CN116310114A - Coronary artery point cloud data processing optimization method of multitask label model - Google Patents

Coronary artery point cloud data processing optimization method of multitask label model Download PDF

Info

Publication number
CN116310114A
CN116310114A CN202310261385.6A CN202310261385A CN116310114A CN 116310114 A CN116310114 A CN 116310114A CN 202310261385 A CN202310261385 A CN 202310261385A CN 116310114 A CN116310114 A CN 116310114A
Authority
CN
China
Prior art keywords
coronary
coronary artery
heart
point cloud
cloud data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310261385.6A
Other languages
Chinese (zh)
Inventor
高琪
鲁云霞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Shengshi Technology Co ltd
Original Assignee
Hangzhou Shengshi Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Shengshi Technology Co ltd filed Critical Hangzhou Shengshi Technology Co ltd
Priority to CN202310261385.6A priority Critical patent/CN116310114A/en
Publication of CN116310114A publication Critical patent/CN116310114A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00Three dimensional [3D] modelling, e.g. data description of 3D objects
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • G06T7/11Region-based segmentation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • G06T7/187Segmentation; Edge detection involving region growing; involving region merging; involving connected component labelling
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/60Analysis of geometric attributes
    • G06T7/68Analysis of geometric attributes of symmetry
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20081Training; Learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20084Artificial neural networks [ANN]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30048Heart; Cardiac
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2210/00Indexing scheme for image generation or computer graphics
    • G06T2210/41Medical
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02TCLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00Road transport of goods or passengers
    • Y02T10/10Internal combustion engine [ICE] based vehicles
    • Y02T10/40Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Geometry (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Graphics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a coronary artery point cloud data processing optimization method of a multi-task label model. Dividing the heart medical image into heart target areas by using the trained heart partition model; performing coronary search on the aortic target area by using the trained import point position model to obtain the import point position of the coronary artery; obtaining the central line of the coronary artery by combining the position of the inlet point and the heart medical image with the central line tracking model; and obtaining new coronary artery point cloud data by combining a deep learning coronary artery reconstruction algorithm with the coronary artery inlet points and feeding back the new coronary artery point cloud data to updating optimization. The invention utilizes the multi-task label to completely realize the multi-task of the full-automatic heart medical image processing, overcomes the difficult manufacturing problem of the multi-task label, and solves the problems of branch loss and adhesion in the coronary artery reconstruction process, thereby improving the accuracy and the efficiency of the coronary artery reconstruction.

