CN109801294A - Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium - Google Patents
Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium Download PDFInfo
- Publication number
- CN109801294A CN109801294A CN201811535118.9A CN201811535118A CN109801294A CN 109801294 A CN109801294 A CN 109801294A CN 201811535118 A CN201811535118 A CN 201811535118A CN 109801294 A CN109801294 A CN 109801294A
- Authority
- CN
- China
- Prior art keywords
- roi region
- dimensional
- convolutional layer
- split
- region
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/24—Aligning, centring, orientation detection or correction of the image
Abstract
The embodiment of the present application is suitable for technical field of medical image processing, discloses a kind of three-dimensional atrium sinistrum dividing method, device, terminal device and computer readable storage medium, and method includes: to obtain cardiac magnetic resonance images to be split；ROI region is partitioned into from cardiac magnetic resonance images to be split, ROI region is the region comprising three-dimensional atrium sinistrum；By ROI region input level converging network model trained in advance, the segmentation result of cardiac magnetic resonance images to be split is obtained；Wherein, level converging network model is the U-Net convolutional neural networks model for including encoder path and decoder-path, encoder path includes at least one layering aggregation module, and layering aggregation module includes the level polymerized unit as trunk branch and the attention unit as mask branch.The efficiency and accuracy rate of three-dimensional atrium sinistrum segmentation can be improved in the embodiment of the present application.
Description
Technical field
The application belongs to technical field of medical image processing more particularly to a kind of three-dimensional atrium sinistrum dividing method, device, end
End equipment and computer readable storage medium.
Background technique
Medical image, which refers to, utilizes computer tomography CT, magnetic resonance imaging MRI, B ultrasound or positron emission meter
The image data that calculation machine tomography PET etc. medical imaging device obtains is usually three formed in two dimension slicing form
Dimensional data image.The segmentation of medical image is the key means for handling medical image, and referring to will be in medical image with special
The different zones of connotation distinguish, these regions be mutually it is Uncrossed, each region meets the consistency of specific region.
Auricular fibrillation also known as atrial fibrillation are common types of arrhythmia, due to being short in understanding to human atrial structure, lead to mesh
Preceding atrial fibrillation therapeutic effect is bad.Gadolinium contrast agent is applied to MRI scan, to improve the clarity of patient's image of internal structure, gadolinium
Enhancing magnetic resonance imaging GE-MRI is the important tool for evaluating Atrial fibrosis.
And in order to become more apparent upon atrium structure, it generally requires to carry out atrium segmentation to MRI image.Currently, from three-dimensional GE-
Dividing atrium sinistrum LA in MRI image can be challenging for.The poor contrast of atrium sinistrum LA and background reduce the atrium sinistrum side LA
The visibility on boundary；During the scanning process, the patient respiratory rhythm and pace of moving things is irregular and heart rate variability, picture quality may be affected.
In recent years, with the rapid development of deep learning, there has been proposed several full automatic atrium sinistrum LA dividing methods.For example, can
With in the multiple view convolutional neural networks with adaptive convergence strategy, respectively by three-dimensional data from axial direction, sagittal plane and hat
Shape face resolves to two dimensional component respectively, then carries out multiple views convolutional neural networks analysis respectively to each component；It can also adopt
With the ConvLSTM extension residual error network constituted and sequential learning network, multiple views learning strategy is extended.But it existing is based on
The room method degraded performance of the three-dimensional left heart of GE-MRI image segmentation.
Summary of the invention
In view of this, the embodiment of the present application provides a kind of three-dimensional atrium sinistrum dividing method, device, terminal device and computer
Readable storage medium storing program for executing, to solve the problems, such as the degraded performance of existing cardiac magnetic resonance images.
The first aspect of the embodiment of the present application provides a kind of three-dimensional atrium sinistrum dividing method, comprising:
Obtain cardiac magnetic resonance images to be split；
ROI region is partitioned into from the cardiac magnetic resonance images to be split, the ROI region is to include three-dimensional atrium sinistrum
Region；
By ROI region input level converging network model trained in advance, the cardiac magnetic resonance to be split is obtained
The segmentation result of image；
Wherein, the level converging network model is the U-Net convolutional Neural for including encoder path and decoder-path
Network model, the encoder path include at least one layering aggregation module, and the layering aggregation module includes being used as trunk
The level polymerized unit of branch and attention unit as mask branch.
With reference to first aspect, in one possible implementation, described from the cardiac magnetic resonance images to be split
It is partitioned into ROI region, comprising:
The cardiac magnetic resonance images to be split are detected by pre-training U-Net convolutional neural networks, obtain ROI
Area detection result；Wherein, every level-one of the pre-training U-Net convolutional neural networks has two convolutional layers；
The cardiac magnetic resonance images to be split are cut according to the ROI region testing result, obtain the ROI
Region.
With reference to first aspect, in one possible implementation, the level polymerized unit includes the first convolutional layer, the
Two convolutional layers, third convolutional layer, Volume Four lamination and the 5th convolutional layer are respectively connected with batch normalization and school after each convolutional layer
Linear positive unit；
Wherein, first convolutional layer and second convolutional layer cascade, and cascade result is inputted the third convolutional layer；
Convolution algorithm is carried out to the third convolutional layer and obtains the Volume Four lamination, the third convolutional layer and the Volume Four lamination
Connection generates the 5th convolutional layer；
The attention unit includes the 6th convolutional layer, the 7th convolutional layer and second shape structure sheaf, and the 6th convolutional layer is successively
It is connect by batch normalization and the correction linear unit with the 7th convolutional layer；The second shape structure sheaf isExp indicates exponential function, χi, c expression cthThe i of Feature Mapping on channelthValue；
The output of the hierarchical fusion module isMi,c(χ) expression is covered
The output of code branch, range are [0,1]；Fi,c(χ) indicates the output of trunk branch,Indicate dot-product,It indicates to press element side
To summation.
With reference to first aspect, in one possible implementation, it is described obtain cardiac magnetic resonance images to be split it
Before, further includes:
Obtain training sample and label information corresponding with the training sample；
According to the label information, corresponding target ROI region is partitioned into from the training sample；
According to the target ROI region, model training is carried out to the level converging network model pre-established.
With reference to first aspect, in one possible implementation, described according to the label information, from the trained sample
Corresponding target ROI region is partitioned into this, comprising:
The training sample is uniformly adjusted to the first preset shape；
By the training sample input of first preset shape U-Net convolutional neural networks trained in advance, target is obtained
ROI region testing result；
Each training sample is adjusted from first preset shape to the original shape of each training sample；
According to the ROI region testing result and the label information, corresponding institute is partitioned into from the training sample
State target ROI region.
With reference to first aspect, in one possible implementation, it is described according to the target ROI region testing result and
The label information is partitioned into the corresponding target ROI region from the training sample, comprising:
Judge whether the target ROI region testing result includes the label information；
When the target ROI region testing result does not include the label information, then the target ROI region is extended
For goal-selling region；
The goal-selling region is cut out from the training sample, using the goal-selling region as the target
ROI region；
When the target ROI region testing result includes the label information, institute is cut out from the training sample
State target ROI region.
The second aspect of the embodiment of the present application provides a kind of three-dimensional atrium sinistrum segmenting device, comprising:
Module is obtained, for obtaining cardiac magnetic resonance images to be split；
ROI region divides module, for being partitioned into ROI region, the ROI from the cardiac magnetic resonance images to be split
Region is the region comprising three-dimensional atrium sinistrum；
Divide module, for the level converging network model that ROI region input is trained in advance, obtains described wait divide
Cut the segmentation result of cardiac magnetic resonance images；
Wherein, the level converging network model is the U-Net convolutional Neural for including encoder path and decoder-path
Network model, the encoder path include at least one layering aggregation module, and the layering aggregation module includes being used as trunk
The level polymerized unit of branch and attention unit as mask branch.
In conjunction with second aspect, in one possible implementation, the ROI region segmentation module includes:
ROI region detection unit, for passing through pre-training U-Net convolutional neural networks to the cardiac magnetic resonance to be split
Image is detected, and ROI region testing result is obtained；Wherein, every level-one of the pre-training U-Net convolutional neural networks has
Two convolutional layers；
ROI region cuts unit, is used for according to the ROI region testing result to the cardiac magnetic resonance images to be split
It is cut, obtains the ROI region.
