-transitory computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, an MO, or a DVD, and executed by being read from the recording medium by a computer.

[0143] According to at least one of the embodiments described above, even when the contrast between blood vessels and tissues in a medical image is not high, blood vessel segmentation can be performed with high accuracy, and the vascular connectivity in the segmentation result can be improved.

[0144] Although some embodiments 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 forms, and various omissions, substitutions, modifications, and combinations of embodiments can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents as described in the claims, as well as in the scope and spirit of the invention. [Explanation of symbols]

[0145] 100, 100A, 100B, 100C, 100D, 100E Vascular segmentation device 101 Image acquisition unit 102 Deep learning model acquisition unit 103 Segmentation Department

Claims

1. an image acquisition unit that acquires a plurality of medical images; A deep learning model acquisition unit that acquires a trained deep learning model obtained by performing deep learning model training using a centerline enhancement loss function and a statistical shape constraint loss function based on the medical image; A segmentation unit that obtains a segmentation result related to blood vessels by segmenting the acquired target medical image based on the trained deep learning model; A blood vessel segmentation apparatus comprising:

2. a feature region highlighting unit that extracts tubular object features from the segmentation result by using tubular object filtering, obtains feature regions of the enhanced tubular object features, and connects the tubular object features by using a region growing method to obtain a segmentation result after the feature region is enhanced; The vessel segmentation apparatus of claim 1 , further comprising:

3. an image region enhancement unit for performing a connectivity restoration process on isolated blood vessel segments in the segmentation result in an image region to obtain a segmentation result after image region enhancement; The vessel segmentation apparatus of claim 1 , further comprising:

4. an image region enhancement unit for performing a connectivity restoration process on isolated blood vessel segments in the segmentation result in an image region to obtain a segmentation result after image region enhancement; The vessel segmentation apparatus of claim 2 , further comprising:

5. A morphological processing section performs morphological image processing on the segmentation result to obtain a final segmentation result. The vessel segmentation apparatus of claim 1 , further comprising:

6. a morphological processing unit that performs morphological image processing on the segmentation result after the feature region enhancement to obtain a final segmentation result; The vessel segmentation apparatus of claim 2 , further comprising:

7. a morphological processing unit that performs morphological image processing on the segmentation result after the image region enhancement to obtain a final segmentation result; The vessel segmentation apparatus according to claim 3 , further comprising:

8. a morphological processing unit that performs morphological image processing on the segmentation result after the image region enhancement to obtain a final segmentation result; The vessel segmentation apparatus according to claim 4 , further comprising:

9. A pre-processing unit that performs pre-processing on the acquired medical images to obtain a standardized medical image. Further equipped with The deep learning model acquisition unit acquires the trained deep learning model obtained by performing deep learning model training using the centerline enhancement loss function and the statistical shape constraint loss function based on the standardized medical image. The blood vessel segmentation apparatus according to claim 1 .

10. The deep learning model acquisition unit acquires the trained deep learning model obtained by training the deep learning model using an integrated loss function obtained by weighting a Dice coefficient loss function, a cross entropy loss function, the centerline enhancement loss function, and the statistical shape constraint loss function. The blood vessel segmentation apparatus according to claim 1 .

11. The feature region enhancement unit extracts the tubular object features by applying a Hessian matrix to the segmentation result. The blood vessel segmentation apparatus according to claim 2 .

12. An image acquisition step of acquiring a plurality of medical images; A deep learning model acquisition step of acquiring a trained deep learning model obtained by performing deep learning model training using a centerline enhancement loss function and a statistical shape constraint loss function based on the medical image; A segmentation step of obtaining a segmentation result related to blood vessels by segmenting the acquired target medical image based on the trained deep learning model; A blood vessel segmentation method comprising:

13. a target image acquisition unit for acquiring a target medical image; A blood vessel segmentation apparatus according to any one of claims 1 to 11, an output unit that outputs the segmentation result; A medical image processing device comprising:

14. an image acquisition unit that acquires a plurality of medical images; a deep learning model acquisition unit that acquires a trained deep learning model obtained by training a deep learning model using a centerline enhancement loss function and a statistical shape constraint loss function based on the medical image, the trained deep learning model being configured to output a segmentation result related to blood vessels in a target medical image by inputting the target medical image; A deep learning model generation device comprising:

Images