CN107292257A - Body part automatic identification magnetic resonance scanning method and device based on deep learning - Google Patents
Body part automatic identification magnetic resonance scanning method and device based on deep learning Download PDFInfo
- Publication number
- CN107292257A CN107292257A CN201710447366.7A CN201710447366A CN107292257A CN 107292257 A CN107292257 A CN 107292257A CN 201710447366 A CN201710447366 A CN 201710447366A CN 107292257 A CN107292257 A CN 107292257A
- Authority
- CN
- China
- Prior art keywords
- magnetic resonance
- image data
- data
- resonance image
- body part
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The present invention relates to the body part automatic identification magnetic resonance scanning method based on deep learning and device, belong to mr imaging technique field.It the described method comprises the following steps：Initial network model is trained using the magnetic resonance image data collection of the body different parts under line, the optimal network model of one automatic identification body part of acquistion is used as grader；Using the grader of acquistion, the data to magnetic resonance imaging body part on line are identified, and obtain classification results；And magnetic resonance imaging sequence is set according to the position recognized automatically, according to the classification results of acquisition, magnetic resonance imaging sequence is called automatically, completes to treat the scanning process for sweeping position.Present invention eliminates complicated positioning and parameter setting time, reduce because of multiple scanning caused by operational error, improve operating efficiency, with higher practical value.
Description
Technical field
The present invention relates to mr imaging technique field, more particularly to the body part automatic identification magnetic based on deep learning
Resonance scan method and apparatus.
Background technology
During magnetic resonance imaging, Scan orientation is a basic operation, is precisely positioned at actual clinical with consistent
There is very important effect in research.Our parts of body are complicated, and clinical diagnosis needs to check substantial amounts of magnetic resonance
View data, workload is big, thus the accurately identification scanning of body different parts, and both having contributed to assisted diagnosis, person quickly recognizes
Scanning area is positioned, the anatomical structure of parts of body is much more familiar with, accumulates experience, while magnetic resonance figure can also be effectively improved
As the utilization rate of data.
At present, artificial intelligence is fast-developing, and it is to work as to solve the problems such as image recognition, speech recognition using deep learning technology
A preceding focus, and achieve good achievement in natural image identification.Such as convolutional neural networks (Convolution
Neural Networks, CNN) successful application, CNN is a kind of deep neural network with convolutional coding structure, and it is generally comprised
Convolutional layer, pond layer, active coating, full articulamentum.Afterwards, with the continuous lifting of algorithm, the number of hardware capabilities lifting and magnanimity
According to deep learning not only substantially increases accuracy applied to image recognition, and improves online computational efficiency.
However, the identification of magnetic resonance imaging body part is recognized by human eye at present, and magnanimity is not under line
It is not fully utilized with body part data, cumbersome in addition, fault rate is high.Applied on 04 28th, 2016
, the Chinese patent CN201610278139.1 of entitled " a kind of magnetic resonance data acquisition method and magnetic resonance imaging system ",
Although there is some to improve in some respects, the defect of needs can not be met in terms of efficiency, rate of precision by also existing.Cause
And, it is necessary to develop a kind of new technology to overcome these defects.
The content of the invention
It is applied successfully based on deep learning on natural image, for above-mentioned deficiency of the prior art, the present invention is proposed
Body part automatic identification magnetic resonance scanning method and device based on deep learning.
Specifically, the present invention proposes a kind of body part automatic identification magnetic resonance imaging side based on deep learning
Method, it comprises the following steps：S1) initial network mould is trained using the magnetic resonance image data collection of the body different parts under line
Type, the optimal network model of acquistion automatic identification body part is used as grader；S2) using in S1) in acquistion grader, it is right
The data of magnetic resonance imaging body part are identified on line, obtain classification results；And S3) automatic according to the position recognized
Magnetic resonance imaging sequence is set, according in S2) the middle classification results obtained, call magnetic resonance imaging sequence, complete to treat automatically
Sweep the scanning process at position.
In a specific embodiment of the method described in the present invention, the step S1) comprise the following steps：S11) receive
Collect the magnetic resonance image data collection of body different parts, the partial data that the magnetic resonance image data being collected into is concentrated enters rower
It is denoted as tally set, using the partial data as checksum set, other data that magnetic resonance image data is concentrated are used as training
Collection；S12) magnetic resonance image data is normalized；S13) create include successively connect input layer, several
The initial network model of hidden layer and output layer, the data inputted in the input layer in the normalised training set will
The result that forward-propagating is exported on network model carries out error measure with the data in the tally set, according to error measure knot
Fruit regulating networks model, determines optimal network model as grader.
In another specific embodiment of the method described in the present invention, in step S11) in, the tally set includes magnetic
The data of all body different parts in resonance image data set, body different parts, institute are marked with the label in tally set
State the data bulk that the data bulk in checksum set is less than in the training set.
In another specific embodiment of the method described in the present invention, by the M for being related to a body part being collected into
M in individual magnetic resonance image dataLIndividual part magnetic resonance image data is marked using as tally set, wherein, the part
The quantity M of magnetic resonance image dataLLess than the 10% of the magnetic resonance image data total quantity M, or less than total quantity M's
5%, or less than the 1% of total quantity M.In certain embodiments, by the N number of magnetic resonance for being related to another body part being collected into
N in view dataLIndividual part magnetic resonance image data is marked using as tally set, wherein, it is related to another body
The quantity N of the part magnetic resonance image data at positionLAccount for magnetic resonance image data total quantity N ratio equal to or different from that
It is related to the quantity M of the part magnetic resonance image data of one body partLAccount for magnetic resonance image data total quantity M's
Ratio.In another specific embodiment of the method described in the present invention, in step S13) in, it is logical to determine optimal network model
Cross using backpropagation and gradient descent method regulation whole network model hyper parameter to obtain final hyper parameter, using it is described most
Whole hyper parameter sets up optimal network model as grader.
In another specific embodiment of the method described in the present invention, in the step S2) in, utilizing grader pair
Before the data of magnetic resonance imaging body different parts are identified on line, data are normalized.In some implementations
In example, the normalized is that image resolution ratio, visual field sizes interested and image angle are unified.
Present invention also offers a kind of computer-readable medium, the computer-readable medium has the journey being stored therein
Sequence, the program is that computer is executable so that computer performs the processing of abovementioned steps.
The invention also provides a kind of body part automatic identification magnetic resonance imaging equipment based on deep learning, it is wrapped
Include：Classifier modules are trained under line, it trains original net using the magnetic resonance image data collection of the body different parts under line
Network model, the optimal network model of acquistion automatic identification body part is used as grader；Line upper bit automatic identification module, its profit
The grader of acquistion in classifier modules is trained under the line, the data to magnetic resonance imaging body part on line are known
Not, classification results are obtained；And automatic setting scanning sequence module, it sets magnetic resonance imaging automatically according to the position recognized
Sequence, according to the classification results obtained in the line upper bit automatic identification module, calls magnetic resonance imaging sequence automatically, complete
Into treating the scanning process of sweeping position.
In some specific embodiments of the described device of the present invention, classifier modules are trained also to include under the line：
Data preparation module, it is configured to the magnetic resonance image data collection of collection of bodily different parts, by the MRI being collected into
Partial data in data set is marked as tally set, using the partial data as checksum set, by MRI number
Training set is used as according to other data of concentration；Data preprocessing module, it is configured to return the magnetic resonance image data
One change is handled；Supervised learning module, it is configured to it and is configured to create input layer, several hidden layers for including successively connecting
With the initial network model of output layer, the data inputted in the input layer in the normalised training set will be in network
The result that forward-propagating is exported on model carries out error measure with the data in the tally set, is adjusted according to error measurement
Network model, determines optimal network model as grader.
In another specific embodiment of the described device of the present invention, the tally set in the data preparation module
The data for all body different parts concentrated including magnetic resonance image data, body difference portion is marked with the label in tally set
Data bulk in position, the checksum set is less than the data bulk in the training set.
In some specific embodiments of the described device of the present invention, by the M for being related to a body part being collected into
M in individual magnetic resonance image dataLIndividual part magnetic resonance image data is marked using as tally set, wherein, the part
The quantity M of magnetic resonance image dataLLess than the 10% of the magnetic resonance image data total quantity M, or less than total quantity M's
5%, or less than the 1% of total quantity M.In certain embodiments, by the N number of magnetic resonance for being related to another body part being collected into
N in view dataLIndividual part magnetic resonance image data is marked using as tally set, wherein, it is related to another body
The quantity N of the part magnetic resonance image data at positionLAccount for magnetic resonance image data total quantity N ratio equal to or different from that
It is related to the quantity M of the part magnetic resonance image data of one body partLAccount for magnetic resonance image data total quantity M's
Ratio.
In another specific embodiment of the described device of the present invention, in the supervised learning module, it is determined that most
Excellent network model is final super to obtain by using the hyper parameter of backpropagation and gradient descent method regulation whole network model
Parameter, sets up optimal network model by the use of the final hyper parameter and is used as grader.
In another specific embodiment of the described device of the present invention, the line upper bit automatic identification module includes：
Data preprocessing module is right before the data of magnetic resonance imaging body different parts on line are identified using grader
Data are normalized.In certain embodiments, the normalized is to image resolution ratio, visual field sizes interested
Unified with image angle.
The body part automatic identification magnetic resonance scanning method based on deep learning of the present invention can be fully sharp with device
With the data learning classification network of magnanimity difference body part under line, so that different body parts on automatic identification line, and improve
The accuracy of different body part identifications, it makes scanning eliminate the reliance on operator's naked eyes identification body part, so as to realize automatic
Change scanning, save complicated positioning and parameter setting time, reduce because of multiple scanning caused by operational error, improve work effect
Rate.
Brief description of the drawings
Fig. 1 is the flow chart of the body part automatic identification magnetic resonance scanning method based on deep learning of the present invention；
Fig. 2 is an exemplary process diagram of the step S1 in Fig. 1.
Embodiment
The embodiment to the present invention is illustrated below in conjunction with the accompanying drawings.It is described below of the invention specific
In embodiment, some very specific technical characteristics are described for better understanding of the present invention, but it will be apparent that
For those skilled in the art, these not all technical characteristics are all to realize that the necessary technology of the present invention is special
Levy.Certain specific embodiments of the invention described below are some exemplary embodiments of the present invention,
It is not construed as limitation of the present invention.
The body part automatic identification magnetic resonance scanning method based on deep learning of the present invention is to utilize data under line
Classification and Identification network is practised to improve the accuracy of identification and shorten recognition time.In one embodiment, as shown in Figure 1 there is provided
A kind of body part automatic identification magnetic resonance scanning method based on deep learning, it is specifically included：
Step S1：Initial network model is trained using the magnetic resonance image data collection of the body different parts under line, is practised
The optimal network model for obtaining automatic identification body part is used as grader.
A step S1 exemplary process diagram is shown in Fig. 2, it comprises the following steps：
(1) data prepare
In this step, the magnetic resonance image data of collection of bodily different parts, based on the MRI number being collected into
According to coming establishing label collection, checksum set and training set.
To make tried to achieve network model more accurate, the magnetic resonance figure for being related to body different parts is preferably collected in large quantities
As data, for example, the view data can be related to including but not limited to cranium brain, ear,nose ＆ throat, neck, cardiopulmonary, liver and spleen, courage kidney,
The image of the concrete positions such as extremities joint, backbone, peripheral vascular.It is further preferred that for a specific portion in body part
Position, such as neck, collected magnetic resonance image data can include the substantial amounts of different scanning angle for being related to the concrete position
The view data of gained.
Using the magnetic resonance image data being collected into come establishing label collection, checksum set and training set.For example, in an implementation
In example, the partial data that can concentrate the magnetic resonance image data being collected into is marked, and is used as tally set, the tally set bag
The data that magnetic resonance image data concentrates all body different parts are included, and body different parts, the part are marked with label
Data are used as training set as checksum set, remaining data.
Preferably, the quantity of the partial data (checksum set) is far smaller than the quantity of remaining data (training set).Example
Such as, in certain embodiments, the quantity of the partial data can be the 10% of collected magnetic resonance image data total amount
Below or less than 5% or less than 1%.
More specifically,, can be to the magnetic resonance image data collection that is collected into by taking neck as an example for a certain concrete position
In M neck magnetic resonance image data in MLIndividual part neck magnetic resonance image data is marked using as tally set,
Remaining MRIndividual neck magnetic resonance image data is used as training set.
In certain embodiments, the number for being selected as the part neck magnetic resonance image data of tally set can be less than all
Neck magnetic resonance image data total amount 10% or less than the neck magnetic resonance image data total amount 5% or less than the neck
The 1% of magnetic resonance image data total amount.
, can be by M neck MRI number being collected into by taking neck and lung as an example for different concrete positions
M inLIndividual part neck magnetic resonance image data is marked using as tally set, and the N number of lung's magnetic being collected into is total to
N in view data of shakingLIndividual part lung magnetic resonance image data is marked to be used as tally set.
Wherein, the number for being selected as the part neck magnetic resonance image data of tally set accounts for all neck MRI numbers
According to the ratio (M of total amountL/ M) and be selected as tally set partial head magnetic resonance image data number account for all head magnetic be total to
Ratio (the N of view data of shaking total amountL/ N) can be with identical, can also be different.
(2) data prediction
Before network training is carried out, the magnetic resonance image data being collected into is normalized, to improve network
The efficiency of training.
The basic normalization thought of view data is to find one group of parameter using the not bending moment of image to eliminate it
The influence that his transforming function transformation function is converted to image.Unique canonical form is namely converted into resist affine transformation.Image normalizing
Change allows image to resist the attack of geometric transformation, can find out those invariants in image, be conducive to follow-up network
Training.
In one embodiment, normalization processing method can carry out resolution ratio, the visual field interested using for example to image
The unification of size and image angle.
For example, for the characteristic block image of different resolution, same image in different resolution can be normalized to, i.e., through normalization
Characteristic block image afterwards will have equal resolution.
(3) supervised learning
The initial network model of the input layer, several hidden layers and the output layer that include successively connecting is created, described defeated
Enter the data progress forward-propagating that layer is inputted in the normalised training set.The process of forward-propagating is to utilize deep learning
Network carries out the extraction of high-level characteristic to input data, is successively trained, the output of first layer as the second layer input, according to
This analogizes.
For example, being to include L hidden layer by the initial network model specification of the deep learning., will be pre- when training beginning
Training set data after processing (after normalization) carries out forward-propagating as input on network model.By forward-propagating,
Input data is handled in layer from input layer by hidden layer, output layer is additionally provided with after l-th hidden layer, through feature
Data after extraction will be sent to the output layer.
Then, the result that forward-propagating is exported on network model is compared with the data in the tally set, root
According to comparative result regulating networks model until determining optimal network model, the optimal network model is regard as grader.
In one embodiment, can be by the output result of forward-propagating on network model and the number in the tally set
According to measurement error, the hyper parameter of whole network is adjusted using backpropagation and gradient descent method so that anticipation error tends to minimum,
So as to obtain final hyper parameter.Final hyper parameter is the parameter for each layer of network for instigating anticipation error minimum, it is possible to use gained
Final hyper parameter create optimal network model.
The main purpose so done is, by supervised learning, and one neutral net of study carrys out automatic identification scanning body
Body different parts.That is, after supervised learning terminates, will learn to the network mould for scanning the minimum automatic identification of anticipation error
Type.
Step S2：Using the grader of acquistion in step sl, the data to magnetic resonance imaging body part on line are carried out
Identification, obtains classification results.
In one embodiment, line upper bit automatic identification is the data by magnetic resonance imaging body different parts on line
Optimal network model is tested, for example, according to defeated to testing with the data prediction identical mode in abovementioned steps S12
Enter data to be pre-processed, that is, being done in the same fashion normalized, then be entered into the grader of acquistion,
So as to obtain classification results, with the accuracy rate of statistics network.
Classification results mentioned herein are to recognize that it is body to carry out a width image with regard to that can automatically identify to image classification
Which position of body.Such as identify it is head, neck, waist or chest etc., point of human body different parts is divided into according to structure
Class result.
Step S3：Magnetic resonance imaging sequence is set according to the position recognized automatically, according to point obtained in step s 2
Class result, calls magnetic resonance imaging sequence automatically, completes to treat the scanning process for sweeping position.
If for example, recognizing neck, the magnetic resonance imaging sequence relevant with neck is set automatically, according in step S2
In obtained classification results, the magnetic resonance imaging sequence relevant with neck in the classification results is called automatically, so as to complete pair
The scanning process of neck.
The method of the present invention is the body part automatic identification magnetic resonance scanning method based on deep learning, and it not only makes to sweep
Retouch and eliminate the reliance on operator's naked eyes identification body part, save complicated positioning and parameter setting time, and advantageously ensure that list
Data consistency between individual tested position Multiple-Scan, avoids the factor and individual subjective factor of scan operation personnel to a certain extent
Influence.
In addition, present invention also offers a kind of computer-readable medium, the computer-readable medium, which has, to be stored therein
Program, the program is that computer can perform, so that computer performs the processing of above-mentioned various steps.
Present invention also offers a kind of body part automatic identification magnetic resonance based on deep learning of utilization above method
Scanning means.The device, which contains disparate modules, is used to realize each step referred in the above method.
For example, the body part automatic identification magnetic resonance imaging equipment based on deep learning of the present invention can include：Line
Lower training classifier modules, it trains initial network mould using the magnetic resonance image data collection of the body different parts under line
Type, the optimal network model of one automatic identification body part of acquistion is used as grader；Line upper bit automatic identification module, its profit
The grader of acquistion in classifier modules is trained under the line, the data to magnetic resonance imaging body part on line are known
Not, classification results are obtained；And automatic setting scanning sequence module, it sets magnetic resonance imaging automatically according to the position recognized
Sequence, according to the classification results obtained in the line upper bit automatic identification module, calls magnetic resonance imaging sequence automatically, complete
Into treating the scanning process of sweeping position.
In one embodiment, in the line upper bit automatic identification module, grader is being utilized to magnetic resonance on line
Before the data of scanning body different parts are identified, data are normalized.
In one embodiment, train classifier modules to include data preparation module, data under the line in advance to locate
Manage module and supervised learning module.
The magnetic being collected into can be total to by the data preparation module with the magnetic resonance image data collection of collection of bodily different parts
The partial data that view data of shaking is concentrated is marked as tally set, using the partial data as checksum set, by magnetic resonance
Other data that view data is concentrated are used as training set.
In a specific embodiment, the tally set in the data preparation module includes magnetic resonance image data collection
In all body different parts data, mark body different parts with the label in tally set, the number in the checksum set
Data bulk is less than the data bulk in the training set.For the selection of tally set and training set, it may be referred in preceding method
Step S1 in specific descriptions.
The data preprocessing module can be configured to that the magnetic resonance image data is normalized.The normalizing
Change processing can be the unification that resolution ratio, visual field sizes interested and image angle are carried out to image.
The supervised learning module can be configured to create input layer, several hidden layers and defeated for including successively connecting
Go out the initial network model of layer, the data inputted in the input layer in the normalised training set compare in network mould
The result that forward-propagating is exported in type and the data in the tally set, according to comparative result regulating networks model until determining most
Excellent network model is used as grader.
In a specific embodiment, in the supervised learning module, determine optimal network model be by using
The hyper parameter of backpropagation and gradient descent method regulation whole network model utilizes the final super ginseng to obtain final hyper parameter
Number sets up optimal network model as grader.
In another specific embodiment of the described device of the present invention, the line upper bit automatic identification module can be wrapped
Include：Data preprocessing module, before the data of magnetic resonance imaging body different parts on line are identified using grader,
Data are normalized to improve processing quality and processing speed.
Although according to preferred embodiment, the present invention is described, exist fall into the scope of the invention it
Interior change, displacement and various replacement equivalents.It should also be noted that there are a variety of methods for realizing the present invention and be
The optional mode of system.It is therefore intended that appended claims are construed to include within the spirit and scope of the invention
All these changes, displacement and various replacement equivalents.
Claims (33)
1. a kind of body part automatic identification magnetic resonance scanning method based on deep learning, it is characterised in that including following step
Suddenly：
S1 initial network model) is trained using the magnetic resonance image data collection of the body different parts under line, acquistion is known automatically
The optimal network model of other body part is used as grader；
S2) using in S1) in acquistion grader, the data to magnetic resonance imaging body part on line are identified, are divided
Class result；And
S3 magnetic resonance imaging sequence) is set according to the position recognized automatically, according in S2) the middle classification results obtained, automatically
Magnetic resonance imaging sequence is called, completes to treat the scanning process for sweeping position.
2. the body part automatic identification magnetic resonance scanning method according to claim 1 based on deep learning, its feature
It is, step S1) comprise the following steps：
S11) the magnetic resonance image data collection of collection of bodily different parts, the part that the magnetic resonance image data being collected into is concentrated
Data are marked as tally set, using the partial data as checksum set, other numbers that magnetic resonance image data is concentrated
According to being used as training set；
S12) magnetic resonance image data is normalized；
S13 the initial network model of the input layer, several hidden layers and the output layer that include successively connecting) is created, described defeated
Enter the data that layer is inputted in the normalised training set, by result and the mark that forward-propagating is exported on network model
Label concentrate data carry out error measure, according to error measurement regulating networks model, determine optimal network model as point
Class device.
3. the body part automatic identification magnetic resonance scanning method according to claim 2 based on deep learning, its feature
It is, in step S11) in, the tally set includes the data for all body different parts that magnetic resonance image data is concentrated, and uses
Data bulk in label mark body different parts in tally set, the checksum set is less than the data number in the training set
Amount.
4. the body part automatic identification magnetic resonance scanning method according to claim 3 based on deep learning, its feature
It is, by the M being related in M magnetic resonance image data of a body part being collected intoLIndividual part magnetic resonance image data
It is marked using as tally set, wherein, the quantity M of the part magnetic resonance image dataLLess than the MRI number
According to the 10% of total quantity M.
5. the body part automatic identification magnetic resonance scanning method according to claim 4 based on deep learning, its feature
It is, the quantity M of the part magnetic resonance image dataLLess than the 5% of the magnetic resonance image data total quantity M.
6. the body part automatic identification magnetic resonance scanning method according to claim 4 based on deep learning, its feature
It is, the quantity M of the part magnetic resonance image dataLLess than the 1% of the magnetic resonance image data total quantity M.
7. the body part automatic identification magnetic resonance imaging side based on deep learning according to any one of claim 4-6
Method, it is characterised in that by the N being related in N number of magnetic resonance image data of another body part being collected intoLIndividual part magnetic resonance
View data is marked using as tally set,
Wherein, it is related to the quantity N of the part magnetic resonance image data of another body partLAccount for magnetic resonance image data
Total quantity N ratio is equal to the quantity M for the part magnetic resonance image data for being related to one body partLAccount for magnetic resonance
View data total quantity M ratio.
8. the body part automatic identification magnetic resonance imaging side based on deep learning according to any one of claim 4-6
Method, it is characterised in that by the N being related in N number of magnetic resonance image data of another body part being collected intoLIndividual part magnetic resonance
View data is marked using as tally set,
Wherein, it is related to the quantity N of the part magnetic resonance image data of another body partLAccount for the MRI
Data count amount N ratio is different from the quantity M for being related to the part magnetic resonance image data of one body partLAccount for
The ratio of the magnetic resonance image data total quantity M.
9. the body part automatic identification magnetic resonance scanning method according to claim 2 based on deep learning, its feature
It is, in step S13) in, it is by using backpropagation and gradient descent method regulation whole network to determine optimal network model
The hyper parameter of model sets up optimal network model to obtain final hyper parameter, by the use of the final hyper parameter and is used as grader.
10. the body part automatic identification magnetic resonance scanning method according to claim 1 based on deep learning, its feature
It is, in step S2) in, before the data of magnetic resonance imaging body different parts on line are identified using grader,
Data are normalized.
11. the body part automatic identification magnetic resonance scanning method based on deep learning according to claim 2 or 10, its
It is characterised by, the normalized is that image resolution ratio, visual field sizes interested and image angle are unified.
12. a kind of body part automatic identification magnetic resonance imaging equipment based on deep learning, it is characterised in that including：
Classifier modules are trained under line, it trains original net using the magnetic resonance image data collection of the body different parts under line
Network model, the optimal network model of acquistion automatic identification body part is used as grader；
Line upper bit automatic identification module, it is using the grader that acquistion in classifier modules is trained under the line, on line
The data of magnetic resonance imaging body part are identified, and obtain classification results；And
Automatic to set scanning sequence module, it sets magnetic resonance imaging sequence automatically according to the position recognized, according to described
The classification results obtained in line upper bit automatic identification module, call magnetic resonance imaging sequence automatically, and completion, which is treated, sweeps position
Scanning process.
13. the body part automatic identification magnetic resonance imaging equipment according to claim 12 based on deep learning, it is special
Levy and be, train classifier modules also to include under the line：
Data preparation module, it is configured to the magnetic resonance image data collection of collection of bodily different parts, by the magnetic resonance being collected into
The partial data that view data is concentrated is marked as tally set, using the partial data as checksum set, by magnetic resonance figure
As other data in data set are used as training set；
Data preprocessing module, it is configured to that the magnetic resonance image data is normalized；
Supervised learning module, its be configured to create include successively connect input layer, several hidden layers and output layer just
Beginning network model, the data inputted in the input layer in the normalised training set positive on network model will be passed
Broadcast the result of output and carry out error measure with the data in the tally set, according to error measurement regulating networks model, really
Optimal network model is determined as grader.
14. the body part automatic identification magnetic resonance imaging equipment according to claim 13 based on deep learning, it is special
Levy and be, the tally set in the data preparation module includes all body different parts that magnetic resonance image data is concentrated
Data, mark body different parts with the label in tally set, the data bulk in the checksum set is less than the training set
In data bulk.
15. the body part automatic identification magnetic resonance imaging equipment according to claim 14 based on deep learning, it is special
Levy and be, by the M being related in M magnetic resonance image data of a body part being collected intoLIndividual part MRI number
According to being marked using as tally set, wherein, the quantity M of the part magnetic resonance image dataLLess than the MRI
The 10% of data count amount M.
16. the body part automatic identification magnetic resonance imaging equipment according to claim 15 based on deep learning, it is special
Levy and be, the quantity M of the part magnetic resonance image dataLLess than the 5% of the magnetic resonance image data total quantity M.
17. the body part automatic identification magnetic resonance imaging equipment according to claim 15 based on deep learning, it is special
Levy and be, the quantity M of the part magnetic resonance image dataLLess than the 1% of the magnetic resonance image data total quantity M.
18. the body part automatic identification magnetic resonance based on deep learning according to any one of claim 15-17 is swept
Imaging apparatus, it is characterised in that by the N being related in N number of magnetic resonance image data of another body part being collected intoLIndividual part magnetic
Resonance image data are marked using as tally set,
Wherein, it is related to the quantity N of the part magnetic resonance image data of another body partLAccount for magnetic resonance image data
Total quantity N ratio is equal to the quantity M for the part magnetic resonance image data for being related to one body partLAccount for magnetic resonance
View data total quantity M ratio.
19. the body part automatic identification magnetic resonance based on deep learning according to any one of claim 15-17 is swept
Imaging apparatus, it is characterised in that by the N being related in N number of magnetic resonance image data of another body part being collected intoLIndividual part magnetic
Resonance image data are marked using as tally set,
Wherein, it is related to the quantity N of the part magnetic resonance image data of another body partLAccount for the MRI
Data count amount N ratio is different from the quantity M for being related to the part magnetic resonance image data of one body partLAccount for
The ratio of the magnetic resonance image data total quantity M.
20. the body part automatic identification magnetic resonance imaging equipment according to claim 13 based on deep learning, it is special
Levy and be, in the supervised learning module, it is by using backpropagation and gradient descent method to determine optimal network model
The hyper parameter of whole network model is adjusted to obtain final hyper parameter, optimal network model is set up using the final hyper parameter and makees
For grader.
21. the body part automatic identification magnetic resonance imaging equipment according to claim 12 based on deep learning, it is special
Levy and be, in the line upper bit automatic identification module, utilizing grader to magnetic resonance imaging body different parts on line
Data be identified before, data are normalized.
22. the body part automatic identification magnetic resonance imaging equipment based on deep learning according to claim 13 or 21,
Characterized in that, the normalized is that image resolution ratio, visual field sizes interested and image angle are unified.
23. a kind of computer-readable medium, the computer-readable medium has the program being stored therein, the program is computer
It is executable so that computer performs the processing comprised the following steps：
S1 initial network model) is trained using the magnetic resonance image data collection of the body different parts under line, acquistion is known automatically
The optimal network model of other body part is used as grader；
S2) using in S1) in acquistion grader, the data to magnetic resonance imaging body part on line are identified, are divided
Class result；And
S3 magnetic resonance imaging sequence) is set according to the position recognized automatically, according in S2) the middle classification results obtained, automatically
Magnetic resonance imaging sequence is called, completes to treat the scanning process for sweeping position.
24. computer-readable medium according to claim 23, it is characterised in that step S1) comprise the following steps：
S11) the magnetic resonance image data collection of collection of bodily different parts, the part that the magnetic resonance image data being collected into is concentrated
Data are marked as tally set, using the partial data as checksum set, other numbers that magnetic resonance image data is concentrated
According to being used as training set；
S12) magnetic resonance image data is normalized；
S13 the initial network model of the input layer, several hidden layers and the output layer that include successively connecting) is created, described defeated
Enter the data that layer is inputted in the normalised training set, by result and the mark that forward-propagating is exported on network model
Label concentrate data carry out error measure, according to error measurement regulating networks model, determine optimal network model as point
Class device.
25. computer-readable medium according to claim 24, it is characterised in that in step S11) in, the tally set
The data for all body different parts concentrated including magnetic resonance image data, body difference portion is marked with the label in tally set
Data bulk in position, the checksum set is less than the data bulk in the training set.
26. computer-readable medium according to claim 25, it is characterised in that be related to a body by what is be collected into
M in M magnetic resonance image data of positionLIndividual part magnetic resonance image data is marked using as tally set, wherein, it is described
The quantity M of part magnetic resonance image dataLLess than the 10% of the magnetic resonance image data total quantity M.
27. computer-readable medium according to claim 26, it is characterised in that the part magnetic resonance image data
Quantity MLLess than the 5% of the magnetic resonance image data total quantity M.
28. computer-readable medium according to claim 26, it is characterised in that the part magnetic resonance image data
Quantity MLLess than the 1% of the magnetic resonance image data total quantity M.
29. the computer-readable medium according to any one of claim 26-28, it is characterised in that by relating to for being collected into
And the N in N number of magnetic resonance image data of another body partLIndividual part magnetic resonance image data is marked to be used as label
Collection,
Wherein, it is related to the quantity N of the part magnetic resonance image data of another body partLAccount for magnetic resonance image data
Total quantity N ratio is equal to the quantity M for the part magnetic resonance image data for being related to one body partLAccount for magnetic resonance
View data total quantity M ratio.
30. the computer-readable medium based on deep learning according to any one of claim 26-28, its feature exists
In by the N being related in N number of magnetic resonance image data of another body part being collected intoLIndividual part magnetic resonance image data enters
Line flag using as tally set,
Wherein, it is related to the quantity N of the part magnetic resonance image data of another body partLAccount for the MRI
Data count amount N ratio is different from the quantity M for being related to the part magnetic resonance image data of one body partLAccount for
The ratio of the magnetic resonance image data total quantity M.
31. computer-readable medium according to claim 24, it is characterised in that in step S13) in, determine optimal net
Network model be by using backpropagation and gradient descent method regulation whole network model hyper parameter to obtain final hyper parameter,
Optimal network model, which is set up, by the use of the final hyper parameter is used as grader.
32. computer-readable medium according to claim 23, it is characterised in that in step S2) in, utilizing grader
Before the data of magnetic resonance imaging body different parts on line are identified, data are normalized.
33. the computer-readable medium according to claim 24 or 32, it is characterised in that the normalized is to figure
As resolution ratio, visual field sizes interested and image angle are unified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710447366.7A CN107292257A (en) | 2017-06-14 | 2017-06-14 | Body part automatic identification magnetic resonance scanning method and device based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710447366.7A CN107292257A (en) | 2017-06-14 | 2017-06-14 | Body part automatic identification magnetic resonance scanning method and device based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107292257A true CN107292257A (en) | 2017-10-24 |
Family
ID=60097139
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710447366.7A Pending CN107292257A (en) | 2017-06-14 | 2017-06-14 | Body part automatic identification magnetic resonance scanning method and device based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107292257A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107669273A (en) * | 2017-11-07 | 2018-02-09 | 上海联影医疗科技有限公司 | Magnetic resonance equipment scanning system, scan method and computer-readable recording medium |
| CN109124635A (en) * | 2018-09-25 | 2019-01-04 | 上海联影医疗科技有限公司 | Model generating method, MRI scan method and system |
| WO2019127550A1 (en) * | 2017-12-29 | 2019-07-04 | Shenzhen United Imaging Healthcare Co., Ltd. | Systems and methods for selecting a scanning device for an object |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1926444A (en) * | 2004-03-01 | 2007-03-07 | 皇家飞利浦电子股份有限公司 | All in one plan scan imaging for optimization of acquisition parameters |
| CN1967562A (en) * | 2005-11-15 | 2007-05-23 | 中华电信股份有限公司 | Facial identification method based on human facial features identification |
| US20080112605A1 (en) * | 2006-11-14 | 2008-05-15 | Siemens Corporate Research, Inc. | Methods and Apparatus for Automatic Body Part Identification and Localization |
| WO2012174495A2 (en) * | 2011-06-17 | 2012-12-20 | Carnegie Mellon University | Physics based image processing and evaluation process of perfusion images from radiology imaging |
| CN104106096A (en) * | 2012-02-17 | 2014-10-15 | 先进穆阿分析公司 | Method of classification of organs from a tomographic image |
| CN106096616A (en) * | 2016-06-08 | 2016-11-09 | 四川大学华西医院 | Magnetic resonance image feature extraction and classification method based on deep learning |
| CN106203432A (en) * | 2016-07-14 | 2016-12-07 | 杭州健培科技有限公司 | A kind of localization method of area-of-interest based on convolutional Neural net significance collection of illustrative plates |
-
2017
- 2017-06-14 CN CN201710447366.7A patent/CN107292257A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1926444A (en) * | 2004-03-01 | 2007-03-07 | 皇家飞利浦电子股份有限公司 | All in one plan scan imaging for optimization of acquisition parameters |
| CN1967562A (en) * | 2005-11-15 | 2007-05-23 | 中华电信股份有限公司 | Facial identification method based on human facial features identification |
| US20080112605A1 (en) * | 2006-11-14 | 2008-05-15 | Siemens Corporate Research, Inc. | Methods and Apparatus for Automatic Body Part Identification and Localization |
| WO2012174495A2 (en) * | 2011-06-17 | 2012-12-20 | Carnegie Mellon University | Physics based image processing and evaluation process of perfusion images from radiology imaging |
| CN104106096A (en) * | 2012-02-17 | 2014-10-15 | 先进穆阿分析公司 | Method of classification of organs from a tomographic image |
| CN106096616A (en) * | 2016-06-08 | 2016-11-09 | 四川大学华西医院 | Magnetic resonance image feature extraction and classification method based on deep learning |
| CN106203432A (en) * | 2016-07-14 | 2016-12-07 | 杭州健培科技有限公司 | A kind of localization method of area-of-interest based on convolutional Neural net significance collection of illustrative plates |
Non-Patent Citations (2)
| Title |
|---|
| HOO-CHANG SHIN等: "Stacked Autoencoders for Unsupervised Feature Learning and Multiple Organ Detection in a Pilot Study Using 4D Patient Data", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 王斌 等主编: "《高校学生社团建设的理论与实践》", 30 June 2014, 四川大学出版社 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107669273A (en) * | 2017-11-07 | 2018-02-09 | 上海联影医疗科技有限公司 | Magnetic resonance equipment scanning system, scan method and computer-readable recording medium |
| WO2019127550A1 (en) * | 2017-12-29 | 2019-07-04 | Shenzhen United Imaging Healthcare Co., Ltd. | Systems and methods for selecting a scanning device for an object |
| CN111527565A (en) * | 2017-12-29 | 2020-08-11 | 上海联影医疗科技有限公司 | System and method for selecting scanning device of subject |
| US11562821B2 (en) | 2017-12-29 | 2023-01-24 | Shanghai United Imaging Healthcare Co., Ltd. | Systems and methods for selecting a scanning device for an object |
| CN111527565B (en) * | 2017-12-29 | 2023-03-31 | 上海联影医疗科技股份有限公司 | System and method for selecting scanning device of subject |
| CN109124635A (en) * | 2018-09-25 | 2019-01-04 | 上海联影医疗科技有限公司 | Model generating method, MRI scan method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110544274B (en) | Multispectral-based fundus image registration method and system | |
| CN109767440A (en) | A kind of imaged image data extending method towards deep learning model training and study | |
| CN107292257A (en) | Body part automatic identification magnetic resonance scanning method and device based on deep learning | |
| CN110464380A (en) | A kind of method that the ultrasound cross-section image of the late pregnancy period fetus of centering carries out quality control | |
| CN108416821B (en) | A kind of CT Image Super-resolution Reconstruction method of deep neural network | |
| CN109671076A (en) | Blood vessel segmentation method, apparatus, electronic equipment and storage medium | |
| CN107529650A (en) | Network model construction and closed loop detection method, corresponding device and computer equipment | |
| CN110992351B (en) | sMRI image classification method and device based on multi-input convolution neural network | |
| CN107369160A (en) | A kind of OCT image median nexus film new vessels partitioning algorithm | |
| CN107506770A (en) | Diabetic retinopathy eye-ground photography standard picture generation method | |
| CN110169770A (en) | The fine granularity visualization system and method for mood brain electricity | |
| CN107292884A (en) | The method and device of oedema and hemotoncus in a kind of identification MRI image | |
| CN109447962A (en) | A kind of eye fundus image hard exudate lesion detection method based on convolutional neural networks | |
| CN107669273A (en) | Magnetic resonance equipment scanning system, scan method and computer-readable recording medium | |
| CN109447976A (en) | A kind of medical image cutting method and system based on artificial intelligence | |
| CN110246137A (en) | A kind of imaging method, device and storage medium | |
| Tu et al. | Automatic measurement algorithm of scoliosis Cobb angle based on deep learning | |
| CN110136147A (en) | A kind of method, apparatus and storage medium of the Medical Image Segmentation based on U-Net model | |
| CN105534534A (en) | Emotion recognition method, device and system based on real-time functional magnetic resonance | |
| CN113707268A (en) | Rehabilitation training evaluation method and system | |
| CN110309692A (en) | Face identification method, apparatus and system, model training method and device | |
| CN110223275A (en) | A kind of cerebral white matter fiber depth clustering method of task-fMRI guidance | |
| CN109864737A (en) | The recognition methods of P wave and system in a kind of electrocardiosignal | |
| CN109215040A (en) | A kind of tumor of breast dividing method based on multiple dimensioned weighting study | |
| CN109658385A (en) | Eye fundus image judgment method and equipment |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171024 |