In conjunction with second aspect, in one possible implementation, the level polymerized unit includes the first convolutional layer, the
Two convolutional layers, third convolutional layer, Volume Four lamination and the 5th convolutional layer are respectively connected with batch normalization and school after each convolutional layer
Linear positive unit；
Wherein, first convolutional layer and second convolutional layer cascade, and cascade result is inputted the third convolutional layer；
Convolution algorithm is carried out to the third convolutional layer and obtains the Volume Four lamination, the third convolutional layer and the Volume Four lamination
Connection generates the 5th convolutional layer；
The attention unit includes the 6th convolutional layer, the 7th convolutional layer and second shape structure sheaf, and the 6th convolutional layer is successively
It is connect by batch normalization and the correction linear unit with the 7th convolutional layer；The second shape structure sheaf isExp indicates exponential function, χi, c expression cthThe i of Feature Mapping on channelthValue；
The output of the hierarchical fusion module isMi,c(χ) expression is covered
The output of code branch, range are [0,1]；Fi,c(χ) indicates the output of trunk branch,Indicate dot-product,It indicates to press element side
To summation.
In conjunction with second aspect, in one possible implementation, further includes:
Training sample obtains module, for obtaining training sample and label information corresponding with the training sample；
Target ROI region divides module, corresponding for being partitioned into from the training sample according to the label information
Target ROI region；
Training module, for being carried out to the level converging network model pre-established according to the target ROI region
Model training.
In conjunction with second aspect, in one possible implementation, the target ROI region segmentation module includes:
The first adjustment unit, for uniformly adjusting the training sample to the first preset shape；
Detection unit, the U-Net convolutional Neural trained in advance for the training sample input by first preset shape
Network obtains target ROI region testing result；
Second adjustment unit, for adjusting each training sample from first preset shape to each training sample
Original shape；
Cutting unit, for dividing from the training sample according to the ROI region testing result and the label information
Cut out the corresponding target ROI region.
In conjunction with second aspect, in one possible implementation, the cutting unit includes:
Judgment sub-unit, for judging whether the target ROI region testing result includes the label information；
Subelement is extended, it, then will be described for when the target ROI region testing result does not include the label information
Target ROI region is extended to goal-selling region；
First cuts subelement, will be described default for cutting out the goal-selling region from the training sample
Target area is as the target ROI region；
Second cuts subelement, for when the target ROI region testing result includes the label information, from described
The target ROI region is cut out in training sample.
The third aspect of the embodiment of the present application provides a kind of terminal device, including memory, processor and is stored in institute
The computer program that can be run in memory and on the processor is stated, the processor executes real when the computer program
Now as described in above-mentioned any one of first aspect the step of three-dimensional atrium sinistrum dividing method.
The fourth aspect of the embodiment of the present application provides a kind of computer readable storage medium, the computer-readable storage medium
Matter is stored with computer program, realizes three as described in above-mentioned any one of first aspect when the computer program is executed by processor
The step of tieing up atrium sinistrum dividing method.
Existing beneficial effect is the embodiment of the present application compared with prior art:
The embodiment of the present application from cardiac magnetic resonance images to be split by being partitioned into the area ROI comprising three-dimensional atrium sinistrum
Domain can greatly reduce calculation amount using the ROI region as the input of network model, while can also substantially reduce image back
The interference of scape, to improve the efficiency and accuracy rate of three-dimensional atrium sinistrum segmentation；Using level converging network model to ROI region
It is split, which is iterated by pantostrat of the layering aggregation module to same stage different depth
Fusion, improves the shallow-layer of network model and the Fusion Features ability of deep layer, can obtain more preferable fuse information, it will also be noted that unit
It is attached to each stage of encoder section as mask branch, has so as to gradually be enhanced using the spatial information of shallow-layer
The deep layer contour feature of abundant semantic information substantially increases a three-dimensional left side by combining level aggregation and attention mechanism
The efficiency and accuracy rate of atrium segmentation.
Detailed description of the invention
It in order to more clearly explain the technical solutions in the embodiments of the present application, below will be to embodiment or description of the prior art
Needed in attached drawing be briefly described, it should be apparent that, the accompanying drawings in the following description is only some of the application
Embodiment for those of ordinary skill in the art without any creative labor, can also be according to these
Attached drawing obtains other attached drawings.
Fig. 1 is a kind of schematic process flow diagram of three-dimensional atrium sinistrum dividing method provided by the embodiments of the present application；
Fig. 2 is the level converging network structural schematic diagram provided by the embodiments of the present application based on attention；
Fig. 3 is a kind of another schematic process flow diagram of three-dimensional atrium sinistrum dividing method provided by the embodiments of the present application；
Fig. 4 is the detailed process schematic block diagram of step S302 provided by the embodiments of the present application；
Fig. 5 is the comparative result schematic diagram of the dice value of UNet-2 provided by the embodiments of the present application and HAANet-3；
Fig. 6 is the comparison schematic diagram of two-dimentional segmentation result and three-dimensional segmentation result provided by the embodiments of the present application；
Fig. 7 is a kind of structural schematic block diagram of three-dimensional atrium sinistrum segmenting device provided by the embodiments of the present application；
Fig. 8 is the schematic diagram of terminal device provided by the embodiments of the present application.
Specific embodiment
In being described below, for illustration and not for limitation, the tool of such as particular system structure, technology etc is proposed
Body details, so as to provide a thorough understanding of the present application embodiment.However, it will be clear to one skilled in the art that there is no these specific
The application also may be implemented in the other embodiments of details.In other situations, it omits to well-known system, device, electricity
The detailed description of road and method, so as not to obscure the description of the present application with unnecessary details.
In order to illustrate technical solution described herein, the following is a description of specific embodiments.
Embodiment one
It referring to Figure 1, is a kind of schematic process flow diagram of three-dimensional atrium sinistrum dividing method provided by the embodiments of the present application, it should
Method may comprise steps of:
Step S101, cardiac magnetic resonance images to be split are obtained.
Step S102, ROI region is partitioned into from cardiac magnetic resonance images to be split, ROI region is comprising the three-dimensional left heart
The region in room.
It should be noted that above-mentioned ROI region (region of interest, region of interest) includes entire three-D volumes
Left atrial region, in the ROI region, percentage shared by atrium sinistrum is bigger, percentage pole shared by image background regions
It is small.Under normal circumstances, in each cardiac magnetic resonance images to be split, percentage very little shared by atrium sinistrum, background area institute
The percentage accounted for is bigger.That is, most of volume data in cardiac magnetic resonance images to be split divides atrium sinistrum
Task is useless, if using whole picture cardiac magnetic resonance images to be split as if the input of network model, allowing many nothings
It participates in calculating with data, calculation amount is larger to influence computational efficiency, while being also the huge waste to computing resource.In addition,
Since the comparison in the picture between the region and background of atrium sinistrum is all very low, if with whole picture cardiac magnetic resonance images to be split
If input as network model, a wide range of background area, surrounding tissue etc. can cause greatly to interfere to atrium sinistrum segmentation,
To reduce segmentation accuracy rate.
And after being partitioned into ROI region in image, then using ROI region as the basis of subsequent calculating and segmentation, such energy
Greatly reduce calculation amount, help to mitigate the influence to segmentation of surrounding tissue, background, thus greatly improve segmentation efficiency and
Accuracy rate.
In some embodiments, it can use U-Net convolutional neural networks to examine cardiac magnetic resonance images to be split
It surveys, to obtain the estimation results of atrium sinistrum, then cuts out corresponding region again as ROI region.Therefore it is above-mentioned from the heart to be split
The detailed process that ROI region is partitioned into dirty magnetic resonance image may include: to be treated by pre-training U-Net convolutional neural networks
Segmentation cardiac magnetic resonance images are detected, and ROI region testing result is obtained；Wherein, pre-training U-Net convolutional neural networks
Every level-one has two convolutional layers；Cardiac magnetic resonance images to be split are cut according to ROI region testing result, are obtained
ROI region.