Description

Coronary artery point cloud data processing optimization method of multitask label model
Technical Field
The invention relates to a heart medical image data processing optimization method, in particular to a coronary artery point cloud data processing optimization method of a multi-task label model.
Background
Along with the continuous development of medicine, medical images provide more visual basis in the treatment process of different diseases. The prior art is generally based on heart CTA data, and three-dimensional reconstruction of tissues and organs in the CTA data is of great importance, so that in the deep learning process, the manufacture of labels is of great importance.
Coronary arteries are branches of blood vessels supplying the heart, and focuses of the coronary arteries can have functional influence on corresponding heart areas, and are the reasons of coronary heart disease, angina and myocardial infarction. Among the heart CTA data, the techniques now mainly involved are coronary artery reconstruction, coronary plaque analysis, coronary stenosis analysis, heart zonal segmentation, etc., wherein reconstruction of the coronary arteries is particularly important. In the process of three-dimensional reconstruction of coronary arteries, doctors realize the reconstruction function more or less semi-automatically.
The current general technology, with respect to the study of cardiac CTA data, most of the functions are semi-automated, relying more on the relevant experience of the operator. In the development process of deep learning, the collection of a sample data set is very important, and corresponding labels need to be collected under different task states. There is currently no fixed, procedural process in the collection and management of cardiac CTA datasets.
In the prior art, after CTA images are simply preprocessed (rotated, normalized and the like), a CNN network is utilized to track the radius and the direction of a central line, and the 3D Uet network is used for directly dividing the images, so that the problems of branch fracture, loss and the like of coronary arteries are caused, the problems of branch loss or adhesion and the like are caused, and the reconstructed coronary artery point cloud cannot meet the requirements of the coronary artery tree labels.
Disclosure of Invention
In order to solve the problems in the background art, the invention provides a coronary artery point cloud data processing optimization method of a multi-task label model, solves the problems of small branch loss, auricle adhesion, pulmonary vein adhesion and the like in the coronary artery reconstruction process, and overcomes the difficult problem that the multi-task label is difficult to manufacture.
In the existing cardiac medical image data, in the state of excessive contrast agents, the problem of adhesion between the auricle and the coronary artery often occurs in the coronary artery reconstruction process. In images where contrast medium is not full, there is data of a depth stenosis in the coronary, and a problem of a branch fracture of the coronary occurs. The invention can also well solve such problems.
The technical scheme of the invention is as follows:
the detailed flow of the method of the invention can refer to the flow chart of the multi-task label training model in fig. 1, wherein the flow 13 is to directly obtain coronary artery point cloud data by giving an original heart medical image, coronary artery inlet point position and central line tracking and deep learning coronary artery reconstruction algorithm.
The CTA type heart medical images can be processed in batch in the later period, and the heart partition model is directly utilized to automatically obtain the aortic tissue, the left ventricle, the left atrium, the auricle and the cardiac muscle; directly using an inlet point position model of coronary artery to automatically search the positions of inlet points of the left crown and the right crown; obtaining coronary point cloud data by using a deep learning coronary reconstruction algorithm; and automatically extracting the coronary tree label by using a central line extraction algorithm. If the quality of various coronary artery tree labels is good, the coronary artery tree labels can be used as a new label adding data set for updating the corresponding model to obtain a new import point position model.
1. A coronary artery point cloud data processing optimization method of a multitask label model is characterized by comprising the following steps of:
step one, heart partitioning:
dividing the heart medical image into heart target areas by using the trained heart partition model;
more specifically, marking is performed on the heart medical image to obtain marked heart region data, the marked heart region data is input into a marking neural network, a heart partition model is obtained through training, and the heart medical image is segmented into heart target regions in the heart partition model.
Step two, identifying coronary artery inlet point positions:
performing coronary search on aortic tissue in a heart target area by using the trained import point position model to obtain an import point position of coronary artery;
step three, coronary artery point cloud data are obtained:
inputting a traditional coronary reconstruction algorithm to obtain initial coronary point cloud data by utilizing the position of the inlet point of the coronary obtained in the step two,
the method specifically comprises the steps of inputting the position of an inlet point of the coronary artery into a traditional coronary artery reconstruction algorithm for regional growth, dividing the coronary artery point cloud, and obtaining the coronary artery point cloud data through restoration and reservation.
Step four, center line tracking:
obtaining coronary tree tag data through coronary point cloud data, wherein the coronary tree tag data are used for calculating a target Q value, and the target Q value is used for training a central line tracking model; simultaneously, carrying out data preprocessing on the heart medical image to obtain image data serving as a first target area; inputting the position of the inlet point of the coronary artery and the first target area into the trained central line tracking model together for prediction processing to obtain the central line of the coronary artery;
step five, deep learning coronary reconstruction: and (3) obtaining new coronary artery point cloud data by adopting a special deep learning coronary artery reconstruction algorithm and combining the central line processing of the coronary artery obtained in the step (IV).
Firstly, importing a heart medical image into marking software to be marked so as to obtain heart areas of aortic tissues, left crowns, right crowns, left ventricles, left atria, auricles and cardiac muscles as first tag data; and inputting the heart medical image and the first label data into a labeled neural network for training to obtain a heart partition model, and inputting the heart medical image to be processed into the trained heart partition model for segmentation and extraction to obtain heart target areas of aortic tissues, left ventricle, left atrium, auricle and cardiac muscle.
The heart region obtained by segmenting the heart medical image data of the unknown heart region by using the trained heart partitioning model does not comprise a left crown and a right crown, and the left crown and the right crown are not segmented, because the left crown and the right crown have smaller tissues, partial adhesion and coronary fracture phenomena exist, and marked coronary has a small branch loss phenomenon if the marked coronary is obtained by direct segmentation.
The left crown and right crown results obtained by carrying out predictive segmentation on the heart medical image data of the unknown heart region by using the trained heart partition model cannot be used as labels of coronary arteries.
The second step further comprises:
and on the aortic tissue result obtained by segmentation in the step one, selecting an image area near a specific aortic tissue, removing non-coronary artery characteristics in the image area by using a method of twice expansion and connected domain marking to obtain sample characteristic points, taking part of characteristic points in the sample characteristic points as centers to intercept image blocks near the aortic tissue, carrying out normalization processing on the intercepted image blocks, taking the intercepted image blocks as input parameters, and inputting the input parameters into the trained import point position model to calculate and obtain the import point position of the coronary artery.
The output parameters of the inlet point position model are a left crown inlet point and a right crown inlet point, and the left crown inlet point and the right crown inlet point which are obtained through data processing of aortic tissue, left crown and right crown obtained through marking processing in the step one are used as labels during training.
The inlet point position model is used for automatically searching the inlet point position of the coronary artery, and the inlet point position of the coronary artery comprises a left coronary inlet point and a right coronary inlet point.
The step four, further includes:
in the prior art, in the coronary central line tracking processing, most of the central line radius and direction tracking is performed by utilizing a CNN network after CTA images are simply preprocessed (rotated, normalized and the like).
The obtaining coronary tree label data through the coronary point cloud data comprises the following steps:
extracting coronary central line data by utilizing the coronary point cloud data, and converting the extracted coronary central line data into coronary tree label data;
the preprocessing the heart medical image to obtain image data as a first target area comprises the following steps:
and (3) removing the heart regions of aortic tissues, left ventricles, left atria, auricles and cardiac muscles by using heart target region results obtained by segmentation in the step (A), determining left crown and right crown related regions in the remaining regions by using geometric features of coronary arteries which are always clung to the surface of cardiac muscles, wherein the left crown and right crown related regions are regions and comprise left crown related regions and right crown related regions, uniformly planning equiaxial intervals on image data of the left crown and right crown related regions, retaining data smaller than a gray threshold value, and normalizing to obtain a first target region.
The center line tracking model adopts a DQN reinforcement learning network and is divided into a main branch and a target branch.
Specifically, in the fourth step, the centerline tracking model adopts an DQN reinforcement learning network, and includes a trunk branch and a target branch, where the topology structures of the trunk branch and the target branch are the same;
taking the position of the inlet point of the coronary artery obtained in the second step as an initial coronary artery point, obtaining the position of the next coronary artery point according to the initial coronary artery point prediction by the central line tracking model, selecting an image with a fixed size from the first target area by taking the position of the next coronary artery point as the center, and inputting the image with the fixed size into the trained central line tracking model to obtain the central line result of the coronary artery.
The target Q value is an intermediate parameter in the DQN reinforcement learning network training of the centerline tracking model.
The implementation can also strengthen the learning central line tracking by utilizing a strengthening learning network through a plurality of sets of data, and acquire the central line tracking.
The network parameters of the target branch do not need to be updated iteratively, but are copied from the current backbone branch at intervals to form a delayed update.
The existing coronary reconstruction is usually carried out by directly carrying out image segmentation through a 3D Uet network. Due to the small size of the coronary artery, the 3D Uet network is directly used for image segmentation, so that the problems of branch fracture, loss and the like of the coronary artery are often caused. Therefore, the existing coronary artery reconstruction algorithm also has the problems of branch loss, adhesion and the like, so that the reconstructed coronary artery point cloud cannot meet the requirements of the coronary artery number label.
The method comprises the steps of firstly obtaining a coronary artery number central line by using a central line tracking model, obtaining coronary artery point cloud data by combining a region growing function, and recording the coronary artery point cloud data as a deep learning coronary artery reconstruction method. The complete coronary point cloud data obtained by the deep learning coronary reconstruction method is converted into coronary tree label data by obtaining coronary central line data through a central line extraction algorithm, and a central line tracking model is retrained, so that the accuracy of a final model is improved, the accuracy of coronary reconstruction is improved, and the problems of adhesion between the left auricle and the coronary are solved.
In the fifth step, the deep learning coronary reconstruction algorithm specifically comprises:
and (3) firstly expanding the central line of the coronary artery to obtain a connected domain range by gray level binarization, extracting according to the inlet point position on the basis of the connected domain range to obtain a non-coronary artery connected domain, removing the non-coronary artery connected domain from the connected domain range, and superposing the connected domain range finally positioned in the central line of the coronary artery on the coronary artery so as to obtain complete coronary artery point cloud data by traversing and searching.
More specifically, in the fifth step, the deep learning coronary reconstruction algorithm specifically includes:
s1, expanding a sphere with a core of 3 x 3 for K times by using the central line of the coronary artery, taking an image range in which the central line is expanded as a first range, binarizing an original medical image in the first range by using a cut-off gray level to obtain a second range, marking a connected domain by using original medical image data in the second range to obtain each connected domain to form a third range, extracting a left crown connected domain and a right crown connected domain according to a left crown inlet point and a right crown inlet point obtained in the step two in the third range to obtain a left crown connected domain and a right crown connected domain as a fourth range, and specifically taking the connected domain close to the left crown inlet point as a left crown connected domain and taking the connected domain close to the right crown inlet point as a right crown connected domain;
s2, judging the number of the connected domains in the third range:
if the number of the connected domains included in the third range exceeds two, the connected domain excluding the fourth range from the third range is taken as a fifth range; otherwise, all the communicating areas are coronary tissues, the third range is directly used as a fifth range without removal;
s3, judging whether the fifth range is on the central line:
if the fifth range is in the central line, the fifth range belongs to the coronary artery organization, the fifth range is overlapped on the coronary artery of the fourth range to obtain a sixth range, each point on the central line is traversed, the point on the central line between the fifth range and the fourth range is searched as a reference point, the reference point is overlapped on the sixth range to obtain a seventh range, and the seventh range is used as complete coronary artery point cloud data.
And then feeding back the coronary artery point cloud data to the coronary artery point cloud data in the third step for updating, and obtaining a new coronary artery central line through calculation processing of a central line extraction algorithm. Therefore, better coronary point cloud data are obtained through continuous iterative processing, the success rate and accuracy of coronary reconstruction are improved, and the problems of adhesion between the left auricle and the coronary are solved.
Preferably, the cardiac medical image may be a CTA image, the cardiac partition model/marker neural network may be a 3D Unet neural network, and the entry point location model may be a Resnet10 neural network, but is not limited thereto.
The method further comprises the step six of: and feeding back the new coronary artery point cloud data to the fourth step, obtaining coronary artery tree label data through the new coronary artery point cloud data, and retraining the central line tracking model. Preferably, the iterative processing from the fourth step to the fifth step can be repeated until the preset iterative times are reached.
2. A coronary tree label data processing optimization system for a multi-tasking label model, comprising:
the heart partitioning module is used for partitioning the heart medical image into a heart target area by using the trained heart partitioning model;
the coronary artery inlet point position module is used for inputting a trained inlet point position model into the heart target area obtained by the heart partitioning module, and carrying out coronary artery search on the aortic target area to obtain the inlet point position of the coronary artery;
the central line tracking module combines the heart target area result obtained by the segmentation of the heart partitioning module with the coronary artery point cloud data, and then inputs the coronary artery point cloud data into a trained central line tracking model for processing to obtain the central line of the coronary artery;
and the deep learning coronary reconstruction module is used for obtaining new coronary point cloud data by combining a deep learning coronary reconstruction algorithm with the position processing of the inlet point of the coronary, and feeding back the new coronary point cloud data to the central line tracking module for further updating and optimizing.
In the invention, a multi-task label model is formed by a heart partitioning module, a coronary artery inlet point position module, a central line tracking module and a deep learning coronary artery reconstruction module.
3. A storage medium storing a computer program which, when executed by a processor, implements the method described above.
Specifically, the computer program is instructions for implementing the above method correspondingly.
The multi-task labels of the present invention include heart zoning labels (aortic tissue, left ventricle, left atrium, left auricle, left crown, right crown, myocardium), left crown and right crown entry point position labels, coronary point cloud labels, coronary tree labels.
The invention can fully process the data by importing the case data to obtain a more regular and unified data set, and can meet a plurality of research tasks. For later development, support of the data set is provided, and technical direction is provided for cardiac medical diagnosis.
The beneficial effects of the invention are as follows:
the invention utilizes the multi-task label to completely realize full-automatic heart partition, automatic coronary artery inlet search, coronary artery central line tracking, three-dimensional coronary artery reconstruction and the like.
The invention provides a set of fixed flow operation, and overcomes the difficulty that the multi-task label is difficult to manufacture.
The invention provides a flow method for processing cardiac medical image data, which solves the problems of branch loss and adhesion in the coronary artery reconstruction process, thereby improving the accuracy of coronary artery reconstruction.
The invention uses full-automatic multitask label generation to optimize coronary reconstruction algorithm, not only solves the problem of adhesion between auricle tissue and coronary artery, but also solves the problem of branch fracture of blood vessel, unifies the processing mode of sample data set, and improves the efficiency of algorithm.
Drawings
FIG. 1 is a flow chart of a data optimization process under the multitasking tag training model of the present invention;
FIG. 2 (a) is a schematic view of a region growing mark;
FIG. 2 (b) is a schematic view of a Spline marker;
FIG. 2 (c) is a schematic representation of a three-dimensional cardiac zonal segmentation;
FIG. 3 is a schematic representation of an image after deboning;
FIG. 4 is a schematic illustration of an automatic search for coronary entry points;
FIG. 5 (a) is a schematic view of a coronary point cloud;
FIG. 5 (b) is a schematic diagram of a coronary point cloud superimposed on a VR map;
FIG. 5 (c) is a schematic view of a coronary centerline;
FIG. 6 is a schematic representation of coronary entry points after removal of a portion of the cardiac zoning results;
FIG. 7 is a schematic view of the region after maximum slice expansion;
FIG. 8 is a schematic diagram of maximum slice rotation and translation;
FIG. 9 is a schematic illustration of different slice preserving coronary artery regions;
FIG. 10 is a schematic diagram of a flow chart for reinforcement learning update parameters;
FIG. 11 is a schematic illustration of centerline tracking;
FIG. 12 (a) is a schematic diagram of coronary trunk and coronary branch point cloud data;
FIG. 12 (b) is a schematic diagram of coronary point cloud data at an extreme stenosis;
FIG. 13 is a schematic illustration of a coronary point cloud and a coronary centerline;
fig. 14 is a flow chart of the overall steps of the method of the present invention.
Detailed Description
The invention will be described in further detail with reference to the accompanying drawings and specific examples.
As shown in fig. 14, the embodiment of the present invention is as follows:
the original cardiac medical image is data of multiple groups of human heart CTA, the size is [ N, C, L ], N represents the number of slices of the image, C, L respectively represent the resolution size of the image, and the CTA data also comprises a series of shooting parameters including interlayer spacing and pixel spacing.
The network training related to the embodiment is to uniformly utilize multiple data sets in the center, and 740 sets of data sets are used. The present example illustrates a set of CTA data, with a size of [224,512,512], an interlayer spacing of 0.625 and a pixel spacing of 0.378906. The specific implementation process is as follows:
1. and (5) heart partitioning.
Firstly, introducing heart CTA image data into marking software, wherein the marking software marks aortic tissues, left crowns, right crowns, left ventricles, left atria and auricles by utilizing a region growing function, as shown in fig. 2 (a); the right and left crowns mark the myocardium using the Spline function, as shown in fig. 2 (b).
And secondly, using the 3D Unet neural network as a labeled neural network training model to obtain a heart partition model, and marking the heart partition model as model1. Wherein the model input parameters are [128,128,128,1], the model output parameters are [128,128,128,5], and the model1 is obtained by segmenting the aortic tissue (Aorta), the Left Ventricle (LV), the Left Atrium (LA), the auricle (LAA) and the myocardium, and the segmentation result is recorded as a heart target region, as shown in fig. 2 (c).
In the third step, according to the auxiliary information of the heart target area, redundant spines can be removed from different angles, only tissue organs near the heart are reserved, and as shown in fig. 3, the heart tissue visualization is performed.
2. Coronary artery entry point location.
First, the original heart CTA image is scaled to ensure that the pixel spacing and the layer spacing between the heart CTA images are unified (0.5,0.5,0.5).
And secondly, selecting an area of the image near the aortic tissue, and removing non-coronary features in the area by using methods such as 2 times of expansion, connected domain marking and the like to obtain a sample feature point I.
Third, the Resnet10 neural network is used as an entry point location model, which is denoted model2.
And taking a part of characteristic points in the sample characteristic points as centers, intercepting nearby image blocks, and carrying out normalization processing on the data to serve as input parameters of the Resnet10 neural network. The data of the aortic tissue, the left crown and the right crown marked by the software can directly obtain the inlet points of the left crown and the right crown, so that the data can be used as a label, wherein the label [1,0] represents the position of the inlet point of the coronary artery, and the label [0,1] represents the position of the inlet point of the coronary artery. The label is the output parameter of the Resnet10 neural network, where the input parameter of the model is [64,64,64,1], the output parameter is [ n,2], where n is the number of non-zero index coordinates.
Fourth, the Resnet10 neural network output value of the present case obtains the coordinates of the left and right entry points of the coronary artery by using a hierarchical clustering method, as shown in FIG. 4. The Resnet10 neural network is used to automatically find the coronary entry point location.
3. And (5) tracking a central line.
And 3.1, acquiring coronary tree label data through the coronary point cloud data.
First, by means of automatic coronary artery import point position and combining with region growth in the traditional coronary artery reconstruction algorithm, a coronary artery point cloud is segmented, and coronary artery point cloud data are obtained through manual restoration and reservation processing and recorded as a second target region, as shown in fig. 5 (a). Here, the second target area is superimposed on the VR display, as shown in fig. 5 (b), and a partial branch loss can be found.
The conventional coronary artery reconstruction algorithm specifically refers to a coronary artery reconstruction algorithm obtained by using region growth through the position of a coronary artery inlet point, and specific details can be referred to an image processing method, an image processing device and a storage medium of CN 201910047390.0.
Secondly, coronary artery central line extraction can be performed by utilizing coronary artery point cloud data and utilizing a central line extraction algorithm, and particularly, the method proposed in a patent of a radius calculation method, a terminal and a storage medium of a coronary artery of CN202110278953.4 can be referred to, as shown in fig. 5 (c), wherein the storage mode of the coronary artery central line data is [ NN, W ], NN is the number of central points on the coronary artery central line, W is the information of the central points on the coronary artery central line, and is recorded as a first target point. Including but not limited to the following: coordinates under a pixel coordinate system, coordinates under a physical coordinate system, tangential direction of a center line, curvature direction, curvature radius, traversal index of the center line, and the like.
Thirdly, the coronary central line data are converted into coronary tree label data along the directions of X, Y and Z at equal intervals (0.5,0.5,0.5), and the coronary tree label data are recorded as a second target point.
The coronary tree is stored in a dictionary which is formed by separately separating a section of coronary artery, wherein a keyword is stored as an index of a starting position and an ending position, the example can be '0, 20)', a value is stored as a central line position coordinate of the section of coronary artery, and a central point of the coronary artery at a bifurcation position is not only the ending position of the previous section of coronary artery, but also the starting point of the next section of coronary artery, so that the coronary tree label data can be obtained.
3.2, preprocessing the original CTA image.
Subsequently, an DQN reinforcement learning network is adopted as a central line tracking model, and when the DQN reinforcement learning is carried out, the original image CTA is subjected to data preprocessing as follows, and the obtained image data is recorded as a first target area:
the left ventricle, left atrium, auricle in the segmentation result of the heart partition model are removed first, and then the aortic tissue retaining the position of the coronary access point is removed, as shown in fig. 6.
And then, there are problems of auricle, pulmonary vein adhesion, etc. due to coronary artery. Typically such problems are left coronary and rarely occur in right coronary. In order to solve such a problem, the coronary left crown related region is defined in advance according to the feature that the coronary left crown always surrounds the surface of the myocardial muscle, while it is possible to enhance the calculation time. The maximum two-dimensional slice is determined by maximizing the common area of the left ventricle and the myocardium, on which a fixed region is determined by a certain threshold T-dilation pixel process (in this case t=50), as shown in fig. 7, and the fixed region is updated simultaneously into three-dimensional space.
Next, the relevant region on the obtained maximum two-dimensional slice is denoted as Plane1, and the centroid point P is calculated 0 ,P 1 Is the right coronary artery inlet point. Will be rotated 90 degrees to obtain P 2 The area that is the centroid is denoted Plane2. Subsequently Plane2 is translated as P 1 The relevant area, denoted Plane3, is the centroid, as shown in fig. 8. Because the coronary artery right crown generally does not have excessive adhesion, the relevant region of Plane3 can be further expanded outward by a certain threshold pixel to ensure that the right crown falls entirely within the relevant region.
Finally, in combination with the above operations, the present invention can frame the coronary arteries in the heart in the first target region as shown in fig. 9, and normalize the region data while maintaining the gray scale value between [ -100,700], and uniformly program the layer spacing and pixel spacing to [0.5,0.5,0.5].
In the image preprocessing process, the spatial characteristics of the coronary artery are combined to make corresponding limitation, and meanwhile, the calculated amount is reduced and the central line of the coronary artery is tracked in a targeted manner.
And 3.3, constructing a model.
And adopting the DQN reinforcement learning network as a topological structure of the central line tracking model.
Taking the coronary artery inlet point position as an initial position, acquiring the next point position according to certain area fluctuation (the box with the area size of 4 x 4 in the example) from the initial position, updating the image data with the size of [19,19,19] selected from the first Target area by taking the initial position as the center, and taking the image data as input parameters of a main branch Policy net and a Target branch Target net of the DQN reinforcement learning network, as shown in fig. 10.
In order to update the block diagram for the parameters of reinforcement learning, the current state is S (t), the Q values of different Actions corresponding to the current state can be predicted through a trunk branch Policy net, and Actions with the largest Q value are selected through a Policy function to be converted. The different behaviors refer to different directions in which the next point falls.
The state of the next moment is set as S (t+1), the Q value Q (t+1) corresponding to the next moment is calculated through the Target branch Target net, loss is calculated, and the trunk branch Policy net is updated.
Loss=((r t +γ*max a′ Q(S t+1 ,a′)-Q(S t ,a)) 2
Wherein r is t For the prize value in the policy function, gamma is the discount factor, a is the current behavior, S t For the current state, a' the next action, S t+1 The next state.
The Q value is calculated by using the coronary tree label data obtained in 3.1 and is used as an output parameter of the DQN reinforcement learning network, a sample data set is enlarged by randomly rotating in X, Y and Z axis directions in the data training process of the DQN reinforcement learning network, a Target branch Target net is updated by using a main branch Policy net parameter after a period of time, and after training is carried out for a plurality of times, a central line tracking model3 is obtained, wherein the model3 is used for tracking central line data.
4. Deep learning coronary reconstruction algorithms.
The heart CTA data to be processed in this example is subjected to the data preprocessing process, and the center line is directly obtained through the center line tracking model3, as shown in fig. 11, so that the branches on the first-stage branch of the anterior descending branch and the branches on the right crown trunk can be found.
The region growing is performed as coronary reconstruction at a certain cut-off gray value (here the cut-off gray value is 180) using the tracked center line. There is often an extreme stenosis of the coronary, where the gray value is lower than the cut-off gray value, but after the coronary stenosis the coronary branches are normal. For such problems, a centerline is generally obtained by centerline tracing, but branches after coronary stenosis cannot be grown when coronary artery is grown in the region growing process.
To solve this problem, the present invention also expands the tracking centerline K times in a deep learning coronary reconstruction algorithm first, the nuclei are spheres of 3 x 3 (in this example k=10), and the coronary arteries are substantially within the range defined by the post-expansion centerline, this region being denoted as the first range. The original image in the first range is then binarized with a cut-off gray scale, the white being noted as the second range. And marking different connected domains by using a measurement function in a skin library in the Python language in the second range, and marking the data as a third range.
At this time, the third range necessarily has not less than 2 connected domains, one is a left crown, the other is a right crown, and the fourth range is recorded. In the case where the connected domain of the third range exceeds two connected domains, there is a possibility that the coronary branch breaks due to coronary stenosis, or that the connected domain is not a tissue on the coronary, the left and right coronary connected domains are obtained as the fourth range by extraction from the left and right coronary entry points, and the connected domain remaining after the fourth range is removed from the third range is the fifth range.
It is determined whether the fifth range is on the center line after tracking, and usually the fifth range is within the center line, and the fifth range is a coronary tissue, which can be superimposed on the fourth range coronary, and is denoted as the sixth range, as shown in fig. 12 (a).
Each point on the center line after the tracking is traversed, and a point on the center line between the fifth range and the fourth range is found and superimposed on the sixth range, denoted as a seventh range, as shown in fig. 12 (b).
The seventh range obtained at this time is a complete coronary point cloud data, and then returns to step 3 to calculate the coronary geometry information through the centerline extraction algorithm, as shown in fig. 13.
Therefore, the success rate and accuracy of coronary reconstruction can be improved. The method also solves the problems of adhesion between the left auricle and the coronary artery.
If the later stage is implemented for other heart CTA image data sets, data processing can be carried out according to the heart partition model, the coronary artery inlet point position model, the central line tracking model, the traditional coronary artery reconstruction algorithm and the deep learning coronary artery reconstruction algorithm, the data obtained in the later stage can meet the requirements of corresponding labels, and can be used as a training set of a network, so that the collection of the data sets can be conveniently enlarged.

Claims (10)

1. A coronary artery point cloud data processing optimization method of a multitask label model is characterized by comprising the following steps of:
step one, heart partitioning: dividing the heart medical image into heart target areas by using the trained heart partition model;
step two, identifying coronary artery inlet point positions: performing coronary search on aortic tissue in a heart target area by using the trained import point position model to obtain an import point position of coronary artery;
step three, coronary artery point cloud data are obtained: inputting a coronary reconstruction algorithm to obtain coronary point cloud data by using the position of the inlet point of the coronary obtained in the step two,
step four, center line tracking: obtaining coronary tree tag data through coronary point cloud data, wherein the coronary tree tag data are used for calculating a target Q value, and the target Q value is used for training a central line tracking model; simultaneously, carrying out data preprocessing on the heart medical image to obtain image data serving as a first target area; inputting the position of the inlet point of the coronary artery and the first target area into the trained central line tracking model together for prediction processing to obtain the central line of the coronary artery;
step five, deep learning coronary reconstruction: and (3) obtaining new coronary artery point cloud data by adopting a deep learning coronary artery reconstruction algorithm and combining the central line processing of the coronary artery obtained in the step (IV).
2. The coronary point cloud data processing optimization method of a multitasking label model of claim 1, wherein: firstly, importing a heart medical image into marking software to be marked so as to obtain heart areas of aortic tissues, left crowns, right crowns, left ventricles, left atria, auricles and cardiac muscles as first tag data; and inputting the heart medical image and the first label data into a labeled neural network for training to obtain a heart partition model, and inputting the heart medical image to be processed into the trained heart partition model for segmentation to obtain heart target areas of aortic tissues, left ventricle, left atrium, auricle and cardiac muscle.
3. The coronary point cloud data processing optimization method of the multitasking label model of claim 2, wherein: the second step comprises the following steps: and (3) selecting an image area of the aortic tissue on the aortic tissue result obtained by segmentation in the step one, removing non-coronary artery characteristics in the image area by using a method of twice expansion and connected domain marking to obtain sample characteristic points, taking part of characteristic points in the sample characteristic points as centers to intercept image blocks near the aortic tissue, carrying out normalization processing on the intercepted image blocks, taking the intercepted image blocks as input parameters, and inputting the input parameters into the trained import point position model to calculate and obtain the import point position of the coronary artery.
4. The coronary point cloud data processing optimization method of a multitasking label model of claim 1, wherein: the obtaining coronary tree label data through the coronary point cloud data comprises the following steps:
extracting coronary central line data by utilizing the coronary point cloud data, and converting the extracted coronary central line data into coronary tree label data;
the preprocessing the heart medical image to obtain image data as a first target area comprises the following steps:
and (3) removing heart regions of aortic tissues, left ventricles, left atria, auricles and cardiac muscles by using heart target region results obtained by segmentation in the step (A), determining left crown and right crown related regions in the remaining regions by using geometric features of coronary arteries which are always closely attached to the surface of cardiac muscles, uniformly planning equiaxial intervals of image data of the left crown and right crown related regions, retaining data smaller than a gray threshold value, and normalizing to obtain a first target region.
5. The coronary point cloud data processing optimization method of the multitasking label model of claim 1 or 4, characterized by: in the fourth step, the center line tracking model adopts a DQN reinforcement learning network, and comprises a main branch and a target branch, wherein the topology structures of the main branch and the target branch are the same;
taking the position of the inlet point of the coronary artery obtained in the second step as an initial coronary artery point, obtaining the position of the next coronary artery point according to the initial coronary artery point prediction by the central line tracking model, selecting an image with a fixed size from the first target area by taking the position of the next coronary artery point as the center, and inputting the image with the fixed size into the trained central line tracking model to obtain the central line result of the coronary artery.
6. The coronary point cloud data processing optimization method of a multitasking label model of claim 1, wherein: in the fifth step, the deep learning coronary reconstruction algorithm specifically comprises:
and (3) firstly expanding the central line of the coronary artery to obtain a connected domain range by gray level binarization, extracting according to the inlet point position on the basis of the connected domain range to obtain a non-coronary artery connected domain, removing the non-coronary artery connected domain from the connected domain range, and superposing the connected domain range finally positioned in the central line of the coronary artery on the coronary artery so as to obtain complete coronary artery point cloud data by traversing and searching.
7. The coronary point cloud data processing optimization method of a multitasking label model of claim 1, wherein: the method further comprises the steps of:
and step six, feeding back new coronary artery point cloud data to the step four, obtaining coronary artery tree label data through the new coronary artery point cloud data, and retraining a central line tracking model.
8. A coronary tree label data processing optimization system for a multi-task label model applied to the method of any one of claims 1-7, comprising:
the heart partitioning module is used for partitioning the heart medical image into a heart target area by using the trained heart partitioning model;
the coronary artery inlet point position module is used for inputting a trained inlet point position model into the heart target area obtained by the heart partitioning module, and carrying out coronary artery search on the aortic target area to obtain the inlet point position of the coronary artery;
the central line tracking module is used for combining the coronary artery point cloud data by utilizing the heart target area result obtained by the heart partition module, and further inputting the coronary artery point cloud data into the central line tracking model for processing to obtain the central line of the coronary artery;
and the deep learning coronary reconstruction module is used for obtaining new coronary point cloud data by combining a deep learning coronary reconstruction algorithm with the position processing of the inlet point of the coronary, and feeding back the new coronary point cloud data to the central line tracking module for further updating and optimizing.
9. A storage medium storing a computer program, which when executed by a processor implements the method of any one of claims 1 to 7.
10. A storage medium according to claim 9, wherein said computer program is instructions for implementing the method according to any one of claims 1 to 7.
CN202310261385.6A 2023-03-17 2023-03-17 Coronary artery point cloud data processing optimization method of multitask label model Pending CN116310114A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310261385.6A CN116310114A (en) 2023-03-17 2023-03-17 Coronary artery point cloud data processing optimization method of multitask label model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310261385.6A CN116310114A (en) 2023-03-17 2023-03-17 Coronary artery point cloud data processing optimization method of multitask label model

Publications (1)

Publication Number Publication Date
CN116310114A true CN116310114A (en) 2023-06-23

Family

ID=86835649

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310261385.6A Pending CN116310114A (en) 2023-03-17 2023-03-17 Coronary artery point cloud data processing optimization method of multitask label model

Country Status (1)

Country Link
CN (1) CN116310114A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117079080A (en) * 2023-10-11 2023-11-17 青岛美迪康数字工程有限公司 Training optimization method, device and equipment for coronary artery CTA intelligent segmentation model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117079080A (en) * 2023-10-11 2023-11-17 青岛美迪康数字工程有限公司 Training optimization method, device and equipment for coronary artery CTA intelligent segmentation model
CN117079080B (en) * 2023-10-11 2024-01-30 青岛美迪康数字工程有限公司 Training optimization method, device and equipment for coronary artery CTA intelligent segmentation model

Similar Documents

Publication Publication Date Title
CN107563983B (en) Image processing method and medical imaging device
CN110458847B (en) Automatic coronary artery segmentation and center line extraction method based on CTA image
Tobon-Gomez et al. Benchmark for algorithms segmenting the left atrium from 3D CT and MRI datasets
CN101425186B (en) Liver subsection method based on CT image and system thereof
CN101393644B (en) Hepatic portal vein tree modeling method and system thereof
CN108765430B (en) Cardiac left cavity region segmentation method based on cardiac CT image and machine learning
CN109215032A (en) The method and system of image segmentation
US20220036561A1 (en) Method for image segmentation, method for training image segmentation model
CN110458848B (en) Method for judging coronary artery branch in image coronary artery segmentation
CN115908297A (en) Topology knowledge-based blood vessel segmentation modeling method in medical image
CN116310114A (en) Coronary artery point cloud data processing optimization method of multitask label model
CN112561917A (en) Image segmentation method and system, electronic device and readable storage medium
CN113160120A (en) Liver blood vessel segmentation method and system based on multi-mode fusion and deep learning
CN116051738A (en) Method for reconstructing coronary artery blood vessel model based on CTA image and readable storage medium
Badano et al. Artificial intelligence and cardiovascular imaging: A win-win combination.
Yong et al. Automatic ventricular nuclear magnetic resonance image processing with deep learning
CN113889238A (en) Image identification method and device, electronic equipment and storage medium
CN113192069A (en) Semantic segmentation method and device for tree structure in three-dimensional tomography image
CN104915989A (en) CT image-based blood vessel three-dimensional segmentation method
Yuan et al. Pulmonary arteries segmentation from CT images using PA‐Net with attention module and contour loss
CN107464243B (en) Aortic valve positioning method, device and equipment
CN110378868B (en) Pericardium segmentation method based on CT sequence and method for roughly positioning pericardium area from pericardium sequence center layer slice
Pan et al. Automatic annotation of liver computed tomography images based on a vessel‐skeletonization method
Deng et al. Automatic segmentation of the left atrium from mr images via semantic information
Ansó Synthetic vascular structure generation for unsupervised pre-training in CTA segmentation tasks

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination