CN110189299A - A kind of cerebrovascular events automatic identifying method and system based on MoileNet - Google Patents
A kind of cerebrovascular events automatic identifying method and system based on MoileNet Download PDFInfo
- Publication number
- CN110189299A CN110189299A CN201910324180.1A CN201910324180A CN110189299A CN 110189299 A CN110189299 A CN 110189299A CN 201910324180 A CN201910324180 A CN 201910324180A CN 110189299 A CN110189299 A CN 110189299A
- Authority
- CN
- China
- Prior art keywords
- image data
- arteria carotis
- layer
- detailed structure
- cerebrovascular events
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- 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
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10132—Ultrasound image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30016—Brain
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30101—Blood vessel; Artery; Vein; Vascular
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Ultra Sonic Daignosis Equipment (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to the crossing domains of computer technology and medical image, disclose a kind of cerebrovascular events automatic identifying method and system based on MoileNet, and wherein method is comprising steps of (1) acquires two-dimentional arteria carotis initial ultrasound image data；(2) divide arteria carotis outer membrane, obtain ROI image data；(3) the deep learning network based on MoileNet is constructed, and is trained；(4) ROI image data are input in trained deep learning network and are tested, prediction result whether the corresponding cerebrovascular events of two-dimentional arteria carotis initial ultrasound image data occur is obtained, thus automatic identification cerebrovascular events.The present invention uses the deep learning network based on MoileNet, automatically extracts ultrasonic carotid images feature using deep learning method, carries out the automatic identification of cerebrovascular events, can effectively solve Manual definition's feature subjectivity is strong and the Redundancy of feature.
Description
Technical field
The invention belongs to the crossing domains of computer technology and medical image, are based on more particularly, to one kind
The cerebrovascular events automatic identifying method and system of MoileNet can be carried out based on MoileNet by carotid ultrasound image
The automatic identification of cerebrovascular events.
Background technique
Cranial vascular disease has become the most disease of Chinese death toll, around the automatic recognition problem of cerebrovascular events
It expands largely and in-depth study.In medicine, the pathologic basis that cerebrovascular events occur is atherosclerosis, is shown as
Interior medial thickening or patch are formed.Plaque rupture will block blood vessel, make cerebrum ischemia anoxic, eventually lead to cerebrovascular events
The generation of (such as cerebral infarction).Arteria carotis connects heart and brain, and position table is shallow, and structure is simple, it has also become observation body artery is athero-
The important window hardened lesion, assess plaque stability.
In the research in the past to ultrasonic carotid images characteristic image, mainly determined by studying the vulnerability of patch
Property patch is described, comprising: ulcer, patch internal haemorrhage, be rich in lipid core, thin fibrous cap etc..In recent years, the method for machine learning
Also it is applied to the field.The method of conventional machines study is mainly by extracting the textural characteristics of patch come quantitative description spot
Then block is classified by classifier.Textural characteristics include: gray level co-occurrence matrixes, grey scale difference matrix, gray scale brigade commander's journey square
Battle array, the analysis of Law's textural characteristics, Fourier spectrum etc..Classifier includes: AdaBoost classifier, K arest neighbors (K-
Nearest Neighbor, KNN), support vector machines (Support Vector Machine, SVM) etc..
Since features described above needs Manual definition, feature type is more, in order to solve feature Redundancy generally require into
Row feature selecting or dimension-reduction treatment, and the application of different classifications device causes result to be also very different.Therefore, one kind is established
The automatic identifying method and system for automatically extracting ultrasonic carotid images feature progress cerebrovascular events are of great significance.
Summary of the invention
Aiming at the above defects or improvement requirements of the prior art, the purpose of the present invention is to provide one kind to be based on MoileNet
Cerebrovascular events automatic identifying method and system, wherein by each in the processing of method overall flow and corresponding system device
The set-up mode of a functional module element improves, and using the deep learning network based on MoileNet, utilizes deep learning
Method automatically extracts ultrasonic carotid images feature, carries out the automatic identification of cerebrovascular events, can effectively solve Manual definition spy
The subjectivity of sign is strong and the Redundancy of feature；Also, this method and corresponding system carry out feature selecting automatically, solve tradition
The problem of artificially carrying out feature selecting is needed in method.
To achieve the above object, according to one aspect of the present invention, a kind of cerebrovascular thing based on MoileNet is provided
Part automatic identifying method, which comprises the following steps:
(1) two-dimentional arteria carotis initial ultrasound image data is acquired；
(2) image-region of part will be surrounded in two-dimentional arteria carotis initial ultrasound image data by arteria carotis outer membrane as sense
Interest region ROI is partitioned into outside arteria carotis from the two-dimentional arteria carotis initial ultrasound image data that the step (1) collects
Film, the ROI image data after obtaining segmentation arteria carotis outer membrane；
(3) the deep learning network based on MoileNet is constructed, and is trained；
Wherein, the deep learning network based on MobileNet has the structure of separation convolution；The deep learning network
It is input with ROI image data, whether cerebrovascular events occur as output；
The training is to actually occur the corresponding training set ROI image data of cerebrovascular events as positive sample, and reality is not
The corresponding training set ROI image data of cerebrovascular events occur as negative sample；
(4) the ROI image data that the step (2) obtains are input to the trained deep learning that step (3) obtains
It is tested in network, obtains the corresponding cerebrovascular events hair of two dimension arteria carotis initial ultrasound image data in the step (1)
Prediction result whether raw, thus automatic identification cerebrovascular events.
As present invention further optimization, in the step (1), the two dimension arteria carotis initial ultrasound image data is
The arteria carotis that one group of two width corresponding with its left and right sides arteria carotis of the patient in arteria carotis containing patch contains patch is horizontal
The two-dimensional ultrasonic image data of section；
Correspondingly, in step (2), it is described obtain segmentation arteria carotis outer membrane after ROI image data specifically refer to obtain
ROI image data after one group corresponding with left and right sides arteria carotis two segmentation arteria carotis outer membranes；
In step (3), the training set ROI image data of the positive sample and the negative sample include multiple groups segmentation neck
The ROI image data containing patch after tunica adventitia of artery, wherein the ROI containing patch after any one group of segmentation arteria carotis outer membrane
Image data is specially the ROI figure containing patch after two segmentations arteria carotis outer membrane corresponding with left and right sides arteria carotis
As data；
In step (4), specifically by the ROI after one group of two segmentation arteria carotis outer membrane obtained in the step (2)
Image data is input in the trained deep learning network simultaneously and is tested.
As present invention further optimization, in the step (3), the deep learning network based on MobileNet
Including 8 layers of detailed structure, wherein the 1st layer of detailed structure and the 8th layer of detailed structure are all convolutional layer；2-7 layers of detailed structure pair
It should be in 3 separation convolutional coding structures；
Preferably, the 1st layer of detailed structure convolutional layer is a shared convolutional layer, and convolution nuclear parameter is used by the training stage
Left and right sides carotid ultrasound image data common iterations obtain；There are two being arranged behind the output of the 1st layer of detailed structure
Operation, one is channel attended operation, and one is shuffle operation, and shuffle operation is specifically with reference in ShuffleNet
The shuffle layer that one channel of shuffle thought setting is 2；Any 1 separation convolution knot in 2-7 layers of detailed structure
Structure is made of 1 depth convolutional layer and 1 point-by-point convolutional layer, wherein the output of the 3rd layer of detailed structure and the 5th layer of detailed structure
Between jump be provided with based on ResNet residual error thought connect, to prevent gradient from disappearing；Also, in the 7th layer of detailed structure
Output after be provided with global pool layer, for carrying out global pool operation；8th layer of detailed structure convolutional layer is an output volume
Lamination represents the two class values that cerebrovascular events whether occur for exporting；
Preferably, the deep learning network based on MobileNet also use weight decaying and
BatchNormalization, and dropout is used after the global pool layer.
As present invention further optimization, in the step (3), for the training, train in solution parameter
Learning rate is preferably arranged to 0.0008.
As present invention further optimization, in the step (1), the arteria carotis described in any one width containing patch is crosscutting
The two-dimensional ultrasonic image data in face are with the perpendicular vertical direction in arteria carotis cross section by ultrasonic probe edge to each arteria carotis
Cross section is scanned, and is taken the cross section two-dimensional ultrasonic image data of wherein carotid plaques area maximum to be used as and is contained patch
Arteria carotis cross section two-dimensional ultrasonic image data；
Preferably, the two-dimentional arteria carotis initial ultrasound image data correspond to total neck, vascular bifurcation inside and outside sinus portion or neck
Place.
As present invention further optimization, in the step (2), the arteria carotis outer membrane that is partitioned into is using semi-automatic
Dividing method specifically comprises the following steps:
(2-1) reads two-dimentional arteria carotis initial ultrasound image data；
The artificial compartment of terrain on arteria carotis epicardial border (2-2) marks at least eight point as initial point；
(2-3) carries out interpolation to initial point using interpolation filter, obtains arteria carotis epicardium contours, includes in the profile
There are middle film and patch in arteria carotis；Cast out the image data of profile exterior domain to get the ROI image after segmentation arteria carotis outer membrane is arrived
Data.
It is another aspect of this invention to provide that the present invention provides a kind of cerebrovascular events automatic identification based on MoileNet
System, which is characterized in that the system includes:
Image capture module, for acquiring two-dimentional arteria carotis initial ultrasound image data；
Image segmentation module, for part will to be surrounded by arteria carotis outer membrane in two-dimentional arteria carotis initial ultrasound image data
It is dynamic to be partitioned into neck as region of interest ROI from the two-dimentional arteria carotis initial ultrasound image data collected for image-region
Arteries and veins outer membrane, the ROI image data after obtaining segmentation arteria carotis outer membrane；
Deep learning network based on MoileNet, for being input with ROI image data, whether cerebrovascular thing occurs
Part is as output；Wherein, the deep learning network based on MobileNet has the structure of separation convolution；This is based on
The deep learning network of MoileNet is particularly used for being tested using ROI image data as input, obtains the two-dimentional neck
Prediction result whether the corresponding cerebrovascular events of artery initial ultrasound image data occur, automatic identification cerebrovascular events.
As present invention further optimization, in described image acquisition module, the two dimension arteria carotis initial ultrasound image
Data are that the neck that one group of two width corresponding with arteria carotis at left and right sides of the patient in arteria carotis containing patch contains patch is dynamic
The two-dimensional ultrasonic image data in arteries and veins cross section；
Correspondingly, in image segmentation module, the ROI image data after obtaining segmentation arteria carotis outer membrane specifically refer to obtain
ROI image data after one group corresponding with left and right sides arteria carotis two segmentation arteria carotis outer membranes；
The deep learning network of the MoileNet is particularly used for after described one group two segmentation arteria carotis outer membranes
ROI image data are tested as input simultaneously.
As present invention further optimization, the deep learning network based on MobileNet, including 8 layers of details knot
Structure, wherein the 1st layer of detailed structure and the 8th layer of detailed structure are all convolutional layer, 2-7 layers of detailed structure correspond to 3 separation volumes
Product structure；
Preferably, the 1st layer of detailed structure convolutional layer is a shared convolutional layer, and convolution nuclear parameter is used by the training stage
Left and right sides carotid ultrasound image data common iterations obtain；There are two behaviour for tool behind the output of the 1st layer of detailed structure
Make, one is channel attended operation, and one is shuffle operation, and shuffle operation is specifically with reference in ShuffleNet
The shuffle layer that one channel of shuffle thought setting is 2；Any 1 separation convolution knot in 2-7 layers of detailed structure
Structure is made of 1 depth convolutional layer and 1 point-by-point convolutional layer, wherein the output of the 3rd layer of detailed structure and the 5th layer of detailed structure
Between jump connection is arranged based on ResNet residual error thought, to prevent gradient from disappearing；Also, in the 7th layer of detailed structure
One global pool layer is set after output, for carrying out global pool operation；8th layer of detailed structure convolutional layer is an output volume
Lamination represents the two class values that cerebrovascular events whether occur for exporting；
Preferably, the deep learning network based on MobileNet also use weight decaying and
BatchNormalization, and dropout is used after the global pool layer.
As present invention further optimization, in described image acquisition module, the neck described in any one width containing patch is dynamic
The two-dimensional ultrasonic image data in arteries and veins cross section are along the vertical direction perpendicular with arteria carotis cross section by ultrasonic probe to each
Arteria carotis cross section is scanned, and is taken the cross section two-dimensional ultrasonic image data of wherein carotid plaques area maximum to be used as and is contained
There are the two-dimensional ultrasonic image data in the arteria carotis cross section of patch；
Preferably, the two-dimentional arteria carotis initial ultrasound image data correspond to total neck, vascular bifurcation inside and outside sinus portion or neck
Place.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, by being based on
The network model automatic identification cerebrovascular events of MoileNet have following characteristic and advantages:
(1) present invention automatically extracts ultrasonic carotid images feature using deep learning network to carry out cerebrovascular events
Automatic identification, can effectively solve Manual definition's feature subjectivity is strong and the Redundancy of feature；Meanwhile the present invention automatically into
Row feature selecting solves the problems, such as to need in conventional method artificially to carry out feature selecting.
(2) it is based on convolution separating thought in MobileNet and preferably sets progress with reference to shuffle thought in ShuffleNet
The design of network structure, training network design is reasonable, and solution parameter setting is reasonable, provides guarantee for accurate training.
Deep learning network of the present invention is that have separation convolution based on being separated designed by convolution thought in MobileNet
Structure.Network in the present invention includes 8 layers of detailed structure, wherein the 1st layer of detailed structure and the 8th layer of detailed structure are all common
Convolutional layer, 2-7 layers of detailed structure correspond to 3 separation convolutional coding structures, and (any 1 separation convolutional coding structure is by 1 depth
Convolutional layer and 1 point-by-point convolutional layer are constituted).The present invention by its detailed structure of deep learning network based on MoileNet into
Row control preferably uses for reference residual error thought and devises jump connection, can prevent gradient from disappearing.Also, preferably make in network
With weight decaying and BatchNormalization, and after global pool layer use dropout, can effectively solve over-fitting and ask
Topic.
In addition, in the training process, learning rate is preferably arranged to 0.0008 by the present invention, good model running can be obtained
As a result.
(3) divide arteria carotis outer membrane, tested with the data for including within outer membrane, it can be with compared to independent segmentation patch
More information are obtained, especially when, there are when multiple patches, solving artificial selection in a carotid ultrasound image
The subjectivity problem of patch.The present invention is input to using ultrasound image data within two-dimentional arteria carotis outer membrane as ROI image data
In deep learning network, for training and testing.
(4) network application of MoileNet deep learning is carried out the cerebrovascular in two-dimentional carotid ultrasound image for the first time by the present invention
Event automatic identification is based on the method, can be modeled with the risk profile of further progress cerebrovascular events, thus to prediction result
The patient of high risk carries out Clinical Alert, formulates therapeutic intervention scheme, prevents the generation of cerebrovascular events.Also, previous
It is typically all to use patient one ultrasonic carotid images characteristic as input, and present invention preferably employs a human body left sides in document
Right two sides ultrasound carotid images characteristic combines available more information, to preferably carry out cerebrovascular events
Automatic identification.
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow charts of the cerebrovascular events automatic identifying method of MoileNet.
Fig. 2 (a) and Fig. 2 (b) is two-dimensional ultrasound carotid images initial data, and wherein Fig. 2 (a) is that neck is dynamic on the left of patient
Arteries and veins cross section patch data, Fig. 2 (b) are patient's right carotid artery cross section patch data.
Fig. 3 (a) and Fig. 3 (b) are ROI image needed for testing, i.e., comprising in interior within outer membrane after segmentation arteria carotis outer membrane
Film and patch data, wherein Fig. 3 (a) is patient's left carotid artery cross section ROI data, and Fig. 3 (b) is patient's right carotid artery
Cross section ROI data.
Fig. 4 is deep learning network structure of the present invention.
Fig. 5 is that deep learning network obtains the ROC curve of classification results.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
Not constituting a conflict with each other can be combined with each other.
Embodiment 1
Cerebrovascular events automatic identifying method provided by the invention, as shown in Figure 1, comprising the following steps:
(1) two-dimentional carotid ultrasound image data is acquired.It is actual two dimension carotid ultrasound image by Ultrasonography doctor into
Row acquisition, the arteria carotis range of acquisition is preferably neck total, sinus portion, inside and outside neck at vascular bifurcation.Doctor is by hanging down ultrasonic probe
It is directly scanned in arteria carotis cross section, takes the patch data of plaque area maximum, it is horizontal to obtain arteria carotis at left and right sides of patient
Section patch image.By taking the present embodiment as an example, the patch image data that the present embodiment finally acquires includes 333 patients, wherein
Occur cerebrovascular events patient 116, there is no cerebrovascular events patient 217.Fig. 2 (a) and Fig. 2 (b) is two-dimensional ultrasound
Carotid images initial data, wherein Fig. 2 (a) is patient's left carotid artery cross section patch data, and Fig. 2 (b) is on the right side of patient
Arteria carotis cross section patch data.
(2) it can be based on dividing method in the prior art, divided from the collected two-dimentional carotid ultrasound image data of institute
Arteria carotis outer membrane is cut out, epicardium contours are obtained.In the present embodiment, being partitioned into arteria carotis outer membrane is semi-automatic point in the prior art
It cuts, the specific method is as follows:
(2-1) reads two-dimentional carotid ultrasound patch image data；
(2-2) selects 8 points as initial point evenly and at intervals in arteria carotis epicardial border；
(2-3) designs interpolation filter: the filter is a balanced-filter, by minimizing interpolation point and ideal point
Between mean square error obtain.The length of filter is 81 points, and cutoff frequency is 0.5.
(2-4) carries out interpolation to initial point using interpolation filter, obtains arteria carotis epicardium contours point, wraps within the profile
0 is set to containing film middle in arteria carotis and patch data, profile with exterior pixel.Fig. 3 (a) and Fig. 3 (b) is after dividing arteria carotis outer membrane
ROI image data, wherein Fig. 3 (a) is patient's left carotid artery cross section ROI data, and Fig. 3 (b) is that patient's right carotid artery is horizontal
Section ROI data.
(3) projected depth learning network.
The deep learning network designs as follows:
(3-1) is using carotid ultrasound image data at left and right sides of patient as the input of network, if cerebrovascular thing occurs
Output of the part as network.Whether cerebrovascular events occur, 0/1 or other Closing Binary Markers can be used, for example, wherein 0 can be with
It represents there is no cerebrovascular events, 1 represents generation cerebrovascular events.
The structure and model parameter of (3-2) projected depth learning network.
The structure of the deep learning network is based on convolution separating thought in MobileNet and with reference in ShuffleNet
The design of shuffle thought.Network include a shared convolutional layer, one shuffle layers, three separation convolutional layers, an overall situations
Pond layer and output convolutional layer.In separation convolutional layer, uses for reference ResNet residual error thought and done a jump connection, to prevent gradient
It disappears.In order to solve overfitting problem, weight decaying and BatchNormalization are used in network, and in global pool
Dropout is used after layer.
Model parameter setting are as follows: input picture size is 112*112, and gray value is normalized between 0-1；Shared convolutional layer
Convolution kernel having a size of 3*3*16, step-length 1；Shuffle layers of channel is set as 2；First separation convolutional layer channel setting
It is 64, step-length 2, second separation convolutional layer channel is set as 64, and step-length 1, third separation convolutional layer channel is set as
128, step-length 2.All using relu6 as excitation layer behind all convolutional layers, and carry out BatchNormalization.Figure
4 be deep learning network structure.
As shown in figure 4, deep learning network is based on designed by separation convolution thought in MobileNet.Network includes 8
Layer detailed structure, wherein the 1st layer of detailed structure and the 8th layer of detailed structure are all convolutional layer, 2-7 layers of detailed structure correspond to 3
A separation convolutional coding structure.The 1st layer of detailed structure convolutional layer is a shared convolutional layer, and convolution nuclear parameter can be by the training stage
Its left and right sides carotid ultrasound image data common iterations of the patient containing patch obtain in used arteria carotis；The 1st
The output of layer detailed structure is operated followed by two, and one is channel attended operation, and one is shuffle operation,
Shuffle operation is the shuffle layer that a channel is 2 with reference to the shuffle thought in ShuffleNet；Described 2-7 layers
Any 1 separation convolutional coding structure is made of 1 depth convolutional layer and 1 point-by-point convolutional layer in detailed structure, wherein the 3rd layer and
5 layers of output uses for reference residual error thought and has done a jump connection, and to prevent gradient from disappearing, the 7th layer of output is followed by the overall situation
Pondization operation；8th layer of detailed structure convolutional layer is the output convolutional layer that an output is 2, represents whether cerebrovascular events occur
Two classes.
(4) ultrasound image data within the resulting two-dimentional arteria carotis outer membrane of step (2) is input to the depth of step (3) design
Degree learning network is trained and tests, and obtains network model automatic identification cerebrovascular events.The instruction of the deep learning network
Practice and test, the specific steps are as follows:
Patient is divided into two groups according to whether patient occurs cerebrovascular events by (4-1), and event group occurs and is recorded as 1 group, does not have
There is generation event group to be recorded as 0 group；
(4-2) carries out 0 group of data down-sampled: random grouping obtains comparable two subset datas of number.Pass through sample
It is down-sampled, imbalanced training sets can be solved the problems, such as, so that finally obtained prediction result has more sensibility and robustness.
Two subset datas in (4-2) respectively with 1 group of data mixing, are obtained two experiment sample data by (4-3).
(4-4) carries out the grouping of five folding cross validations to two experiment sample data respectively, and training data is inputted network
In, training solution parameter and initialization network parameter are set, and iteration updates weight, obtains the network model and carries out brain blood
It runs affairs the automatic identification of part.
Loss value selects two-value cross entropy in the trained solution parameter, and learning rate is preferably arranged to 0.0008, prison
The loss value of test card collection, when the loss value of verifying collection no longer changes, 20 iterative learning rates of every progress fall to original
0.1.Weight decaying punishment weight_decay is set as 0.01.Trained batch_size is set as 15, the batch_size of test
It is set as 32.In order to keep result more accurate, every time when training, training set data, random initializtion network parameter will be upset.
The deep learning network is tested, evaluates the correctness and accuracy of cerebrovascular events automatic identifying method, specifically
It is as follows:
A loss value network structure as small as possible is chosen, performance test data are tested.Test data is inputted
The true tag value of network, obtained prediction label value and test data compares, and calculates and predicts correct data amount check, from
And the accuracy of every group of test data is obtained, finally calculate average accuracy (ACC).Meanwhile saving the label value of prediction,
ROC curve is drawn together with true tag value, and calculates area AUC under ROC curve, and the value of AUC is bigger to illustrate classifying quality more
Good, output probability is more reasonable, and the range of AUC is between 0-1.
When test data, due to the limitation of number of patients, five five folding cross validations can be used, i.e., it is data are random
It is divided into five foldings, uses wherein that a broken number is according to as test set every time, remaining four fold data is as training set.Number is tested for every folding
According to, it carries out five repetition trainings and tests, result of the obtained average result as the folding test set.It can guarantee to train in this way
The independence of sample and test sample, so that classifying quality has more robustness.
Table 1 is first group of five folding cross validation results of arteria carotis cross section ROI image data, and table 2 is arteria carotis cross section
Second group of five folding cross validation results of ROI image data, table 3 are the average value of two groups of results.
1 first group of five folding cross validation results of table
2 second group of five folding cross validation results of table
The average value of 3 two groups of results of table
Accuracy range is between 82%-89% when five folding cross validation of test set progress known to table 1, table 2, AUC model
It encloses between 88%-95%.This is because grouping is grouped at random, every group of training set and test set quantity be not identical, and
And data have a certain difference.Simultaneously, it can be seen that the result of accuracy and AUC are with uniformity, and there are positively related passes
System, matches with actual conditions, illustrates that result has certain correctness and reliability.
The value that the final average result accuracy of test set is 85.2% or so, AUC as shown in Table 3 is 91.5% or so,
Illustrate that the method has preferable classification results.
Fig. 5 is the wherein value of the ROC curve and corresponding A UC of one group of test data in group result.It can be seen that the value of AUC
It is 91.29%, illustrates that the method has high sensitivity and specificity, have good robustness.
Whether midbrain event of the present invention occurs to acquire patient's two dimension arteria carotis initial ultrasound in arteria carotis containing patch
Image data is judgement instantly.As ultrasound image used by the embodiment of the present invention is the image checked when patient is hospitalized, it is
Cerebrovascular events whether occur when being hospitalized using patient as standard, that is to say, that ultrasound image is that moment in hospital, it is general come
Say, if this patient in hospital when there is no an apoplexy, below if apoplexy, time interval takes 1 year even more long at least
Time, ultrasound image at that time may change again, need to resurvey ultrasound image data.Brain event in the present invention
Generally just refer to cerebral infarction, including acute cerebral infarction, chamber stalk etc.；Although there is no cerebrovascular events for some cases, go out
It, is still classified as that there is no in cerebrovascular events one kind by the symptoms such as existing glossolalia, dark and dim eyesight.
The unspecified place of the present invention can refer to related art, for example, it separates convolution about MobileNet
Thought can refer to the prior art, such as Howard A G, Zhu M, Chen B, et al.Mobilenets:Efficient
convolutional neural networks for mobile vision applications[J].arXiv
preprint arXiv:1704.04861,2017；Shuffle thought is see also the prior art, such as Zhang X, Zhou X,
Lin M,et al.Shufflenet:An extremely efficient convolutional neural network
for mobile devices[C]//Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition.2018:6848-6856；Weight decaying and BatchNormalization can be referred to respectively
Loshchilov I,Hutter F.Fixing weight decay regularization in adam[J].arXiv
Preprint arXiv:1711.05101,2017. and Ioffe S, Szegedy C.Batch normalization:
Accelerating deep network training by reducing internal covariate shift[J]
.arXiv preprint arXiv:1502.03167,2015.；Dropout can refer to Krizhevsky A, Sutskever I,
Hinton G E.Imagenet classification with deep convolutional neural networks
[C] //Advances in neural information processing systems.2012:1097-1105. etc..
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (10)
1. a kind of cerebrovascular events automatic identifying method based on MoileNet, which comprises the following steps:
(1) two-dimentional arteria carotis initial ultrasound image data is acquired；
(2) image-region of part is surrounded as interested by arteria carotis outer membrane using in two-dimentional arteria carotis initial ultrasound image data
Region ROI is partitioned into arteria carotis outer membrane from the two-dimentional arteria carotis initial ultrasound image data that the step (1) collects,
ROI image data after obtaining segmentation arteria carotis outer membrane；
(3) the deep learning network based on MoileNet is constructed, and is trained；
Wherein, the deep learning network based on MobileNet has the structure of separation convolution；The deep learning network with
ROI image data are input, whether cerebrovascular events occur as output；
The training is practical not occur to actually occur the corresponding training set ROI image data of cerebrovascular events as positive sample
The corresponding training set ROI image data of cerebrovascular events are as negative sample；
(4) the ROI image data that the step (2) obtains are input to the trained deep learning network that step (3) obtains
In tested, obtain in the step (1) the corresponding cerebrovascular events of two-dimentional arteria carotis initial ultrasound image data occur with
No prediction result, thus automatic identification cerebrovascular events.
2. the cerebrovascular events automatic identifying method based on MoileNet as described in claim 1, which is characterized in that the step
(1) in, the two dimension arteria carotis initial ultrasound image data is that its left and right sides neck is dynamic with the patient in arteria carotis containing patch
Arteries and veins respectively corresponding one group of two width contain patch arteria carotis cross section two-dimensional ultrasonic image data；
Correspondingly, in step (2), the ROI image data obtained after segmentation arteria carotis outer membrane specifically refer to obtain with it is left
ROI image data after the corresponding one group of two segmentation arteria carotis outer membrane of two bilateral common carotid arteries of the right side；
In step (3), the training set ROI image data of the positive sample and the negative sample include multiple groups segmentation arteria carotis
The ROI image data containing patch after outer membrane, wherein the ROI image containing patch after any one group of segmentation arteria carotis outer membrane
Data are specially the ROI image number containing patch after two segmentations arteria carotis outer membrane corresponding with left and right sides arteria carotis
According to；
In step (4), specifically by the ROI image after one group of two segmentation arteria carotis outer membrane obtained in the step (2)
Data are input in the trained deep learning network simultaneously and are tested.
3. the cerebrovascular events automatic identifying method based on MoileNet as described in claim 1, which is characterized in that the step
(3) in, the deep learning network based on MobileNet includes 8 layers of detailed structure, wherein the 1st layer of detailed structure and the 8th layer
Detailed structure is all convolutional layer, and 2-7 layers of detailed structure correspond to 3 separation convolutional coding structures；
Preferably, the 1st layer of detailed structure convolutional layer is a shared convolutional layer, convolution nuclear parameter left side as used by the training stage
Right two sides carotid ultrasound image data common iterations obtain；The operation there are two settings behind the output of the 1st layer of detailed structure,
One is channel attended operation, and one is shuffle operation, and shuffle operation is specifically with reference in ShuffleNet
The shuffle layer that one channel of shuffle thought setting is 2；Any 1 separation convolution knot in 2-7 layers of detailed structure
Structure is made of 1 depth convolutional layer and 1 point-by-point convolutional layer, wherein the output of the 3rd layer of detailed structure and the 5th layer of detailed structure
Between jump connection is arranged based on ResNet residual error thought, to prevent gradient from disappearing；Also, in the 7th layer of detailed structure
One global pool layer is set after output, for carrying out global pool operation；8th layer of detailed structure convolutional layer is an output volume
Lamination represents the two class values that cerebrovascular events whether occur for exporting；
Preferably, the deep learning network based on MobileNet also use weight decaying and
BatchNormalization, and dropout is used after the global pool layer.
4. the cerebrovascular events automatic identifying method based on MoileNet as described in claim 1, which is characterized in that the step
(3) in, for the training, learning rate is preferably arranged to 0.0008 in training solution parameter.
5. the cerebrovascular events automatic identifying method based on MoileNet as claimed in claim 2, which is characterized in that the step
(1) in, the two-dimensional ultrasonic image data in the arteria carotis cross section described in any one width containing patch are by ultrasonic probe edge and neck
The perpendicular vertical direction in artery cross section is scanned each arteria carotis cross section, takes wherein carotid plaques area maximum
Two-dimensional ultrasonic image data of the cross section two-dimensional ultrasonic image data at place as the arteria carotis cross section containing patch；
Preferably, the two-dimentional arteria carotis initial ultrasound image data corresponds to total neck, inside and outside sinus portion or neck at vascular bifurcation.
6. the cerebrovascular events automatic identifying method based on MoileNet as described in claim 1, which is characterized in that the step
(2) in, the arteria carotis outer membrane that is partitioned into is specifically comprised the following steps: using semi-automatic partition method
(2-1) reads two-dimentional arteria carotis initial ultrasound image data；
The artificial compartment of terrain on arteria carotis epicardial border (2-2) marks at least eight point as initial point；
(2-3) carries out interpolation to initial point using interpolation filter, obtains arteria carotis epicardium contours, includes neck in the profile
Film and patch in intra-arterial；Cast out the image data of profile exterior domain to get the ROI image number after segmentation arteria carotis outer membrane is arrived
According to.
7. a kind of cerebrovascular events automatic recognition system based on MoileNet, which is characterized in that the system includes:
Image capture module, for acquiring two-dimentional arteria carotis initial ultrasound image data；
Image segmentation module, for the image of part will to be surrounded in two-dimentional arteria carotis initial ultrasound image data by arteria carotis outer membrane
Region is partitioned into outside arteria carotis from the two-dimentional arteria carotis initial ultrasound image data collected as region of interest ROI
Film, the ROI image data after obtaining segmentation arteria carotis outer membrane；
Deep learning network based on MoileNet, for being input with ROI image data, whether cerebrovascular events work occurs
For output；Wherein, the deep learning network based on MobileNet has the structure of separation convolution；It should be based on MoileNet
Deep learning network, be particularly used for being tested using ROI image data as input, it is initial to obtain the two-dimentional arteria carotis
Prediction result whether the corresponding cerebrovascular events of ultrasound image data occur, automatic identification cerebrovascular events.
8. the cerebrovascular events automatic recognition system based on MoileNet as claimed in claim 7, which is characterized in that described image
In acquisition module, the two dimension arteria carotis initial ultrasound image data is and neck at left and right sides of the patient in arteria carotis containing patch
Corresponding one group of two width of artery contains the two-dimensional ultrasonic image data in the arteria carotis cross section of patch；
Correspondingly, in described image segmentation module, the ROI image data after obtaining segmentation arteria carotis outer membrane specifically refer to obtain
ROI image data after one group corresponding with left and right sides arteria carotis two segmentation arteria carotis outer membranes；
The deep learning network of the MoileNet is particularly used for the ROI after described one group two segmentation arteria carotis outer membranes
Image data is tested as input simultaneously.
9. the cerebrovascular events automatic recognition system based on MoileNet as claimed in claim 7, which is characterized in that described to be based on
The deep learning network of MobileNet, including 8 layers of detailed structure, wherein the 1st layer of detailed structure and the 8th layer of detailed structure are all
Convolutional layer, 2-7 layers of detailed structure correspond to 3 separation convolutional coding structures；
Preferably, the 1st layer of detailed structure convolutional layer is a shared convolutional layer, convolution nuclear parameter left side as used by the training stage
Right two sides carotid ultrasound image data common iterations obtain；The operation there are two tools behind the output of the 1st layer of detailed structure, one
A is channel attended operation, and one is shuffle operation, and shuffle operation is specifically with reference to the shuffle in ShuffleNet
The shuffle layer that one channel of thought setting is 2；Any 1 separation convolutional coding structure in 2-7 layers of detailed structure is by 1
A depth convolutional layer and 1 point-by-point convolutional layer are constituted, wherein base between the 3rd layer of detailed structure and the output of the 5th layer of detailed structure
ResNet residual error thought is provided with a jump connection, to prevent gradient from disappearing；Also, in the output of the 7th layer of detailed structure
After be provided with global pool layer, for carrying out global pool operation；8th layer of detailed structure convolutional layer is an output convolutional layer,
The two class values that cerebrovascular events whether occur are represented for exporting；
Preferably, the deep learning network based on MobileNet also use weight decaying and
BatchNormalization, and dropout is used after the global pool layer.
10. the cerebrovascular events automatic recognition system based on MoileNet as claimed in claim 8, which is characterized in that the figure
As in acquisition module, the two-dimensional ultrasonic image data in the arteria carotis cross section described in any one width containing patch are by ultrasonic probe
Along being scanned with the perpendicular vertical direction in arteria carotis cross section to each arteria carotis cross section, wherein carotid plaques face is taken
Two-dimensional ultrasonic image data of the cross section two-dimensional ultrasonic image data of product maximum as the arteria carotis cross section containing patch；
Preferably, the two-dimentional arteria carotis initial ultrasound image data corresponds to total neck, inside and outside sinus portion or neck at vascular bifurcation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910324180.1A CN110189299B (en) | 2019-04-22 | 2019-04-22 | Cerebrovascular event automatic identification method and system based on MobileNet |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910324180.1A CN110189299B (en) | 2019-04-22 | 2019-04-22 | Cerebrovascular event automatic identification method and system based on MobileNet |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110189299A true CN110189299A (en) | 2019-08-30 |
| CN110189299B CN110189299B (en) | 2020-12-08 |
Family
ID=67714849
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910324180.1A Active CN110189299B (en) | 2019-04-22 | 2019-04-22 | Cerebrovascular event automatic identification method and system based on MobileNet |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110189299B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111724365A (en) * | 2020-06-16 | 2020-09-29 | 中国科学院自动化研究所 | Interventional instrument detection method, system and device for endovascular aneurysm repair operation |
| CN113838028A (en) * | 2021-09-24 | 2021-12-24 | 无锡祥生医疗科技股份有限公司 | Carotid artery ultrasonic automatic Doppler method, ultrasonic equipment and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102800088A (en) * | 2012-06-28 | 2012-11-28 | 华中科技大学 | Automatic dividing method of ultrasound carotid artery plaque |
| CN103996194A (en) * | 2014-05-23 | 2014-08-20 | 华中科技大学 | Automatic intima-media membrane partitioning method based on ultrasound carotid artery image |
| CN108388833A (en) * | 2018-01-15 | 2018-08-10 | 阿里巴巴集团控股有限公司 | A kind of image-recognizing method, device and equipment |
| CN109259784A (en) * | 2018-08-27 | 2019-01-25 | 上海铱硙医疗科技有限公司 | AI prediction technique, device, equipment and the storage medium of cerebral infarction |
-
2019
- 2019-04-22 CN CN201910324180.1A patent/CN110189299B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102800088A (en) * | 2012-06-28 | 2012-11-28 | 华中科技大学 | Automatic dividing method of ultrasound carotid artery plaque |
| CN103996194A (en) * | 2014-05-23 | 2014-08-20 | 华中科技大学 | Automatic intima-media membrane partitioning method based on ultrasound carotid artery image |
| CN108388833A (en) * | 2018-01-15 | 2018-08-10 | 阿里巴巴集团控股有限公司 | A kind of image-recognizing method, device and equipment |
| CN109259784A (en) * | 2018-08-27 | 2019-01-25 | 上海铱硙医疗科技有限公司 | AI prediction technique, device, equipment and the storage medium of cerebral infarction |
Non-Patent Citations (4)
| Title |
|---|
| HOWARD A G: "MobileNets: efficient convolutional networks for mobile vision applications", 《ARXIV PREPRINT ARXIV:1704.04861》 * |
| 孙夏: "基于卷积神经网络的颈动脉斑块超声图像特征识别", 《中国医疗器械信息》 * |
| 杨星: "基于深度置信网络的脑血管病风险预警研究", 《中国优秀硕士学位论文全文数据库 医药卫生科技辑》 * |
| 赵媛: "一种基于深度学习的颈动脉斑块超声图像识别方法", 《中国医疗器械信息》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111724365A (en) * | 2020-06-16 | 2020-09-29 | 中国科学院自动化研究所 | Interventional instrument detection method, system and device for endovascular aneurysm repair operation |
| CN111724365B (en) * | 2020-06-16 | 2021-11-09 | 中国科学院自动化研究所 | Interventional instrument detection method, system and device for endovascular aneurysm repair operation |
| CN113838028A (en) * | 2021-09-24 | 2021-12-24 | 无锡祥生医疗科技股份有限公司 | Carotid artery ultrasonic automatic Doppler method, ultrasonic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110189299B (en) | 2020-12-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109886273B (en) | CMR image segmentation and classification system | |
| Tian et al. | Multi-path convolutional neural network in fundus segmentation of blood vessels | |
| CN114926477B (en) | Brain tumor multi-mode MRI image segmentation method based on deep learning | |
| Joshi et al. | Glaucoma detection using image processing and supervised learning for classification | |
| CN111667490B (en) | Fundus picture cup optic disc segmentation method | |
| Khandouzi et al. | Retinal vessel segmentation, a review of classic and deep methods | |
| WO2021212715A1 (en) | Schizophrenia classification and identification method, operation control apparatus, and medical equipment | |
| Abbasi-Sureshjani et al. | Boosted exudate segmentation in retinal images using residual nets | |
| Xu et al. | Retinal vessel segmentation via a semantics and multi-scale aggregation network | |
| CN115272295A (en) | Dynamic brain function network analysis method and system based on time domain-space domain combined state | |
| CN110189299A (en) | A kind of cerebrovascular events automatic identifying method and system based on MoileNet | |
| CN115337000A (en) | Machine learning method for evaluating brain aging caused by diseases based on brain structure images | |
| Manan et al. | A residual encoder-decoder network for segmentation of retinal image-based exudates in diabetic retinopathy screening | |
| Alam et al. | Benchmarking deep learning frameworks for automated diagnosis of OCULAR TOXOPLASMOSIS: A comprehensive approach to classification and segmentation | |
| CN112861881A (en) | Honeycomb lung recognition method based on improved MobileNet model | |
| Sangamesh et al. | A New Approach to Recognize a Patient with Diabetic Retinopathy using Pre-trained Deep Neural Network EfficientNetB0 | |
| Zhou | [Retracted] Design of Intelligent Diagnosis and Treatment System for Ophthalmic Diseases Based on Deep Neural Network Model | |
| CN113378794A (en) | Information correlation method for elephant and symptom information | |
| Meshram et al. | MCBM: Implementation Of Multiclass And Transfer Learning Algorithm Based On Deep Learning Model For Early Detection Of Diabetic Retinopathy | |
| Anand et al. | Transformer-based Multiclass Classification of Cervicogram Images for Improved Cervical Cancer Detection | |
| Celik et al. | Automated Retinal Image Analysis to Detect Optic Nerve Hypoplasia | |
| Ramaraj et al. | Identification of Blood Vessel Clot Region using Fuzzy C Means Clustering Based Artificial Bee Colony Algorithm | |
| Manjushree et al. | Automated detection of diabetic retinopathy using deep learning in retinal fundus images: analysis | |
| Khan et al. | Deep Learning Unveiled: Investigating Retina Eye Segmentation for Glaucoma Diagnosis | |
| Zehra et al. | Dr-net: Cnn model to automate diabetic retinopathy stage diagnosis |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |