CN109584209A - Vascular wall patch identifies equipment, system, method and storage medium - Google Patents
Vascular wall patch identifies equipment, system, method and storage medium Download PDFInfo
- Publication number
- CN109584209A CN109584209A CN201811269040.0A CN201811269040A CN109584209A CN 109584209 A CN109584209 A CN 109584209A CN 201811269040 A CN201811269040 A CN 201811269040A CN 109584209 A CN109584209 A CN 109584209A
- Authority
- CN
- China
- Prior art keywords
- patch
- vascular wall
- deep learning
- network
- initial pictures
- 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
- 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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- 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
- G06T2207/30101—Blood vessel; Artery; Vein; Vascular
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
Suitable medical technical field of the present invention provides a kind of vascular wall patch identification equipment, system, method and storage medium, first acquisition vascular wall magnetic resonance MRI image；The patch in the MRI image is identified using deep learning method.In this way, carrying out the identification of vascular wall patch using deep learning method, artificial, the accuracy of raising patch identification, to improve recognition efficiency and can guarantee recognition accuracy can be greatly reduced.Comprehensive, accurate image quided is carried out to cerebral arterial thrombosis relevant blood vessel bed patch using MRI, and is quick and precisely diagnosed using artificial intelligence, high risk population of stroke screening and the cause of disease are detected to prevent from sending out again and have a very important significance.
Description
Technical field
The invention belongs to field of medical technology more particularly to a kind of vascular wall patch identification equipment, system, method and storages
Medium.
Background technique
Magnetic resonance imaging (Magnetic Resonance Imaging, MRI) is currently the only can clearly to show whole body
The non-invasive imaging method of atherosclerotic plaque.Vascular wall MRI not only can be to entocranial artery, arteria carotis and aorta etc.
System vascular patch carries out quantitative analysis, can also accurately identify fibrous cap, the bleeding, calcification, lipid core, inflammation of vulnerable plaque
Etc. unstability feature, be best patch imaging method generally acknowledged at present.
With the unique advantage of the universal and MRI patch imaging of the production domesticization and social application of MRI machine, using MRI
Comprehensive patch screening is carried out to high risk population of stroke and the cerebral apoplexy cause of disease is sought, the prevention and treatment of China's future cerebral apoplexy will be become
Important means.Moreover, because the data volume of three-dimensional high definition MRI vascular wall imaging is huge, the image of every examiner be can reach
500 width, experienced medical practitioner at least need to spend the diagnosis that could complete an examiner in 30 minutes, and workload is bigger than normal,
Efficiency is relatively low, and recognition accuracy can not be effectively ensured because of the appearance of situations such as doctor's fatigue.
Summary of the invention
The purpose of the present invention is to provide a kind of vascular wall patch identification equipment, system, method and storage mediums, it is intended to solve
Certainly present in the prior art, efficiency caused by manual identified vascular wall patch is relatively low and recognition accuracy can not be effectively ensured
Problem.
On the one hand, the present invention provides a kind of vascular wall patches to identify equipment, comprising: memory and processor, the place
Reason device realizes following steps when executing the computer program stored in the memory:
Obtain vascular wall magnetic resonance MRI image；
The patch in the MRI image is identified using deep learning method.
Further, the patch in the MRI image is identified using deep learning method, specifically includes following steps
It is rapid:
The MRI image is pre-processed, initial pictures are obtained；
The initial pictures are input to the identification that deep learning neural network carries out the patch, obtain recognition result.
Further, the initial pictures are input to the identification that deep learning neural network carries out the patch, specifically
Include the following steps:
Feature extraction processing is carried out to the initial pictures, obtains convolution characteristic image；
Candidate region is determined to the convolution characteristic image, accordingly obtains full connection features figure；
Classified based on the full connection features figure, obtains the recognition result.
Further, feature extraction processing is carried out to the initial pictures, obtains convolution characteristic image, specifically:
Feature extraction processing is carried out to the initial pictures using several residual error convolutional neural networks,
It wherein, include that convolutional network layer, activation primitive network layer and batch normalize in the residual error convolutional neural networks
Network layer.
Further, feature extraction processing is carried out to the initial pictures using several residual error convolutional neural networks, specifically
Include the following steps:
Network layer is normalized by the batch to average to the batch data of input；
The variance of the batch data is sought according to the mean value；
According to the mean value and the variance, the batch data is standardized, obtains batch normal data；
The batch normal data is handled using Dynamic gene, obtains having and the batch data of input
Same or similar adjustment in batches data are distributed to be exported.
On the other hand, the present invention provides a kind of vascular wall patch identifying system, the system comprises:
Module is obtained, for obtaining vascular wall magnetic resonance MRI image；And
Identification module, for being identified using deep learning method to the patch in the MRI image.
Further, the identification module specifically includes:
Preprocessing module obtains initial pictures for pre-processing to the MRI image；And
Deep learning module, for the initial pictures to be input to the knowledge that deep learning neural network carries out the patch
Not, recognition result is obtained.
Further, the deep learning module specifically includes:
Convolution module obtains convolution characteristic image for carrying out feature extraction processing to the initial pictures；
Candidate frame module accordingly obtains full connection features figure for determining candidate region to the convolution characteristic image；With
And
Full link block obtains the recognition result for classifying based on the full connection features figure.
On the other hand, the present invention also provides a kind of recognition methods of vascular wall patch, the method includes the following steps:
Obtain vascular wall magnetic resonance MRI image；
The patch in the MRI image is identified using deep learning method.
On the other hand, the present invention also provides a kind of computer readable storage medium, the computer readable storage mediums
It is stored with computer program, is realized when the computer program is executed by processor such as the step in the above method.
In the present invention, vascular wall magnetic resonance MRI image is obtained first；Using deep learning method in the MRI image
Patch identified.In this way, carrying out the identification of vascular wall patch using deep learning method, it can greatly reduce manually, mention
The accuracy of high patch identification, to improve recognition efficiency and can guarantee recognition accuracy.Using MRI to cerebral arterial thrombosis
Relevant blood vessel bed patch carries out comprehensive, accurate image quided, and is quick and precisely diagnosed using artificial intelligence, to cerebral apoplexy
People at highest risk's screening and the cause of disease are detected to prevent from sending out again and have a very important significance.
Detailed description of the invention
Fig. 1 is the structural schematic diagram for the vascular wall patch identification equipment that the embodiment of the present invention one provides；
Fig. 2 is the flow chart of processor institute implementation method in the embodiment of the present invention two；
Fig. 3 is the configuration diagram of deep learning neural network in the embodiment of the present invention three；
Fig. 4 is the process flow diagram of deep learning neural network in the embodiment of the present invention three；
Fig. 5 is the configuration diagram of residual error convolutional neural networks in the embodiment of the present invention four；
Fig. 6 is the process flow diagram for normalizing network layer in the embodiment of the present invention five in batches；
Fig. 7 is the structural schematic diagram for the vascular wall patch identifying system that the embodiment of the present invention six provides；
Fig. 8 is the structural schematic diagram of identification module in the embodiment of the present invention seven；
Fig. 9 is the structural schematic diagram of deep learning module in the embodiment of the present invention eight；
Figure 10 is the process flow diagram for the vascular wall patch recognition methods that the embodiment of the present invention ten provides；
Figure 11 is the configuration diagram of the deep learning neural network of an application example of the invention.
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.
Specific implementation of the invention is described in detail below in conjunction with specific embodiment:
Embodiment one:
Fig. 1 shows the vascular wall patch identification equipment of the offer of the embodiment of the present invention one, which is mainly used for: utilizing people
Work intelligence (Artificial Intelligence, AI) technology, carries out intelligent recognition to the patch in vascular wall MRI image, should
Equipment can be individual computer, chip, can also physically integrate with other equipment, such as: it is integrated with MRI machine,
It can also show as the forms such as cloud server.Vascular wall patch is broadly divided into stable patch and Vulnerable plaque, unstable
Patch is easy to fall off from vascular wall and cause thrombus, and Vulnerable plaque has fibrous cap, bleeding, calcification, lipid core, inflammation etc. no
Stability features, using AI technology carry out vascular wall patch identify when, can not only recognise that there are vascular wall patch,
It can recognize that vascular wall plaque type.For ease of description, only parts related to embodiments of the present invention are shown, is described in detail such as
Under:
Vascular wall patch identification equipment includes: memory 101 and processor 102, and processor executes to be deposited in memory 101
Following steps are realized when the computer program 103 of storage: then acquisition vascular wall MRI image first utilizes deep learning method pair
Patch in MRI image is identified.It in the present embodiment, is the transmission for realizing the data, signaling such as image, which can be with
Including network module；For the output such as display for realizing recognition result, the equipment can also including display screen etc. output modules；For reality
Existing manual control, the equipment can also include the input modules such as mouse, keyboard.Vascular wall MRI image typically refers to vascular wall and cuts
Picture.In the present embodiment, suitable any deep learning method can be used, realize to spot in vascular wall MRI image
The identification of block, such as: region convolutional neural networks (Regions with Convolutional Neural Network, R-
CNN), fast area convolutional neural networks (Fast R-CNN), more classification single pole detector (Single Shot MultiBox
Detector, SSD) etc..
Implement the present embodiment, the identification of vascular wall patch is carried out using deep learning method, can greatly reduce manually, mention
The accuracy of high patch identification, to improve recognition efficiency and can guarantee recognition accuracy.Using MRI to cerebral arterial thrombosis
Relevant blood vessel bed patch carries out comprehensive, accurate image quided, and is quick and precisely diagnosed using artificial intelligence, to cerebral apoplexy
People at highest risk's screening and the cause of disease are detected to prevent from sending out again and have a very important significance.
Embodiment two:
The present embodiment is on the basis of embodiment one, it is further provided following content:
In the present embodiment, when processor 102 executes the computer program 103 stored in memory 101, specific implementation is as schemed
Step in method shown in 2:
In step s 201, above-mentioned MRI image is pre-processed, obtains initial pictures.In the present embodiment, it pre-processes
It can be related to the cutting to image, to reduce redundant computation.
In step S202, initial pictures are input to the identification that deep learning neural network carries out patch, are identified
As a result.In the present embodiment, the framework of deep learning neural network can accordingly use R-CNN framework, Fast R-CNN framework, add
Fast region convolutional neural networks (Faster R-CNN) framework, SSD framework, masked areas convolutional neural networks (Mask R-CNN)
Framework etc..
Embodiment three:
The present embodiment is on the basis of embodiment two, it is further provided following content:
In the present embodiment, as shown in figure 3, deep learning neural network specifically includes: convolution sub-network 301, candidate frame
Network 302 and full connection sub-network 303.Wherein, the processing of each sub-network is approximately as and each sub-network processing correspondence is above-mentioned
The refinement process of step S202:
Step S401 as shown in Figure 4 can be performed in convolution sub-network 301, carries out feature extraction processing to initial pictures, obtains
To convolution characteristic image.In the present embodiment, convolution sub-network 301 may include multistage convolutional neural networks, every section of convolutional Neural
Residual error convolutional neural networks can be used to slow down the problems such as gradient is disappeared with gradient explosion in network, can also use non-residual error convolution
Neural network, certainly, convolution sub-network 301 can also use the group of non-residual error convolutional neural networks and residual error convolutional neural networks
It closes.
Candidate frame sub-network 302, can be performed step S402 as shown in Figure 4, determine candidate region to convolution characteristic image,
Accordingly obtain full connection features figure.In the present embodiment, the sliding window of predetermined size, base can be used in candidate frame sub-network 302
In the central point of each sliding window, predetermined quantity, candidate frame with predetermined size are generated on initial pictures, each
Candidate frame central point is corresponding with the central point of sliding window.It is corresponding to can get candidate region corresponding with each candidate frame.Often
One candidate region is corresponding to generate a candidate region characteristic pattern.Candidate region characteristic pattern can also be carried out accordingly at pool area
Reason, obtains full connection features figure.
Full connection sub-network 303, can be performed step S403 as shown in Figure 4, is classified etc. based on full connection features figure
Processing, obtains recognition result, recognition result indicates whether vascular wall patch.It in the present embodiment, can be in connection subnet entirely
The Liang Ge branch of network 303 such as is classified accordingly respectively, is returned at the processing, corresponding full connection sub-network 303 can correspond to comprising point
Class network layer and Recurrent networks layer.Sorter network layer can be used for judging that candidate region is prospect or back according to full connection features figure
Scape, namely judge that, with the presence or absence of vascular wall patch in candidate region, Recurrent networks layer can be used for correcting the coordinate of candidate frame, finally
Determine patch position.
Implement the present embodiment, realizes the identification to vascular wall patch using the convolutional neural networks based on region, can be improved
The accuracy of identification is conducive to the application of the AI artificial intelligence diagnosis using medical image.
Example IV:
The present embodiment is on the basis of embodiment three, it is further provided following content:
In convolution sub-network 301 and corresponding step S401, several residual error convolutional neural networks can be used to initial graph
As carrying out feature extraction processing, and residual error convolutional neural networks may include multiple network layers as shown in Figure 5: convolutional network layer
501, activation primitive network layer 502 and batch normalization network layer 503.Wherein, each network layer handles approximately as:
Convolutional network layer 501 can be realized using default convolution kernel carries out process of convolution to input picture.
Activation primitive network layer 502 can be linear using S type (Sigmoid) function, tanh (Tahn) function or rectification
Unit (The Rectified LinearUnit, ReLU) function etc. carries out activation processing.
Batch normalization network layer 503 is not only able to achieve traditional standardized processing, but also can enable the network to accelerate convergence,
Further slow down the problem of gradient disappears with gradient explosion.
Embodiment five:
The present embodiment is on the basis of example IV, it is further provided following content:
In the present embodiment, the processing of batch normalization network layer 503 specifically may include step as shown in Figure 6:
In step s 601, to input, handle resulting batch data via convolutional network layer 501 and average.
In step S602, the variance of batch data is sought according to mean value.
In step S603, according to mean and variance, batch data is standardized, obtains batch criterion numeral
According to.
In step s 604, batch normal data is handled using Dynamic gene, is obtained with the batch with input
The same or similar adjustment in batches data of the distribution of data are to be exported.In the present embodiment, Dynamic gene is in initialization
With corresponding initial value, it is then based on the initial value, Dynamic gene can parameter in reverse transfer, with network layer handles
It is trained together, Dynamic gene is enabled to learn the distribution of the batch data of input, the batch data of input is returned by batch
After one change processing, still retain the distribution of the batch data inputted originally.
Embodiment six:
Fig. 7 correspondingly shows the vascular wall patch identifying system of the offer of the embodiment of the present invention six, which equally mainly uses
In: AI technology is utilized, intelligent recognition is carried out to the patch in vascular wall MRI image, which can be individual computer, core
Piece is also possible to the group being made of computer, or the forms such as chipset as made of chip cascade.For ease of description,
Only parts related to embodiments of the present invention are shown, and details are as follows:
The vascular wall patch identifying system includes:
Module 701 is obtained, for obtaining vascular wall magnetic resonance MRI image；And
Identification module 702, for being identified using deep learning method to the patch in MRI image.
Module 701 and 702 phase of identification module are obtained in requisition for the content of explaination, class has been carried out in above-mentioned other embodiments
Like statement, details are not described herein again.
Embodiment seven:
The present embodiment is on the basis of embodiment six, it is further provided following content:
In the present embodiment, identification module 702 specifically includes structure as shown in Figure 8:
Preprocessing module 801 obtains initial pictures for pre-processing to MRI image；And
Deep learning module 802 is obtained for initial pictures to be input to the identification that deep learning neural network carries out patch
To recognition result.
Preprocessing module 801 and 802 phase of deep learning module in requisition for explaination content, in above-mentioned other embodiments into
It has gone similar to statement, details are not described herein again.
Embodiment eight:
The present embodiment is on the basis of embodiment seven, it is further provided following content:
In the present embodiment, deep learning module 802 specifically includes structure as shown in Figure 9:
Convolution module 901 obtains convolution characteristic image for carrying out feature extraction processing to initial pictures；
Candidate frame module 902 accordingly obtains full connection features figure for determining candidate region to convolution characteristic image；With
And
Full link block 903 obtains recognition result for classifying based on full connection features figure.
Equally, convolution module 901, candidate frame module 902 and complete 903 phase of link block in requisition for explaination content, upper
It states and has carried out similar statement in other embodiments, details are not described herein again.
Embodiment nine:
The present embodiment is on the basis of embodiment eight, it is further provided following content:
In the present embodiment, convolution module 901 specifically can be used several residual error convolutional neural networks and carry out to initial pictures
Feature extraction processing.It wherein, may include convolutional network layer 501 still as shown in Figure 5, activation primitive in residual error convolutional neural networks
Network layer 502 and batch normalization network layer 503.The processing of specific each network layer repeats no more.
Embodiment ten:
Figure 10 correspondingly show the embodiment of the present invention ten offer vascular wall patch recognition methods, this method specifically include as
Lower step:
In step S1001, vascular wall MRI image is obtained.
In step S1002, the patch in MRI image is identified using deep learning method.
Wherein, the content that the content of each step can be described with corresponding position in the various embodiments described above is similar, no longer superfluous herein
It states.
Embodiment 11:
In embodiments of the present invention, a kind of computer readable storage medium is provided, which deposits
Computer program is contained, the step in above method embodiment is realized when which is executed by processor, for example, Figure 10
Shown step S1001 to S1002.Alternatively, the computer program is realized when being executed by processor in above-mentioned each system embodiment
Described function, such as: the function of above-mentioned deep learning neural network.
The computer readable storage medium of the embodiment of the present invention may include can carry computer program code any
Entity or device, recording medium, for example, the memories such as ROM/RAM, disk, CD, flash memory.
Below by an application example, deep learning neural network involved in the various embodiments described above is carried out specific
Explanation.
The deep learning neural network can be used for identifying vascular wall patch, specifically may include frame as shown in figure 11
Structure:
Entire depth learning neural network includes: convolution sub-network 301, candidate frame sub-network 302 and connects sub-network entirely
303。
Convolution sub-network 501 includes: first segment convolutional neural networks 1101, pond layer 1102, second segment convolutional Neural net
Network 1103, third section convolutional neural networks 1104 and the 4th section of convolutional neural networks 1105.Wherein, first segment convolutional neural networks
1101 use non-residual error convolutional neural networks, and second segment convolutional neural networks 1103, third section convolutional neural networks 1104 and
4th section of convolutional neural networks 1105 use residual error convolutional neural networks.It include multiple network layers in residual error convolutional neural networks,
Still as shown in Figure 5: convolutional network layer 501, activation primitive network layer 502 and batch normalization network layer 503.
Candidate frame sub-network 302 includes: region candidate network (RegionProposal Network, RPN) 1106 and area
Domain pond network 1107.
Full connection sub-network 303 includes: sorter network layer 1108 and Recurrent networks layer 1109.
It further include the 5th section of convolutional neural networks 1111 between candidate frame sub-network 302 and full connection sub-network 303.
A mask network layer 1110 is also set after 5th section of convolutional neural networks 1111.
The treatment process of the above deep learning neural network is approximately as stating:
1, handled after gained vascular wall MRI image cut etc. pretreatment by each projected image, obtain size be 224 ×
224 initial pictures.Vascular wall MRI image referred to herein is usually sectioning image.
2, initial pictures are input to the initial characteristics extraction that first segment convolutional neural networks 1101 carry out convolutional calculation, gained
The characteristic pattern arrived is after the processing of pond layer 1102, then exports to second segment convolutional neural networks 1103, third section convolutional Neural net
Network 1104 and the 4th section of convolutional neural networks 1105 carry out further feature extraction.First segment convolutional neural networks 1101 carry out
Convolution kernel size used by convolutional calculation is 7 × 7, and step-length 2 can be such that data size halves, first segment convolutional neural networks
The characteristic pattern of 1101 outputs is having a size of 112 × 112.The characteristic pattern that first segment convolutional neural networks 1101 export is through pond layer 1102
After processing, characteristic pattern is obtained having a size of 56 × 56.
Convolutional network layer 501 in used residual error convolutional neural networks can be used following formula (1) and be calculated:
Wherein, i, j are the pixel coordinate position of input picture, and I is input image data, and K is convolution kernel, and p, n are respectively
The width and height of convolution kernel, S (i, j) are the convolved data of output.
Batch normalization network layer 503 can be calculated as follows:
Firstly, being averaged μ using following formula (2) to the batch data of inputβ.Batch data β=x of input1...mIt is
The output data of convolutional network layer 501.
Wherein, m is data count.
Secondly, seeking the variances sigma of batch data according to mean value using following formula (3)β 2。
Then, batch data is standardized according to mean value and variance using following formula (4), obtains batch
Normal data
Wherein, ∈ is the small positive number for avoiding divisor from being zero.
Then, using following formula (5), batch normal data is handled using Dynamic gene α, ω, is had
For the same or similar adjustment in batches data of distribution with the batch data of input to be exported, output can be used as next activation letter
The input of number network layer 502.
Wherein, α is zoom factor, and ω is shift factor, and Dynamic gene α, ω have corresponding initial value in initialization,
In this application example, the initial value that the initial value of α is approximately equal to 1, ω is approximately equal to 0, is then based on the initial value, Dynamic gene α,
ω can be trained together with the parameter of network layer handles in reverse transfer, thus, α, ω have just learnt the batch of input
The distribution of data, the batch data of input still retain the distribution of the batch data inputted originally after batch normalized.
Activation primitive network layer 502 can carry out calculating shown in following formula (6):
Wherein, x is the output data that batch normalizes network layer 503, and f (x) is the output of activation primitive network layer 502.
Three kinds of operations of above-mentioned convolutional network layer 501, activation primitive network layer 502 and batch normalization network layer 503
Constitute a neural network block.Second segment convolutional neural networks 1103 have 3 neural network blocks, wherein a kind of neural network
Convolution kernel size employed in block is 1 × 1, and convolution nuclear volume is 64；Convolution kernel employed in another neural network block
Size is 3 × 3, and convolution nuclear volume is 64；There are also the convolution kernel sizes used in a kind of neural network block for 1 × 1, convolution nucleus number
Amount is 256.Third section convolutional neural networks 1104 have 4 neural network blocks, wherein volume employed in a kind of neural network block
Product core size is 1 × 1, and convolution nuclear volume is 128；Convolution kernel size employed in another neural network block is 3 × 3, volume
Product nuclear volume is 128；It is 1 × 1 there are also the convolution kernel size used in a kind of neural network block, convolution nuclear volume is 512.4th
Section convolutional neural networks 1105 have 23 neural network blocks, wherein convolution kernel size employed in a kind of neural network block is 1
× 1, convolution nuclear volume is 256；Convolution kernel size employed in another neural network block is 3 × 3, and convolution nuclear volume is
256；It is 1 × 1 there are also the convolution kernel size used in a kind of neural network block, convolution nuclear volume is 1024.Eventually by first
To the 4th section of convolutional neural networks, the convolution characteristic image of output is 14 × 14 × 1024, indicates that output convolution characteristic image is big
Small is 14 × 14, and convolution nuclear volume is 1024.
3, resulting convolution characteristic image is handled through convolution sub-network 301 to subsequently input to RPN1106, pool area network
Respective handling is carried out in 1107.
RPN1106 is for extracting candidate region, specifically, using the sliding window of predetermined size 3 × 3, is based on each
The central point of sliding window, it is 9, the candidate frame with predetermined size, each candidate that predetermined quantity is generated on initial pictures
Frame central point is corresponding with the central point of sliding window.It is corresponding to can get candidate region corresponding with each candidate frame.Each
Candidate region is corresponding to generate a candidate region characteristic pattern.Due to passing through first to fourth section of convolutional neural networks, the convolution of output
Characteristic image is 14 × 14 × 1024, and sliding window predetermined size is 3 × 3, and candidate frame predetermined quantity is 9, then, it can be corresponding
It obtains 256 candidate regions, and accordingly obtains 256 candidate region characteristic patterns, i.e., the 256 full connection features of dimension.Part is waited
Select the area size of frame identical, the area size of the part candidate frame is different from the area size of other parts candidate frame, candidate
The area of frame, length-width ratio can be according to obtained from settings.
Pool area network 1107 is used for the position coordinates according to candidate frame, and candidate region characteristic pattern pond is turned to fixed ruler
Very little pond characteristic pattern.The optional RoiAlign network of pool area network 1107.Candidate frame is obtained by regression model, generally floating
Points, RoiAlign network do not quantify floating number.To each candidate frame, candidate region characteristic pattern is divided into 7 × 7 lists
Member, fixes four coordinate positions in each cell, the value of four positions is calculated by bilinear interpolation method, then carries out most
Great Chiization operation.To each candidate frame, 7 × 7 × 1024 pond characteristic pattern is obtained, all pond characteristic patterns constitute initially complete
Connection features figure.
4, for initial full connection features figure after the 5th section of convolutional neural networks 1111 are handled, output phase answers final 7
× 7 × 2048 full connection features figure.5th section of convolutional neural networks 1111 have 3 neural network blocks, wherein a kind of nerve net
Convolution kernel size employed in network block is 1 × 1, and convolution nuclear volume is 512；Convolution employed in another neural network block
Core size is 3 × 3, and convolution nuclear volume is 512；There are also the convolution kernel sizes used in a kind of neural network block for 1 × 1, convolution
Nuclear volume is 2048.
The 5th section of final full connection features figure of the processing gained of convolutional neural networks 1111 enters full connection sub-network 803
Three branches: sorter network layer 1108, Recurrent networks layer 1109 and mask network layer 1110.Wherein, sorter network layer 1108 is used
Before inputting the 5th section of final full connection features figure of the processing gained of convolutional neural networks 1111, and judge that candidate region is with this
Scape or background export the array for 14 × 14 × 18, wherein " 18 " indicate that 9 candidate frames can export prospect or two kinds of background knots
Fruit.Recurrent networks layer 1209 is used for coordinate, the Gao Yukuan of predicting candidate frame center anchor point, corrects the coordinate of candidate frame, exports and is
14 × 14 × 36, wherein " 36 " indicate four endpoint values of 9 candidate frames.Mask network layer 1110 utilizes certain size 2 × 2
Candidate region characteristic pattern of the convolution kernel to calcification is accordingly determined as and Jing Guo position correction up-samples, obtain 14 × 14 ×
256 characteristic pattern carries out subsequent process of convolution to this feature figure, obtains 14 × 14 × 2 characteristic pattern, then carries out at exposure mask
Reason, is split prospect and background.In this application example, categorical measure 2 indicates whether or not there is Breast Calcifications stove, in addition, also
Calcification position can further be obtained.
Wherein, sorter network layer loss function used, for optimizing to classification in sub-network 303 is connected entirely
The following formula of calculating (7) shown in, for when classification results be there are when calcification, to the Recurrent networks layer that optimizes of recurrence
Shown in the following formula of the calculating of loss function (8).
Lcls=-logq ... formula (7)
Wherein, q is the probability really classified.
Wherein, b value is (ti-ti '), and ti is prediction coordinate, and ti ' is true coordinate.
And the optimization processing of mask process can be related to: in classification processing, carry out after activation primitive Sigmoid processing
The calculating of cross entropy.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. a kind of vascular wall patch identifies equipment characterized by comprising memory and processor, the processor execute institute
Following steps are realized when stating the computer program stored in memory:
Obtain vascular wall magnetic resonance MRI image；
The patch in the MRI image is identified using deep learning method.
2. equipment as described in claim 1, which is characterized in that using deep learning method to the patch in the MRI image
It is identified, specifically include the following steps:
The MRI image is pre-processed, initial pictures are obtained；
The initial pictures are input to the identification that deep learning neural network carries out the patch, obtain recognition result.
3. equipment as claimed in claim 2, which is characterized in that by the initial pictures be input to deep learning neural network into
The identification of the row patch, specifically include the following steps:
Feature extraction processing is carried out to the initial pictures, obtains convolution characteristic image；
Candidate region is determined to the convolution characteristic image, accordingly obtains full connection features figure；
Classified based on the full connection features figure, obtains the recognition result.
4. equipment as claimed in claim 3, which is characterized in that carry out feature extraction processing to the initial pictures, rolled up
Product characteristic image, specifically:
Feature extraction processing is carried out to the initial pictures using several residual error convolutional neural networks,
It wherein, include convolutional network layer, activation primitive network layer and batch normalization network in the residual error convolutional neural networks
Layer.
5. equipment as claimed in claim 4, which is characterized in that using several residual error convolutional neural networks to the initial pictures
Feature extraction processing is carried out, specifically include the following steps:
Network layer is normalized by the batch to average to the batch data of input；
The variance of the batch data is sought according to the mean value；
According to the mean value and the variance, the batch data is standardized, obtains batch normal data；
The batch normal data is handled using Dynamic gene, is obtained with the distribution with the batch data of input
Same or similar adjustment in batches data are to be exported.
6. a kind of vascular wall patch identifying system, which is characterized in that the system comprises:
Module is obtained, for obtaining vascular wall magnetic resonance MRI image；And
Identification module, for being identified using deep learning method to the patch in the MRI image.
7. system as claimed in claim 6, which is characterized in that the identification module specifically includes:
Preprocessing module obtains initial pictures for pre-processing to the MRI image；And
Deep learning module, for the initial pictures to be input to the identification that deep learning neural network carries out the patch,
Obtain recognition result.
8. system as claimed in claim 7, which is characterized in that the deep learning module specifically includes:
Convolution module obtains convolution characteristic image for carrying out feature extraction processing to the initial pictures；
Candidate frame module accordingly obtains full connection features figure for determining candidate region to the convolution characteristic image；And
Full link block obtains the recognition result for classifying based on the full connection features figure.
9. a kind of recognition methods of vascular wall patch, which is characterized in that the method includes the following steps:
Obtain vascular wall magnetic resonance MRI image；
The patch in the MRI image is identified using deep learning method.
10. a kind of computer readable storage medium, the computer-readable recording medium storage has computer program, and feature exists
In the computer program realizes the step in method as claimed in claim 9 when being executed by processor.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811269040.0A CN109584209B (en) | 2018-10-29 | 2018-10-29 | Vascular wall plaque recognition apparatus, system, method, and storage medium |
| PCT/CN2019/078488 WO2020087838A1 (en) | 2018-10-29 | 2019-03-18 | Blood vessel wall plaque recognition device, system and method, and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811269040.0A CN109584209B (en) | 2018-10-29 | 2018-10-29 | Vascular wall plaque recognition apparatus, system, method, and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109584209A true CN109584209A (en) | 2019-04-05 |
| CN109584209B CN109584209B (en) | 2023-04-28 |
Family
ID=65921085
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811269040.0A Active CN109584209B (en) | 2018-10-29 | 2018-10-29 | Vascular wall plaque recognition apparatus, system, method, and storage medium |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109584209B (en) |
| WO (1) | WO2020087838A1 (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859201A (en) * | 2019-02-15 | 2019-06-07 | 数坤（北京）网络科技有限公司 | A kind of noncalcified plaques method for detecting and its equipment |
| CN110215232A (en) * | 2019-04-30 | 2019-09-10 | 南方医科大学南方医院 | Ultrasonic patch analysis method in coronary artery based on algorithm of target detection |
| CN110738643A (en) * | 2019-10-08 | 2020-01-31 | 上海联影智能医疗科技有限公司 | Method for analyzing cerebral hemorrhage, computer device and storage medium |
| CN110852987A (en) * | 2019-09-24 | 2020-02-28 | 西安交通大学 | Vascular plaque detection method and device based on deep morphology and storage medium |
| CN111178369A (en) * | 2019-12-11 | 2020-05-19 | 中国科学院苏州生物医学工程技术研究所 | Medical image identification method and system, electronic device and storage medium |
| CN112216391A (en) * | 2020-10-22 | 2021-01-12 | 深圳市第二人民医院(深圳市转化医学研究院) | Method and device for assessing cerebral stroke morbidity risk based on carotid atherosclerosis condition |
| CN113096126A (en) * | 2021-06-03 | 2021-07-09 | 四川九通智路科技有限公司 | Road disease detection system and method based on image recognition deep learning |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111862049B (en) * | 2020-07-22 | 2024-03-29 | 齐鲁工业大学 | Brain glioma segmentation network system and brain glioma segmentation method based on deep learning |
| CN111951268B (en) * | 2020-08-11 | 2024-06-07 | 深圳蓝湘智影科技有限公司 | Brain ultrasound image parallel segmentation method and device |
| CN114469174A (en) * | 2021-12-17 | 2022-05-13 | 上海深至信息科技有限公司 | Artery plaque identification method and system based on ultrasonic scanning video |
| CN116012367B (en) * | 2023-02-14 | 2023-09-12 | 山东省人工智能研究院 | Deep learning-based stomach mucosa feature and position identification method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107507148A (en) * | 2017-08-30 | 2017-12-22 | 南方医科大学 | The method that the down-sampled artifact of MRI is removed based on convolutional neural networks |
| US20180068198A1 (en) * | 2016-09-06 | 2018-03-08 | Carnegie Mellon University | Methods and Software for Detecting Objects in an Image Using Contextual Multiscale Fast Region-Based Convolutional Neural Network |
| CN108542390A (en) * | 2018-03-07 | 2018-09-18 | 清华大学 | Vascular plaque ingredient recognition methods based on more contrast nuclear magnetic resonance images |
| CN108710829A (en) * | 2018-04-19 | 2018-10-26 | 北京红云智胜科技有限公司 | A method of the expression classification based on deep learning and the detection of micro- expression |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9730643B2 (en) * | 2013-10-17 | 2017-08-15 | Siemens Healthcare Gmbh | Method and system for anatomical object detection using marginal space deep neural networks |
| US10748277B2 (en) * | 2016-09-09 | 2020-08-18 | Siemens Healthcare Gmbh | Tissue characterization based on machine learning in medical imaging |
| CN107067396B (en) * | 2017-04-26 | 2021-11-16 | 中国人民解放军总医院 | Nuclear magnetic resonance image processing device and method based on self-encoder |
-
2018
- 2018-10-29 CN CN201811269040.0A patent/CN109584209B/en active Active
-
2019
- 2019-03-18 WO PCT/CN2019/078488 patent/WO2020087838A1/en active Application Filing
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180068198A1 (en) * | 2016-09-06 | 2018-03-08 | Carnegie Mellon University | Methods and Software for Detecting Objects in an Image Using Contextual Multiscale Fast Region-Based Convolutional Neural Network |
| CN107507148A (en) * | 2017-08-30 | 2017-12-22 | 南方医科大学 | The method that the down-sampled artifact of MRI is removed based on convolutional neural networks |
| CN108542390A (en) * | 2018-03-07 | 2018-09-18 | 清华大学 | Vascular plaque ingredient recognition methods based on more contrast nuclear magnetic resonance images |
| CN108710829A (en) * | 2018-04-19 | 2018-10-26 | 北京红云智胜科技有限公司 | A method of the expression classification based on deep learning and the detection of micro- expression |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859201A (en) * | 2019-02-15 | 2019-06-07 | 数坤（北京）网络科技有限公司 | A kind of noncalcified plaques method for detecting and its equipment |
| CN110215232A (en) * | 2019-04-30 | 2019-09-10 | 南方医科大学南方医院 | Ultrasonic patch analysis method in coronary artery based on algorithm of target detection |
| CN110852987A (en) * | 2019-09-24 | 2020-02-28 | 西安交通大学 | Vascular plaque detection method and device based on deep morphology and storage medium |
| CN110852987B (en) * | 2019-09-24 | 2022-04-22 | 西安交通大学 | Vascular plaque detection method and device based on deep morphology and storage medium |
| CN110738643A (en) * | 2019-10-08 | 2020-01-31 | 上海联影智能医疗科技有限公司 | Method for analyzing cerebral hemorrhage, computer device and storage medium |
| CN111178369A (en) * | 2019-12-11 | 2020-05-19 | 中国科学院苏州生物医学工程技术研究所 | Medical image identification method and system, electronic device and storage medium |
| CN111178369B (en) * | 2019-12-11 | 2023-12-19 | 中国科学院苏州生物医学工程技术研究所 | Medical image recognition method and system, electronic equipment and storage medium |
| CN112216391A (en) * | 2020-10-22 | 2021-01-12 | 深圳市第二人民医院(深圳市转化医学研究院) | Method and device for assessing cerebral stroke morbidity risk based on carotid atherosclerosis condition |
| CN112216391B (en) * | 2020-10-22 | 2024-05-10 | 深圳市第二人民医院(深圳市转化医学研究院) | Method and device for evaluating cerebral apoplexy incidence risk based on carotid atherosclerosis |
| CN113096126A (en) * | 2021-06-03 | 2021-07-09 | 四川九通智路科技有限公司 | Road disease detection system and method based on image recognition deep learning |
| CN113096126B (en) * | 2021-06-03 | 2021-09-24 | 四川九通智路科技有限公司 | Road disease detection system and method based on image recognition deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109584209B (en) | 2023-04-28 |
| WO2020087838A1 (en) | 2020-05-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109584209A (en) | Vascular wall patch identifies equipment, system, method and storage medium | |
| CN109300121B (en) | A kind of construction method of cardiovascular disease diagnosis model, system and the diagnostic device | |
| Kumar et al. | Resnet-based approach for detection and classification of plant leaf diseases | |
| CN110490850B (en) | Lump region detection method and device and medical image processing equipment | |
| CN107154043B (en) | Pulmonary nodule false positive sample inhibition method based on 3DCNN | |
| Van Grinsven et al. | Fast convolutional neural network training using selective data sampling: Application to hemorrhage detection in color fundus images | |
| CN109584254A (en) | A kind of heart left ventricle's dividing method based on the full convolutional neural networks of deep layer | |
| CN107657612A (en) | Suitable for full-automatic the retinal vessel analysis method and system of intelligent and portable equipment | |
| CN108830826A (en) | A kind of system and method detecting Lung neoplasm | |
| CN109325942A (en) | Eye fundus image Structural Techniques based on full convolutional neural networks | |
| CN107862694A (en) | A kind of hand-foot-and-mouth disease detecting system based on deep learning | |
| CN110459319A (en) | The assistant diagnosis system of breast molybdenum target image based on artificial intelligence | |
| CN113361623B (en) | Medical image classification method combining lightweight CNN with transfer learning | |
| CN112614133B (en) | Three-dimensional pulmonary nodule detection model training method and device without anchor point frame | |
| CN112329871B (en) | Pulmonary nodule detection method based on self-correction convolution and channel attention mechanism | |
| CN110689118A (en) | Improved target detection method based on YOLO V3-tiny | |
| CN111161278A (en) | Deep network aggregation-based fundus image focus segmentation method | |
| CN110363072A (en) | Tongue image recognition method, apparatus, computer equipment and computer readable storage medium | |
| CN109871869A (en) | A kind of Lung neoplasm classification method and its device | |
| CN110288574A (en) | A kind of adjuvant Ultrasonographic Diagnosis hepatoncus system and method | |
| Zhang et al. | Robust image segmentation method for cotton leaf under natural conditions based on immune algorithm and PCNN algorithm | |
| CN113269737A (en) | Method and system for calculating diameter of artery and vein of retina | |
| Miao et al. | Classification of diabetic retinopathy based on multiscale hybrid attention mechanism and residual algorithm | |
| Dong et al. | An improved YOLOv5 network for lung nodule detection | |
| CN112669319B (en) | Multi-view multi-scale lymph node false positive inhibition modeling method |
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 |