It is understood that U-Net convolutional neural networks refer to the overall network structure convolutional Neural similar with alphabetical " U "
Network is considered as the deformation of convolutional neural networks, specifically includes constricted path and extensions path.U-Net network herein
Every level-one all have two convolutional layers.Specific net structure about U-Net is by as it is known to those skilled in the art that herein
It repeats no more.
It advances with training sample to be trained the U-Net convolutional neural networks, after the completion of training, by this U-Net volumes
Product neural network detects network as ROI, to be partitioned into corresponding ROI region.
In concrete application, cardiac magnetic resonance images to be split can be adjusted to a fixed shape, then be inputted again
To trained U-Net convolutional neural networks, the output of neural network is obtained as a result, the output is the result is that a rough prediction
As a result, left atrial region can be oriented in the regional location of whole picture cardiac magnetic resonance images to be split according to the prediction result.
After orienting atrium sinistrum region, cardiac magnetic resonance images to be split can be adjusted into back its original-shape, then again will
Corresponding region is cut, and the region cut is ROI region.
As can be seen that network is detected using U-Net convolutional neural networks as ROI for the first time herein, by U-Net in medicine
Feature on image procossing can be further improved the efficiency and accuracy rate of atrium sinistrum segmentation.
It is of course also possible to no longer be limited herein using other ROI region dividing methods.
Step S103, ROI region input is trained in advance level converging network model, obtains cardiac magnetic resonance to be split
The segmentation result of image.Wherein, level converging network model is the U-Net convolution mind for including encoder path and decoder-path
Through network model, encoder path includes at least one layering aggregation module, and layering aggregation module includes as trunk branch
Level polymerized unit and attention unit as mask branch.
It should be noted that above-mentioned level converging network model refers to the three-dimensional volume in conjunction with hierarchical fusion and attention mechanism
The network naming can be level converging network model (the Attention based based on attention by product neural network
Hierarchical aggregation network, HAANet).
The level converging network model is the model based on U-Net convolutional neural networks comprising the reconciliation of encoder path
Code device path.Encoder path includes at least one layering aggregation module (Attention based Hierarchical
Aggregaition Module, HAAM), layering aggregation module includes the level polymerized unit as trunk branch
(Hierarchical Aggregaition Unit, HAU) and as mask branch attention unit (Attention Unit,
AU).Its decoder-path is identical as U-Net, is specifically made of multiple duplicate convolutional layers, followed by batch normalization (Batch
Normalization, BN) and the linear unit R eLU of correction.
Above-mentioned level polymerized unit HAU may include the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four product
Layer and the 5th convolutional layer are respectively connected with batch normalization BN after each convolutional layer and correct linear unit R eLU；Wherein, the first volume
Lamination and the second convolutional layer cascade, and cascade result is inputted third convolutional layer, surrenders third convolutional layer；To third convolutional layer into
Row convolution algorithm obtains Volume Four lamination, and third convolutional layer is connected with Volume Four lamination generates the 5th convolutional layer.
In the level polymerized unit, the core size of all convolution algorithms can be set to 3x3x3, step-length can for (2,
2,2).Each stage is sampled using BN and ReLU.
It should be noted that the relatively deep in neural network includes more semantic informations, it include more empty compared with shallow-layer
Between information, using hierarchical fusion can be improved layered characteristic indicate ability.And due to the calculation amount of three-dimensional image segmentation task compared with
Greatly, the different numbers of plies can be being assembled every the stage, is having accumulated the layer of three different depths, in each stage to form HAU.
Above-mentioned attention unit AU includes the 6th convolutional layer, the 7th convolutional layer and second shape structure sheaf, and the 6th convolutional layer successively leads to
It crosses batch normalization BN and the linear unit R eLU of correction is connect with the 7th convolutional layer；Second shape structure sheaf isExp indicates exponential function, χi, c expression cthThe i of Feature Mapping on channelthValue.
It should be noted that the atrium sinistrum shape of different patients may have very big difference and convolutional neural networks
Down-sampling operation be easy with depth intensification and lose spatial information, the two factors can influence segmentation precision.In order to
Alleviate influence of above-mentioned two factor to segmentation precision, this is in each stage using attention mechanism as a mask branch
It is integrated into encoder network.By the attention unit, the value of Feature Mapping can be normalized, to obtain attention mask.
The output of above-mentioned hierarchical fusion module HAAM isMi,c(χ) is indicated
The output of mask branch, range are [0,1]；Fi,c(χ) indicates the output of trunk branch,Indicate dot-product,It indicates to press element
Direction summation.
It should be noted that execution point can be produced between mask branch and trunk branch after obtaining attention mask
Raw operation.And since the value of attention mask is to differ from 0 to 1, therefore can reduce feature using mask branch repetition making point and reflect
The value penetrated, and attention mask may destroy the superperformance of trunk branch.In order to solve this problem, residual error is this time used
The method of study carries out completely the same mapping between the outputting and inputting of trunk branch.
In order to preferably introduce the specific network structure of HAANet, below in conjunction with the layer shown in Fig. 2 based on attention
Explanation is introduced in secondary converging network structural schematic diagram.
As shown in Fig. 2, (a) is the overall network structure chart of HAANet, it (b) is the network structure of HAU, (c) for AU's
Network structure, (d) network structure for being HAAM.Wherein, HAAM includes HAU and AU, and AU includes the convolutional layer of 3x3x3, criticizes
The convolutional layer and sigmoid of normalization, ReLU, 1x1x1 are measured, sigmoid is also Logistic function, is used for hidden neuron
Output, value range are (0,1), and a real number can be mapped to the section of (0,1) by it, can be used to do two classification, in spy
Effect is relatively good when sign difference is more complicated or difference is not especially big.Second shape structure sheaf is being not shown in the figure.HAU includes volume
Lamination l1、l2、l3、l4、l5, each convolutional layer Conv is respectively connected with BN and ReLU.Concat is common in convolutional neural networks
One of function.
In the present embodiment, by being partitioned into the ROI region comprising three-dimensional atrium sinistrum from cardiac magnetic resonance images to be split,
Using the ROI region as the input of network model, calculation amount can be greatly reduced, while image background can also be substantially reduced
Interference, to improve the efficiency and accuracy rate of three-dimensional atrium sinistrum segmentation；ROI region is carried out using level converging network model
Segmentation, which is iterated by pantostrat of the layering aggregation module to same stage different depth melts
It closes, improves the shallow-layer of network model and the Fusion Features ability of deep layer, more preferable fuse information can be obtained, it will also be noted that unit is made
It is attached to each stage of encoder section, for mask branch so as to gradually enhance using the spatial information of shallow-layer with rich
The deep layer contour feature of rich semantic information substantially increases the three-dimensional left heart by combining level aggregation and attention mechanism
The efficiency and accuracy rate of room segmentation.
Embodiment two
Fig. 3 is referred to, for a kind of another process signal of three-dimensional atrium sinistrum dividing method provided by the embodiments of the present application
Block diagram, this method may comprise steps of:
Step S301, training sample and label information corresponding with training sample are obtained.
It should be noted that above-mentioned training sample refers to the cardiac magnetic resonance images comprising three-dimensional left atrial region, label
Information refers to the corresponding correct segmentation result information of training sample.
Step S302, according to label information, corresponding target ROI region is partitioned into from training sample.
Specifically, it is based on the corresponding correct segmentation result information of the training sample, judges be partitioned into target ROI region
Whether meet training requirement, if met, can use be partitioned into target ROI region and carry out subsequent training, if not
Meet, then can relocate target area, cuts the region of repositioning as target ROI region.
In some embodiments, the idiographic flow block diagram of the step S302 shown referring to fig. 4, it is above-mentioned according to label information,
The process that corresponding target ROI region is partitioned into from training sample specifically includes:
Step S401, training sample is uniformly adjusted to the first preset shape.
It is understood that above-mentioned first preset shape can be set according to actual needs, shape is by (z-axis, height
Degree, width) it determines.For example, can be set to (64,128,128).
Step S402, the training sample input of the first preset shape is trained in advance U-Net convolutional neural networks, obtain
Target ROI region testing result.
Step S403, each training sample is adjusted from the first preset shape to the original shape of each training sample.
Step S404, according to ROI region testing result and label information, corresponding target is partitioned into from training sample
ROI region.
That is, after the estimation results for obtaining target ROI region, it can be by the Adjusting Shape of each training sample
At respective original-shape, i.e., the parameters such as the z-axis, height, width of each sample are adjusted into back original parameter.Then, in basis
The ROI region testing result and label information, segmentation object ROI region.
Wherein, there may be error, that is, the region positioned is not the output result of pre-training U-Net convolutional neural networks
Actual atrium sinistrum region can make institute if also continuing cutting the region as the subsequent training of target ROI region progress
Trained network model inaccuracy, at this point, can then relocate target area, that is, redefines atrium sinistrum region, so
Corresponding region is cut again afterwards as target ROI region, to guarantee the training accuracy of following model.
In some embodiments, above-mentioned according to target ROI region testing result and label information, divide from training sample
The detailed process of corresponding target ROI region includes: to judge whether target ROI region testing result includes label information out；When
When target ROI region testing result does not include label information, then target ROI region is extended to goal-selling region；From training
Goal-selling region is cut out in sample, using goal-selling region as target ROI region；When target ROI region testing result
Comprising label information, target ROI region is cut out from training sample.
It is understood that since note is the corresponding correct segmentation result of training sample, if identified target ROI
Not comprising the note in region, it may be considered that positioning mistake, the region oriented is not actual left atrial region.?
When estimating mistake, target area can be extended, to ensure that the i.e. three-dimensional left atrial region of entire original object is specifically tailored out
Come.If in target ROI region including the note, it may be considered that positioning is correct, that is, the left atrial region estimated is correct
, estimated region can be directly cut at this time as target ROI region.
It is above-mentioned target ROI region extended into goal-selling region refer to expanded based on the target ROI region estimated
Exhibition, after extending to certain region, then can cut the goal-selling region, as target ROI region, i.e., this is pre-
If input of the target area as training pattern.
As can be seen that being trained the accuracy that can guarantee model training by note information.
Step S303, according to target ROI region, model training is carried out to the level converging network model pre-established.
It certainly, after training, can be by being tested using test sample, when the relevant parameter of model meets centainly
Requirement after recycle the model carry out cardiac magnetic resonance images to be split three-dimensional atrium sinistrum segmentation.
Step S304, cardiac magnetic resonance images to be split are obtained.
Step S305, ROI region is partitioned into from cardiac magnetic resonance images to be split, ROI region is comprising the three-dimensional left heart
The region in room.
Step S306, ROI region input is trained in advance level converging network model, obtains cardiac magnetic resonance to be split
The segmentation result of image.
It should be noted that step S304~S306 is identical as the step S101~S103 of above-described embodiment one, it is specific to be situated between
It continues and refers to corresponding contents above, details are not described herein.
In order to compare the embodiment of the present application three-dimensional atrium sinistrum dividing method and other dividing methods accuracy rate, below will
It is illustrated in conjunction with a specific experiment.
The data set that this experiment uses provides 100 trained numbers to divide in atrium segmentation challenge in 2018 for atrium sinistrum
According to.The original resolution of data is 0.625 × 0.625 × 0.625mm3, and each sample has 88 on Z axis.Because not sending out
Cloth verify data, from training data concentration, isolating 10 patient datas as verify data tests proposed mould at random
Therefore type has 90 patient datas to give training, 10 patient datas verify.In the training process, we are by the area ROI
The Adjusting Shape in domain is fixed size (88,80,128), to input HAANet model.Data enhancing can be used for extending ours
Training set, to mitigate the over adaptation in training process.On the basis of random, data rotate 0~2 π degree along Z axis, 0.8~
It is zoomed in and out in 1.2 ranges, and carries out mirror image and translation transformation along Z axis.In addition, additionally use gamma transformation, in the range of 0.8~
1.3.Finally, after completing above-mentioned conversion, use Z-score standard it is worth noting that, all conversions be not with 50% it is general
What rate executed.By applying these random data Enhancement Methods, can theoretically there be unlimited training data, these random data
Enhancement Method was executed in the training stage of each iteration.
Learnership about experiment: in this experiment, for three-dimensional U-net and HAANet network, the of network
One stage set 8 for the port number of convolutional layer, and the number of channel is increased one times.Down-sampling operation each time is performed until
5th stage.Training batch size is set as 8, HAANet network and is optimized by Adam algorithm.Each period has 100 times
Iteration.Learning rate is initialized as 1e-3, and decays with the factor 0.1 when there is no when any update within 5 periods.Dice
Loss is used as loss function, and indicates are as follows:
Wherein, ytrueIndicate training label, ypredIndicate the output of our networks.
HAANet network is realized that it uses TensorFlow 1.4 as rear end by Keras 2.1.5, which exists
It is trained and has been tested on 1080 Ti GPU of NVIDIA GeForce GTX, which is the ubuntu at one 64
It is developed on 16.04 platforms, uses Intel R CoreTM i5-7640X CPU@4.00GHZ × 4,32GB memory (RAM).
Setting about this experiment: the purpose of experiment is to assess the validity of HAU and AU, and evaluation criterion is dice system
Number shows the matching degree of predicted value and true value.Six LA split-run tests, i.e. UNET-2, Hanet-2, HAANet-2 have been done,
UNET-3, Hanet-3, HAANet-3, suffix number indicate the convolution number of plies in each stage in network, by the way, UNET-2
It is also used for ROI detection, but it is different to input shape.HANet indicates the architecture without paying attention to unit, and HAANet indicates our bases
In the layering aggregation network of attention.In order to coequally compare the difference of these models, all six networks are all provided with
Identical hyper parameter and identical learnership mentioned above.
About quantitative result: the identical 6 kinds of network methods of above-mentioned experimental setup are divided applied to LA.In this six networks
In, UNet-2 represents baseline, and each stage, there are two convolutional layer, UNet-3 to indicate volume of each stage three relative to UNet-2
Lamination.HANet-2 and HANet-3 indicate the network of proposal, as shown in Fig. 2 (a), but without paying attention to module, HANet-2 with
In HANet-3, HAU has 2 layers and 3 layers of convolutional layer respectively, and to be represented in Fig. 2 (b) include the HAU with two layers of convolutional layer
l1、l2And l3Submodule inside.By the way that AU is integrated into acquisition HAANet-2 and HAANet-3 in HANet-2 and HANet-3,
This six networks will be gradually tested to advanced optimize HAANet-2 and HAANet-3., to prove the validity of HAU and AU.
Following table 1 shows comparison result of six kinds of network methods in verify data.Six the experimental results showed that, with warp
Allusion quotation medical image segmentation structure U-net is combined, and HAU and AU, which are combined, can obtain better segmentation effect.The result shows that base
In the Aggregation Model of attention be a kind of up-and-coming LA segmentation strategy.
Comparison result of the 1 six kinds of network methods of table to verify data
Performance about layering aggregation unit: as shown in table 1, no matter each stage has 2 layers or 3 layers, the dice of HANets
Value is all higher than the dice value of UNets.Traditional stacking convolutional layer is fused to tree construction by layering aggregation module, richer by learning
Rich feature improves the performance of network.The Feature Mapping of different levels is compared, shallow-layer feature can be retained, is integrated not
With the convolutional layer of received field size, this is very important semantic segmentation.
Performance about attention unit: attention mechanism is that Web Cams is forced to have in one kind of this target of atrium sinistrum
Efficacious prescriptions method.Normalization intermediate mask is generated by second shape structure function, the HAANets proposed learns plan using remaining attention
Slightly, the attention mapping of shallow-layer convolutional layer is attached in the output of entire block in each stage, not only can avoids having
The potentiality for breaking through backbone-stub superperformance, also further improve the performance of HAANet network.In table 1, HAANets ratio
HANets shows more preferably, it is shown that remnants pay attention to the validity of unit.
About segmentation result: by combining HAU and AU, HAANet-3 obtains 93.00 highest dice.It is right
The dice value of ten verifying patient datas of UNET-2 and HAANet-3 is as shown in figure 5, its vertical axis indicates dice value, trunnion axis
Indicate 10 patient's samples (A~J).It is observed that the dice ratio of our HAANet-3 of almost all of verify data
UNET-2 wants higher, in addition to the dice value of HAANet-3 is slightly below the UNET-2 of sample I.This comparison result further illustrates
Application prospect of the HAANet proposed in the three-dimensional segmentation of atrium sinistrum.
Fig. 6 (a) is the segmentation result comparison schematic diagram of UNet-2 and HAANet-3, which illustrate the HAANet-3 of proposal and
The automatic 2D LA segmentation result of UNET-2.It can be seen that HAANet-3 ratio UNET-2 has good advantage, especially in the hard of MRI
Part.Fig. 6 (b) is three-dimensional segmentation result schematic diagram, is the three-dimensional LA segmentation that 6 kinds of different patients rebuild which depict ITK-SNAP
As a result, the column of top one indicate ground truth, a following column indicate the three-dimensional segmentation result of HAANet-3.
In the present embodiment, ROI region is cut using U-Net network model and label information, model training can be made more
It is accurate to add.And by being partitioned into the ROI region comprising three-dimensional atrium sinistrum, the area Jiang Gai ROI from cardiac magnetic resonance images to be split
Input of the domain as network model can greatly reduce calculation amount, while can also substantially reduce the interference of image background, thus
Improve the efficiency and accuracy rate of three-dimensional atrium sinistrum segmentation；ROI region is split using level converging network model, the layer
Secondary converging network model is iterated fusion by pantostrat of the layering aggregation module to same stage different depth, improves net
The shallow-layer of network model and the Fusion Features ability of deep layer, can obtain more preferable fuse information, it will also be noted that unit is as mask branch
It is attached to each stage of encoder section, there is abundant semantic information so as to gradually enhance using the spatial information of shallow-layer
Deep layer contour feature substantially increase the effect of three-dimensional atrium sinistrum segmentation by combining level aggregation and attention mechanism
Rate and accuracy rate.
It should be understood that the size of the serial number of each step is not meant that the order of the execution order in above-described embodiment, each process
Execution sequence should be determined by its function and internal logic, the implementation process without coping with the embodiment of the present application constitutes any limit
It is fixed.
Embodiment three
Fig. 7 is referred to, it, should for a kind of structural schematic block diagram of three-dimensional atrium sinistrum segmenting device provided by the embodiments of the present application
Device may include:
Module 71 is obtained, for obtaining cardiac magnetic resonance images to be split；
ROI region divides module 72, and for being partitioned into ROI region from cardiac magnetic resonance images to be split, ROI region is
Region comprising three-dimensional atrium sinistrum；
Divide module 73, for the level converging network model that ROI region input is trained in advance, obtains heart to be split
The segmentation result of magnetic resonance image；
Wherein, level converging network model is the U-Net convolutional neural networks for including encoder path and decoder-path
Model, encoder path include at least one layering aggregation module, layering aggregation module include as trunk branch level it is poly-
Close unit and the attention unit as mask branch.
In one possible implementation, above-mentioned ROI region segmentation module includes:
ROI region detection unit, for passing through pre-training U-Net convolutional neural networks to cardiac magnetic resonance images to be split
It is detected, obtains ROI region testing result；Wherein, every level-one of pre-training U-Net convolutional neural networks has two convolution
Layer；
ROI region cuts unit, for being cut out according to ROI region testing result to cardiac magnetic resonance images to be split
It cuts, obtains ROI region.
In one possible implementation, level polymerized unit includes the first convolutional layer, the second convolutional layer, third convolution
Layer, Volume Four lamination and the 5th convolutional layer are respectively connected with batch after each convolutional layer and normalize and correction linear unit；
Wherein, the first convolutional layer and the second convolutional layer cascade, and cascade result is inputted third convolutional layer；To third convolutional layer
It carries out convolution algorithm and obtains Volume Four lamination, third convolutional layer is connected with Volume Four lamination generates the 5th convolutional layer；
Notice that unit includes the 6th convolutional layer, the 7th convolutional layer and second shape structure sheaf, the 6th convolutional layer passes sequentially through batch
Normalization and correction linear unit are connect with the 7th convolutional layer；Second shape structure sheaf isExp is indicated
Exponential function, χi, c expression cthThe i of Feature Mapping on channelthValue；
The output of hierarchical fusion module isMi,c(χ) indicates mask branch
Output, range be [0,1]；Fi,c(χ) indicates the output of trunk branch,Indicate dot-product,Expression is asked by element direction
With.
In one possible implementation, above-mentioned apparatus further include:
Training sample obtains module, for obtaining training sample and label information corresponding with training sample；
Target ROI region divides module, for being partitioned into corresponding target ROI from training sample according to label information
Region；
Training module, for carrying out model instruction to the level converging network model pre-established according to target ROI region
Practice.
In one possible implementation, above-mentioned target ROI region segmentation module includes:
The first adjustment unit, for uniformly adjusting training sample to the first preset shape；
Detection unit, the U-Net convolutional neural networks trained in advance for the training sample input by the first preset shape,
Obtain target ROI region testing result；
Second adjustment unit, for adjusting each training sample from the first preset shape to the script shape of each training sample
Shape；
Cutting unit, for being partitioned into corresponding mesh from training sample according to ROI region testing result and label information
Mark ROI region.
In one possible implementation, above-mentioned cutting unit includes:
Judgment sub-unit, for judging whether target ROI region testing result includes label information；
Subelement is extended, for when target ROI region testing result does not include label information, then by target ROI region
It is extended to goal-selling region；
First cut subelement, for cutting out goal-selling region from training sample, using goal-selling region as
Target ROI region；
Second cuts subelement, for including label information when target ROI region testing result, cuts from training sample
Target ROI region out.
It should be noted that the present embodiment is corresponding with the embodiment of above-mentioned each three-dimensional atrium sinistrum dividing method, correlation is situated between
It continues and refers to corresponding contents above, details are not described herein.
In the present embodiment, by being partitioned into the ROI region comprising three-dimensional atrium sinistrum from cardiac magnetic resonance images to be split,
Using the ROI region as the input of network model, calculation amount can be greatly reduced, while image background can also be substantially reduced
Interference, to improve the efficiency and accuracy rate of three-dimensional atrium sinistrum segmentation；ROI region is carried out using level converging network model
Segmentation, which is iterated by pantostrat of the layering aggregation module to same stage different depth melts
It closes, improves the shallow-layer of network model and the Fusion Features ability of deep layer, more preferable fuse information can be obtained, it will also be noted that unit is made
It is attached to each stage of encoder section, for mask branch so as to gradually enhance using the spatial information of shallow-layer with rich
The deep layer contour feature of rich semantic information substantially increases the three-dimensional left heart by combining level aggregation and attention mechanism
The efficiency and accuracy rate of room segmentation.
Example IV
Fig. 8 is the schematic diagram for the terminal device that one embodiment of the application provides.As shown in figure 8, the terminal of the embodiment is set
Standby 8 include: processor 80, memory 81 and are stored in the meter that can be run in the memory 81 and on the processor 80
Calculation machine program 82.The processor 80 realizes that above-mentioned each three-dimensional atrium sinistrum dividing method is real when executing the computer program 82
Apply the step in example, such as step S101 to S103 shown in FIG. 1.Alternatively, the processor 80 executes the computer program
Each module or the function of unit in above-mentioned each Installation practice, such as the function of module 71 to 73 shown in Fig. 7 are realized when 82.
Illustratively, the computer program 82 can be divided into one or more modules or unit, it is one or
The multiple modules of person or unit are stored in the memory 81, and are executed by the processor 80, to complete the application.It is described
One or more modules or unit can be the series of computation machine program instruction section that can complete specific function, which uses
In implementation procedure of the description computer program 82 in the terminal device 8.For example, the computer program 82 can be by
It is divided into and obtains module, ROI region segmentation module and divide module, each module concrete function is as follows:
Module is obtained, for obtaining cardiac magnetic resonance images to be split；ROI region divides module, is used for from the heart to be split
ROI region is partitioned into dirty magnetic resonance image, ROI region is the region comprising three-dimensional atrium sinistrum；Divide module, is used for ROI
Region input level converging network model trained in advance, obtains the segmentation result of cardiac magnetic resonance images to be split；Wherein, layer
Secondary converging network model is the U-Net convolutional neural networks model for including encoder path and decoder-path, encoder path
Including at least one layering aggregation module, layering aggregation module includes as the level polymerized unit of trunk branch and as mask
The attention unit of branch.
The terminal device 8 can be the calculating such as desktop PC, notebook, palm PC and cloud server and set
It is standby.The terminal device may include, but be not limited only to, processor 80, memory 81.It will be understood by those skilled in the art that Fig. 8
The only example of terminal device 8 does not constitute the restriction to terminal device 8, may include than illustrating more or fewer portions
Part perhaps combines certain components or different components, such as the terminal device can also include input-output equipment, net
Network access device, bus etc..
Alleged processor 80 can be central processing unit (Central Processing Unit, CPU), can also be
Other general processors, digital signal processor (Digital Signal Processor, DSP), specific integrated circuit
(Application Specific Integrated Circuit, ASIC), ready-made programmable gate array (Field-
Programmable Gate Array, FPGA) either other programmable logic device, discrete gate or transistor logic,
Discrete hardware components etc..General processor can be microprocessor or the processor is also possible to any conventional processor
Deng.
The memory 81 can be the internal storage unit of the terminal device 8, such as the hard disk or interior of terminal device 8
It deposits.The memory 81 is also possible to the External memory equipment of the terminal device 8, such as be equipped on the terminal device 8
Plug-in type hard disk, intelligent memory card (Smart Media Card, SMC), secure digital (Secure Digital, SD) card dodge
Deposit card (Flash Card) etc..Further, the memory 81 can also both include the storage inside list of the terminal device 8
Member also includes External memory equipment.The memory 81 is for storing needed for the computer program and the terminal device
Other programs and data.The memory 81 can be also used for temporarily storing the data that has exported or will export.
It is apparent to those skilled in the art that for convenience of description and succinctly, only with above-mentioned each function
Can unit, module division progress for example, in practical application, can according to need and by above-mentioned function distribution by different
Functional unit, module are completed, i.e., the internal structure of described device is divided into different functional unit or module, more than completing
The all or part of function of description.Each functional unit in embodiment, module can integrate in one processing unit, can also
To be that each unit physically exists alone, can also be integrated in one unit with two or more units, it is above-mentioned integrated
Unit both can take the form of hardware realization, can also realize in the form of software functional units.In addition, each function list
Member, the specific name of module are also only for convenience of distinguishing each other, the protection scope being not intended to limit this application.Above system
The specific work process of middle unit, module, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, is not described in detail or remembers in some embodiment
The part of load may refer to the associated description of other embodiments.
Those of ordinary skill in the art may be aware that list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and algorithm steps can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, the specific application and design constraint depending on technical solution.Professional technician
Each specific application can be used different methods to achieve the described function, but this realization is it is not considered that exceed
Scope of the present application.
In embodiment provided herein, it should be understood that disclosed device, terminal device and method, it can be with
It realizes by another way.For example, device described above, terminal device embodiment are only schematical, for example, institute
The division of module or unit is stated, only a kind of logical function partition, there may be another division manner in actual implementation, such as
Multiple units or components can be combined or can be integrated into another system, or some features can be ignored or not executed.Separately
A bit, shown or discussed mutual coupling or direct-coupling or communication connection can be through some interfaces, device
Or the INDIRECT COUPLING or communication connection of unit, it can be electrical property, mechanical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
It, can also be in addition, each functional unit in each embodiment of the application can integrate in one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
If the integrated module or unit are realized in the form of SFU software functional unit and sell as independent product
Or it in use, can store in a computer readable storage medium.Based on this understanding, the application realizes above-mentioned reality
The all or part of the process in a method is applied, relevant hardware can also be instructed to complete by computer program, it is described
Computer program can be stored in a computer readable storage medium, the computer program when being executed by processor, it can be achieved that
The step of above-mentioned each embodiment of the method.Wherein, the computer program includes computer program code, the computer program
Code can be source code form, object identification code form, executable file or certain intermediate forms etc..Computer-readable Jie
Matter may include: can carry the computer program code any entity or device, recording medium, USB flash disk, mobile hard disk,
Magnetic disk, CD, computer storage, read-only memory (ROM, Read-Only Memory), random access memory (RAM,
Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium etc..It should be noted that described
The content that computer-readable medium includes can carry out increasing appropriate according to the requirement made laws in jurisdiction with patent practice
Subtract, such as does not include electric carrier signal and electricity according to legislation and patent practice, computer-readable medium in certain jurisdictions
Believe signal.
Embodiment described above is only to illustrate the technical solution of the application, rather than its limitations；Although referring to aforementioned reality
Example is applied the application is described in detail, those skilled in the art should understand that: it still can be to aforementioned each
Technical solution documented by embodiment is modified or equivalent replacement of some of the technical features；And these are modified
Or replacement, the spirit and scope of each embodiment technical solution of the application that it does not separate the essence of the corresponding technical solution should all
Comprising within the scope of protection of this application.
Claims (10)
1. a kind of three-dimensional atrium sinistrum dividing method characterized by comprising
Obtain cardiac magnetic resonance images to be split；
ROI region is partitioned into from the cardiac magnetic resonance images to be split, the ROI region is the area comprising three-dimensional atrium sinistrum
Domain；
By ROI region input level converging network model trained in advance, the cardiac magnetic resonance images to be split are obtained
Segmentation result；
Wherein, the level converging network model is the U-Net convolutional neural networks for including encoder path and decoder-path
Model, the encoder path include at least one layering aggregation module, and the layering aggregation module includes being used as trunk branch
Level polymerized unit and attention unit as mask branch.
2. three-dimensional atrium sinistrum according to claim 1 dividing method, which is characterized in that described from the heart magnetic to be split
ROI region is partitioned into resonance image, comprising:
The cardiac magnetic resonance images to be split are detected by pre-training U-Net convolutional neural networks, obtain ROI region
Testing result；Wherein, every level-one of the pre-training U-Net convolutional neural networks has two convolutional layers；
The cardiac magnetic resonance images to be split are cut according to the ROI region testing result, obtain the area ROI
Domain.
3. three-dimensional atrium sinistrum dividing method according to claim 1, which is characterized in that the level polymerized unit includes the
One convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolutional layer, are respectively connected with after each convolutional layer batch
Amount normalization and correction linear unit；
Wherein, first convolutional layer and second convolutional layer cascade, and cascade result is inputted the third convolutional layer；To institute
It states third convolutional layer progress convolution algorithm and obtains the Volume Four lamination, the third convolutional layer is connected with the Volume Four lamination
Generate the 5th convolutional layer；
The attention unit includes the 6th convolutional layer, the 7th convolutional layer and second shape structure sheaf, and the 6th convolutional layer passes sequentially through
Batch normalization and the correction linear unit are connect with the 7th convolutional layer；The second shape structure sheaf isExp indicates exponential function, χi, c expression cthThe i of Feature Mapping on channelthValue；
The output of the hierarchical fusion module isMi,c(χ) indicates mask branch
Output, range be [0,1]；Fi,c(χ) indicates the output of trunk branch,Indicate dot-product,Expression is asked by element direction
With.
4. three-dimensional atrium sinistrum dividing method according to any one of claims 1 to 3, which is characterized in that it is described obtain to
Before segmentation cardiac magnetic resonance images, further includes:
Obtain training sample and label information corresponding with the training sample；
According to the label information, corresponding target ROI region is partitioned into from the training sample；
According to the target ROI region, model training is carried out to the level converging network model pre-established.
5. three-dimensional atrium sinistrum dividing method according to claim 4, which is characterized in that it is described according to the label information,
Corresponding target ROI region is partitioned into from the training sample, comprising:
The training sample is uniformly adjusted to the first preset shape；
By the training sample input of first preset shape U-Net convolutional neural networks trained in advance, the area target ROI is obtained
Domain testing result；
Each training sample is adjusted from first preset shape to the original shape of each training sample；
According to the ROI region testing result and the label information, the corresponding mesh is partitioned into from the training sample
Mark ROI region.
6. three-dimensional atrium sinistrum according to claim 5 dividing method, which is characterized in that described according to the area the target ROI
Domain testing result and the label information are partitioned into the corresponding target ROI region from the training sample, comprising:
Judge whether the target ROI region testing result includes the label information；
When the target ROI region testing result does not include the label information, then the target ROI region is extended to pre-
If target area；
The goal-selling region is cut out from the training sample, using the goal-selling region as the target ROI
Region；
When the target ROI region testing result includes the label information, the mesh is cut out from the training sample
Mark ROI region.
7. a kind of three-dimensional atrium sinistrum segmenting device characterized by comprising
Module is obtained, for obtaining cardiac magnetic resonance images to be split；
ROI region divides module, for being partitioned into ROI region, the ROI region from the cardiac magnetic resonance images to be split
For the region comprising three-dimensional atrium sinistrum；
Divide module, for the level converging network model that ROI region input is trained in advance, obtains the heart to be split
The segmentation result of dirty magnetic resonance image；
Wherein, the level converging network model is the U-Net convolutional neural networks for including encoder path and decoder-path
Model, the encoder path include at least one layering aggregation module, and the layering aggregation module includes being used as trunk branch
Level polymerized unit and attention unit as mask branch.
8. three-dimensional atrium sinistrum according to claim 7 segmenting device, which is characterized in that the ROI region divides module packet
It includes:
ROI region detection unit, for passing through pre-training U-Net convolutional neural networks to the cardiac magnetic resonance images to be split
It is detected, obtains ROI region testing result；Wherein, every level-one of the pre-training U-Net convolutional neural networks has two
Convolutional layer；
ROI region cuts unit, for being carried out according to the ROI region testing result to the cardiac magnetic resonance images to be split
It cuts, obtains the ROI region.
9. a kind of terminal device, which is characterized in that in the memory and can be in institute including memory, processor and storage
The computer program run on processor is stated, the processor realizes such as claim 1 to 6 times when executing the computer program
The step of one three-dimensional atrium sinistrum dividing method.
10. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has computer journey
Sequence realizes the three-dimensional atrium sinistrum dividing method as described in any one of claim 1 to 6 when the computer program is executed by processor
The step of.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811535118.9A CN109801294A (en) | 2018-12-14 | 2018-12-14 | Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium |
| PCT/CN2019/124311 WO2020119679A1 (en) | 2018-12-14 | 2019-12-10 | Three-dimensional left atrium segmentation method and apparatus, terminal device, and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811535118.9A CN109801294A (en) | 2018-12-14 | 2018-12-14 | Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109801294A true CN109801294A (en) | 2019-05-24 |
Family
ID=66556774
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811535118.9A Withdrawn CN109801294A (en) | 2018-12-14 | 2018-12-14 | Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109801294A (en) |
| WO (1) | WO2020119679A1 (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210487A (en) * | 2019-05-30 | 2019-09-06 | 上海商汤智能科技有限公司 | A kind of image partition method and device, electronic equipment and storage medium |
| CN110288609A (en) * | 2019-05-30 | 2019-09-27 | 南京师范大学 | A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance |
| CN110310280A (en) * | 2019-07-10 | 2019-10-08 | 广东工业大学 | Hepatic duct and the image-recognizing method of calculus, system, equipment and storage medium |
| CN110428431A (en) * | 2019-07-12 | 2019-11-08 | 广东省人民医院(广东省医学科学院) | A kind of dividing method, device, equipment and the storage medium of cardiologic medical image |
| CN110570416A (en) * | 2019-09-12 | 2019-12-13 | 杭州海睿博研科技有限公司 | Method for visualization and 3D printing of multi-modal cardiac images |
| CN110599502A (en) * | 2019-09-06 | 2019-12-20 | 江南大学 | Skin lesion segmentation method based on deep learning |
| CN110853045A (en) * | 2019-09-24 | 2020-02-28 | 西安交通大学 | Vascular wall segmentation method and device based on nuclear magnetic resonance image and storage medium |
| CN110910364A (en) * | 2019-11-16 | 2020-03-24 | 应急管理部沈阳消防研究所 | Method for detecting electrical equipment easy to cause fire in three-section fire scene based on deep neural network |
| CN111281387A (en) * | 2020-03-09 | 2020-06-16 | 中山大学 | Left atrium and atrial scar segmentation method and device based on artificial neural network |
| WO2020119679A1 (en) * | 2018-12-14 | 2020-06-18 | 深圳先进技术研究院 | Three-dimensional left atrium segmentation method and apparatus, terminal device, and storage medium |
| CN111553895A (en) * | 2020-04-24 | 2020-08-18 | 中国人民解放军陆军军医大学第二附属医院 | Magnetic resonance left atrium segmentation method based on multi-scale fine granularity |
| CN112435247A (en) * | 2020-11-30 | 2021-03-02 | 中国科学院深圳先进技术研究院 | Patent foramen ovale detection method, system, terminal and storage medium |
| CN112508949A (en) * | 2021-02-01 | 2021-03-16 | 之江实验室 | Method for automatically segmenting left ventricle of SPECT three-dimensional reconstruction image |
| CN114066913A (en) * | 2022-01-12 | 2022-02-18 | 广东工业大学 | Heart image segmentation method and system |
| CN114155208A (en) * | 2021-11-15 | 2022-03-08 | 中国科学院深圳先进技术研究院 | Atrial fibrillation assessment method and device based on deep learning |
| CN116385468A (en) * | 2023-06-06 | 2023-07-04 | 浙江大学 | System based on zebra fish heart parameter image analysis software generation |
| CN117456191A (en) * | 2023-12-15 | 2024-01-26 | 武汉纺织大学 | Semantic segmentation method based on three-branch network structure under complex environment |
Families Citing this family (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111754534A (en) * | 2020-07-01 | 2020-10-09 | 杭州脉流科技有限公司 | CT left ventricle short axis image segmentation method and device based on deep neural network, computer equipment and storage medium |
| CN111986204B (en) * | 2020-07-23 | 2023-06-16 | 中山大学 | Polyp segmentation method, device and storage medium |
| CN111968122B (en) * | 2020-08-27 | 2023-07-28 | 广东工业大学 | Textile material CT image segmentation method and device based on convolutional neural network |
| CN112348780A (en) * | 2020-10-26 | 2021-02-09 | 首都医科大学附属北京安贞医院 | Fetal heart measuring method and device |
| CN112766313B (en) * | 2020-12-29 | 2023-11-14 | 厦门贝启科技有限公司 | Crystal segmentation and positioning method, device, equipment and medium based on U-net structure |
| CN112966687B (en) * | 2021-02-01 | 2024-01-19 | 深圳市优必选科技股份有限公司 | Image segmentation model training method and device and communication equipment |
| CN112802034B (en) * | 2021-02-04 | 2024-04-12 | 精英数智科技股份有限公司 | Image segmentation and identification methods, model construction methods and devices, and electronic equipment |
| CN112967294A (en) * | 2021-03-11 | 2021-06-15 | 西安智诊智能科技有限公司 | Liver CT image segmentation method and system |
| CN113112475B (en) * | 2021-04-13 | 2023-04-18 | 五邑大学 | Traditional Chinese medicine ear five-organ region segmentation method and device based on machine learning |
| CN113223014B (en) * | 2021-05-08 | 2023-04-28 | 中国科学院自动化研究所 | Brain image analysis system, method and equipment based on data enhancement |
| CN113592771B (en) * | 2021-06-24 | 2023-12-15 | 深圳大学 | Image segmentation method |
| CN113505535B (en) * | 2021-07-08 | 2023-11-10 | 重庆大学 | Motor life prediction method based on gating self-adaptive hierarchical attention unit network |
| CN114663431B (en) * | 2022-05-19 | 2022-08-30 | 浙江大学 | Pancreatic tumor image segmentation method and system based on reinforcement learning and attention |
| CN116797787B (en) * | 2023-05-22 | 2024-01-02 | 中国地质大学（武汉） | Remote sensing image semantic segmentation method based on cross-modal fusion and graph neural network |
| 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 |
| CN117349714B (en) * | 2023-12-06 | 2024-02-13 | 中南大学 | Classification method, system, equipment and medium for medical image of Alzheimer disease |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108537784B (en) * | 2018-03-30 | 2021-08-24 | 四川元匠科技有限公司 | CT image pulmonary nodule detection method based on deep learning |
| CN108537793A (en) * | 2018-04-17 | 2018-09-14 | 电子科技大学 | A kind of pulmonary nodule detection method based on improved u-net networks |
| CN108765369B (en) * | 2018-04-20 | 2023-05-02 | 平安科技（深圳）有限公司 | Method, apparatus, computer device and storage medium for detecting lung nodule |
| CN108615236A (en) * | 2018-05-08 | 2018-10-02 | 上海商汤智能科技有限公司 | A kind of image processing method and electronic equipment |
| CN109801294A (en) * | 2018-12-14 | 2019-05-24 | 深圳先进技术研究院 | Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium |
-
2018
- 2018-12-14 CN CN201811535118.9A patent/CN109801294A/en not_active Withdrawn
-
2019
- 2019-12-10 WO PCT/CN2019/124311 patent/WO2020119679A1/en active Application Filing
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020119679A1 (en) * | 2018-12-14 | 2020-06-18 | 深圳先进技术研究院 | Three-dimensional left atrium segmentation method and apparatus, terminal device, and storage medium |
| CN110288609A (en) * | 2019-05-30 | 2019-09-27 | 南京师范大学 | A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance |
| CN110210487A (en) * | 2019-05-30 | 2019-09-06 | 上海商汤智能科技有限公司 | A kind of image partition method and device, electronic equipment and storage medium |
| CN110310280B (en) * | 2019-07-10 | 2021-05-11 | 广东工业大学 | Image recognition method, system, equipment and storage medium for hepatobiliary duct and calculus |
| CN110310280A (en) * | 2019-07-10 | 2019-10-08 | 广东工业大学 | Hepatic duct and the image-recognizing method of calculus, system, equipment and storage medium |
| CN110428431A (en) * | 2019-07-12 | 2019-11-08 | 广东省人民医院(广东省医学科学院) | A kind of dividing method, device, equipment and the storage medium of cardiologic medical image |
| CN110428431B (en) * | 2019-07-12 | 2022-12-16 | 广东省人民医院(广东省医学科学院) | Method, device and equipment for segmenting cardiac medical image and storage medium |
| CN110599502A (en) * | 2019-09-06 | 2019-12-20 | 江南大学 | Skin lesion segmentation method based on deep learning |
| CN110599502B (en) * | 2019-09-06 | 2023-07-11 | 江南大学 | Skin lesion segmentation method based on deep learning |
| CN110570416A (en) * | 2019-09-12 | 2019-12-13 | 杭州海睿博研科技有限公司 | Method for visualization and 3D printing of multi-modal cardiac images |
| CN110853045A (en) * | 2019-09-24 | 2020-02-28 | 西安交通大学 | Vascular wall segmentation method and device based on nuclear magnetic resonance image and storage medium |
| CN110853045B (en) * | 2019-09-24 | 2022-02-11 | 西安交通大学 | Vascular wall segmentation method and device based on nuclear magnetic resonance image and storage medium |
| CN110910364A (en) * | 2019-11-16 | 2020-03-24 | 应急管理部沈阳消防研究所 | Method for detecting electrical equipment easy to cause fire in three-section fire scene based on deep neural network |
| CN111281387B (en) * | 2020-03-09 | 2024-03-26 | 中山大学 | Segmentation method and device for left atrium and atrial scar based on artificial neural network |
| CN111281387A (en) * | 2020-03-09 | 2020-06-16 | 中山大学 | Left atrium and atrial scar segmentation method and device based on artificial neural network |
| CN111553895B (en) * | 2020-04-24 | 2022-08-02 | 中国人民解放军陆军军医大学第二附属医院 | Magnetic resonance left atrium segmentation method based on multi-scale fine granularity |
| CN111553895A (en) * | 2020-04-24 | 2020-08-18 | 中国人民解放军陆军军医大学第二附属医院 | Magnetic resonance left atrium segmentation method based on multi-scale fine granularity |
| CN112435247B (en) * | 2020-11-30 | 2022-03-25 | 中国科学院深圳先进技术研究院 | Patent foramen ovale detection method, system, terminal and storage medium |
| WO2022110445A1 (en) * | 2020-11-30 | 2022-06-02 | 中国科学院深圳先进技术研究院 | Patent foramen ovale detection method and system, and terminal and storage medium |
| CN112435247A (en) * | 2020-11-30 | 2021-03-02 | 中国科学院深圳先进技术研究院 | Patent foramen ovale detection method, system, terminal and storage medium |
| CN112508949A (en) * | 2021-02-01 | 2021-03-16 | 之江实验室 | Method for automatically segmenting left ventricle of SPECT three-dimensional reconstruction image |
| CN114155208B (en) * | 2021-11-15 | 2022-07-08 | 中国科学院深圳先进技术研究院 | Atrial fibrillation assessment method and device based on deep learning |
| CN114155208A (en) * | 2021-11-15 | 2022-03-08 | 中国科学院深圳先进技术研究院 | Atrial fibrillation assessment method and device based on deep learning |
| CN114066913B (en) * | 2022-01-12 | 2022-04-22 | 广东工业大学 | Heart image segmentation method and system |
| CN114066913A (en) * | 2022-01-12 | 2022-02-18 | 广东工业大学 | Heart image segmentation method and system |
| CN116385468A (en) * | 2023-06-06 | 2023-07-04 | 浙江大学 | System based on zebra fish heart parameter image analysis software generation |
| CN116385468B (en) * | 2023-06-06 | 2023-09-01 | 浙江大学 | System based on zebra fish heart parameter image analysis software generation |
| CN117456191A (en) * | 2023-12-15 | 2024-01-26 | 武汉纺织大学 | Semantic segmentation method based on three-branch network structure under complex environment |
| CN117456191B (en) * | 2023-12-15 | 2024-03-08 | 武汉纺织大学 | Semantic segmentation method based on three-branch network structure under complex environment |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020119679A1 (en) | 2020-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109801294A (en) | Three-dimensional atrium sinistrum dividing method, device, terminal device and storage medium | |
| Sun et al. | An adversarial learning approach to medical image synthesis for lesion detection | |
| Babier et al. | Knowledge‐based automated planning with three‐dimensional generative adversarial networks | |
| Ren et al. | Interleaved 3D‐CNN s for joint segmentation of small‐volume structures in head and neck CT images | |
| Chen et al. | MMFNet: A multi-modality MRI fusion network for segmentation of nasopharyngeal carcinoma | |
| CN110110617A (en) | Medical image dividing method, device, electronic equipment and storage medium | |
| CN109785334A (en) | Cardiac magnetic resonance images dividing method, device, terminal device and storage medium | |
| CN110689543A (en) | Improved convolutional neural network brain tumor image segmentation method based on attention mechanism | |
| CN109300107A (en) | Patch processing method, device and the calculating equipment of magnetic resonance vascular wall imaging | |
| CN108257134A (en) | Nasopharyngeal Carcinoma Lesions automatic division method and system based on deep learning | |
| Li et al. | Attention based hierarchical aggregation network for 3D left atrial segmentation | |
| CN109829880A (en) | A kind of CT image detecting method based on deep learning, device and control equipment | |
| CN109998532A (en) | Electrocardiosignal recognition methods and device based on the more topology convergence networks of multi-lead | |
| CN107292884A (en) | The method and device of oedema and hemotoncus in a kind of identification MRI image | |
| CN108846829A (en) | Diseased region recognition methods and device, computer installation and readable storage medium storing program for executing | |
| CN109920512B (en) | Training method and device for three-dimensional dose distribution network model | |
| CN110188767A (en) | Keratonosus image sequence feature extraction and classifying method and device based on deep neural network | |
| Pontoriero et al. | Automated data quality control in FDOPA brain PET imaging using deep learning | |
| Fan et al. | Verification of the machine delivery parameters of a treatment plan via deep learning | |
| CN109919932A (en) | The recognition methods of target object and device | |
| Zhan et al. | CFNet: A medical image segmentation method using the multi-view attention mechanism and adaptive fusion strategy | |
| Grossiord et al. | Shaping for PET image analysis | |
| WO2021184195A1 (en) | Medical image reconstruction method, and medical image reconstruction network training method and apparatus | |
| CN112101438A (en) | Left and right eye classification method, device, server and storage medium | |
| CN109223001A (en) | Hyperactivity prediction technique, system, storage medium and equipment based on nuclear magnetic resonance |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190524 |
|
| WW01 | Invention patent application withdrawn after publication |