CN108961229A - Cardiovascular OCT image based on deep learning easily loses plaque detection method and system - Google Patents
Cardiovascular OCT image based on deep learning easily loses plaque detection method and system Download PDFInfo
- Publication number
- CN108961229A CN108961229A CN201810678850.5A CN201810678850A CN108961229A CN 108961229 A CN108961229 A CN 108961229A CN 201810678850 A CN201810678850 A CN 201810678850A CN 108961229 A CN108961229 A CN 108961229A
- Authority
- CN
- China
- Prior art keywords
- network
- region
- oct image
- image
- oct
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 66
- 238000013135 deep learning Methods 0.000 title claims abstract description 31
- 230000002526 effect on cardiovascular system Effects 0.000 title claims abstract description 24
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 50
- 238000000034 method Methods 0.000 claims abstract description 26
- 238000013434 data augmentation Methods 0.000 claims abstract description 11
- 239000000284 extract Substances 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims description 29
- 238000012360 testing method Methods 0.000 claims description 27
- 238000005070 sampling Methods 0.000 claims description 24
- 238000000605 extraction Methods 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 12
- 230000001629 suppression Effects 0.000 claims description 10
- 239000000203 mixture Substances 0.000 claims description 9
- 238000013500 data storage Methods 0.000 claims description 5
- 230000008602 contraction Effects 0.000 claims description 4
- 210000000352 storage cell Anatomy 0.000 claims description 4
- 241001573881 Corolla Species 0.000 claims description 3
- 238000012217 deletion Methods 0.000 claims description 3
- 230000037430 deletion Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000009826 distribution Methods 0.000 claims description 2
- 238000003475 lamination Methods 0.000 claims 1
- 238000013480 data collection Methods 0.000 abstract description 2
- 238000012014 optical coherence tomography Methods 0.000 description 64
- 239000000523 sample Substances 0.000 description 13
- 238000003745 diagnosis Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 5
- 238000003384 imaging method Methods 0.000 description 3
- 238000007689 inspection Methods 0.000 description 3
- 208000004476 Acute Coronary Syndrome Diseases 0.000 description 2
- 208000007536 Thrombosis Diseases 0.000 description 2
- 230000001154 acute effect Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 239000012141 concentrate Substances 0.000 description 2
- 210000004351 coronary vessel Anatomy 0.000 description 2
- 230000003902 lesion Effects 0.000 description 2
- 238000000638 solvent extraction Methods 0.000 description 2
- 208000037260 Atherosclerotic Plaque Diseases 0.000 description 1
- 208000024172 Cardiovascular disease Diseases 0.000 description 1
- 206010020608 Hypercoagulation Diseases 0.000 description 1
- 206010042434 Sudden death Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000000454 anti-cipatory effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000033228 biological regulation Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000001269 cardiogenic effect Effects 0.000 description 1
- 230000007211 cardiovascular event Effects 0.000 description 1
- 208000029078 coronary artery disease Diseases 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 230000006806 disease prevention Effects 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000013100 final test Methods 0.000 description 1
- 230000002962 histologic effect Effects 0.000 description 1
- 238000002608 intravascular ultrasound Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 208000010125 myocardial infarction Diseases 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 208000011580 syndromic disease Diseases 0.000 description 1
- 201000005665 thrombophilia Diseases 0.000 description 1
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
-
- G06T5/70—
-
- 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
- G06T2207/10101—Optical tomography; Optical coherence tomography [OCT]
-
- 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
Abstract
The present invention provides the cardiovascular OCT images based on deep learning easily to lose plaque detection method and system.Described method includes following steps: obtaining angiocarpy OCT image to be detected from OCT image documentation equipment instrument first and establishes image data set；By the patch region in specialist mark OCT image；Data augmentation is carried out to original digital image data collection by the way of scalloping；Original OCT image is denoised using a kind of full convolutional neural networks and extracts area-of-interest.Using object detection field, state-of-the-art Faster R-CNN frame positions and identifies the easy loss patch in OCT image at present；Output marks out the easily OCT image of loss plaque location and stores to image data.The experiment proved that the OCT image proposed by the present invention based on deep learning easily loses, plaque detection method and system discrimination is high, and accurate positioning, performance is stablized, and detection speed is exceedingly fast, and has stronger robustness and higher clinical value.
Description
Technical field
The present invention relates to the field of medical instrument technology, are related to a kind of based on depth learning technology detection angiocarpy OCT image
In easy loss patch method and system.
Background technique
The patch that vulnerable plaque (Vulnerable Plaque) refers to that those are unstable and has thrombophilia.Rapid wear
Patch and acute cardiovascular event are related closely, are to cause thrombus, cause acute coronary artery syndrome, coronary heart disease and even lead
The main reason for causing sudden death.There are about 19,000,000 patients because of acute coronary syndrome (ACS) and cardiogenic sudden every year in the world
It is dead and dead, because unstable Coronary Atherosclerotic Plaque rupture is further formed thrombus and leads to the patient of myocardial infarction
70% is accounted for, therefore cardiovascular easily loss plaque detection is of great significance in terms of cardiovascular disease prevention.
Optical coherence tomography (OCT) developed in recent years is image in current domestic and international newer coronary artery
Technology, resolution ratio are about 10 μm, are 10 times of IVUS resolution ratio.And OCT has tissue resolving ability, that is, passes through OCT image
Can basic understanding patch main component and histologic characteristics, therefore cardiovascular optical coherence tomography (IVOCT) technology at
For at present clinically for detecting the prefered method of cardiovascular vulnerable plaque.
The cardiovascular easily loss patch of clinical detection is that analysis OCT image is manually read by doctor at present, observes its texture
Feature and elasticity.It takes time and effort in this way and the case where mistaken diagnosis easily occurs.
The algorithm that patch is easily lost in current existing automatic detection IVOCT image mainly uses SVM classifier, uses
Shallow-layer CNN neural network etc..
Due to the characteristic of OCT image imaging, the more speckle noise that image contains, to the interference of testing result compared with
Greatly, therefore before detection easily loss patch it is very necessary the operations such as image accurately to be denoised, is divided.It is existing at present automatic
Denoising, partitioning algorithm are mainly the algorithm designed based on image grayscale, Gradient Features.
The method of the artificial read tablet of doctor used clinical at present needs doctor to have professional experiences knowledge abundant, and by
It is high in OCT image resolution, 200 multiple images will be generated to the bracket scanning of one section of 4cm, taken so that doctor diagnosed is time-consuming
Power is also easy to the case where failing to pinpoint a disease in diagnosis, mistaken diagnosis occur.
The algorithm that patch is easily lost in existing automatic detection is based primarily upon traditional algorithm such as SVM classifier or shallow-layer volume
Product neural network (CNN), the accuracy detected in this way is unsatisfactory, and detection speed also has to be hoisted.
Due to the imaging characteristic of OCT image, image contains more speckle noise, produces interference to detection lesion.
Current Denoising Algorithm is mainly based upon the algorithm of image grayscale, Gradient Features design, such as the Denoising Algorithm based on kernel function,
But can only realize and remove partial noise, it cannot still achieve satisfactory results.
Summary of the invention
Of the existing technology in order to solve the problems, such as, the present invention proposes that a kind of OCT image based on deep learning easily loses
Plaque detection method and system realize the easy loss patch in quick, accurate detection OCT image.
The specific technical proposal of the invention is:
Cardiovascular OCT image based on deep learning easily loses plaque detection method, comprises the following steps that
Step 1: obtaining angiocarpy OCT image from OCT image documentation equipment, data augmentation is carried out to image data set；
Step 1-1 obtains the data set being made of m OCT angiocardiograms, including containing easy loss patch
Positive sample and negative sample without easy loss patch, positive sample quantity are greater than negative sample quantity, and all positive samples are by Special Medical
Teacher marks out the specific location of easily loss patch；
Step 1-2 carries out data augmentation to the data set of m OCT angiocardiogram compositions, to obtain 4m~10m
Sample；The data augmentation mode includes that appropriate part expansion, contraction, torsion, corolla and the torsion of flag shape are carried out to original image
It is bent；
Step 2: being denoised and extracted area-of-interest to cardiovascular OCT image with full convolutional neural networks；
Data set is divided into training set, verifying collection and test set, training set and verifying each image pair of collection by step 2-1
A label figure is answered, the label figure is to remove background, conduit redundance, only retains interested region (RoI)；Test set
There is no label figure；By training set, the verifying collection input full convolutional network training, with the performance of test set test network；
After the full convolutional neural networks of step 2-2, step 2-1 show well on test set, by all images of data set
Area-of-interest is denoised and extracted using the full convolutional neural networks；
Step 3: being positioned and identified in cardiovascular OCT image with Faster R-CNN target detection frame and easily lose patch；
Faster R-CNN is the frame that current object detection field is most advanced, most popular.It walks four of target detection substantially
Suddenly within (candidate region generates, feature extraction, classification, position refine) unification to a depth network frame, all calculating do not have
There is repetition, completed in GPU completely, substantially increases the speed of service.
Training dataset image passes through full convolutional network processing described in step 2, and every figure correspondence is marked by specialist
Patch region co-ordinate position information；FasterR-CNN algorithm is made of two big modules: (1) network (Region is suggested in region
Proposal Network, RPN), (2) Fast R-CNN detects network；
Step 3-1 suggests that network extracts detection zone using region, it can be with the convolution of entire detection network share full figure
Feature, so that region suggests hardly taking time.Region suggests that network carries out feature extraction first, then generates candidate region
(anchor), window classification and position refine are finally carried out；
Step 3-2 suggests network for subsequent trained region, gives each candidate region distribution class label { target, non-targeted }；
The label setting in network candidates region, which is described as follows, to be suggested to region:
Target: it is greater than x with the registration of any reference standard (ground truth) bounding box (bounding box)
The candidate region of (IoU > x)；
It is non-targeted: to be less than the candidate region of (1-x) with the registration of all reference standard bounding boxs；
Since feature extraction network extracts large number of candidate frame, there are many overlapping regions between each candidate frame, adopt
Realize that detection block merges and deletion with non-maxima suppression method (NMS, non-maximum suppression)；Specifically, by institute
There is the score of candidate frame to sort, chooses best result and its corresponding frame；X (IoU will be greater than with the registration of top score candidate frame
> x) candidate frame all delete, be only left this top score candidate frame；After non-maxima suppression method, candidate region is arranged
Sequence takes the top n after sorting to detect；
Step 3-3 is realized most after region suggests that network (RPN) extraction obtains candidate region using Fast R-CNN module
Whole detection and identification makes the convolutional layer that network is suggested in region and Fast R-CNN shares ResNet101 by training；
Step 4: system stores detection data, output test result；
To each OCT image of OCT image documentation equipment acquisition, can be obtained after step 2, the processing of 3 deep learning methods
To marking out the output result for easily losing plaque region domain and its probability on image.
Further, the full convolutional network model framework that above-mentioned steps 2 use: the full convolutional neural networks by left side receipts
The extensions path on contracting path and right side composition, it is U-shaped integral into one.N times down-sampling is carried out to input picture and constitutes left side contraction
Path carries out n times up-sampling to image and constitutes right extension path.Wherein, constricted path by two repeated applications convolution kernel
Size be 3*3 convolutional layer form, each convolutional layer follow an amendment linear unit (rectified linear unit,
ReLU), down-sampling (down sampling) of A rear progress of every convolution, the maximum for the 2*2 that down-sampling operation is 2 by step-length
Pond layer (max pooling) is completed, and each down-sampling step doubles feature number of active lanes；Right extension path repeated application
The convolutional layer that two convolution kernel sizes are 3*3 each follows an amendment linear unit, carries out after every convolution A times primary
Up-sampling, up-sampling (up sampling) halve the quantity in feature channel；The convolution kernel of right extension path the last layer is big
It is small be 1*1, by characteristic pattern be converted into certain depth as a result, if the pixel of an image is divided into B classification, certain depth
Equal to B.
Further, in above-mentioned steps 3-1 feature extraction using deep learning network ResNet101；It generates candidate
Region is each position to the image, considers that x kind ratio, y kind size amount to the candidate window of x*y size；It carries out
When window classification and position refine, classification layer exports the probability that x*y candidate region on each position belongs to foreground and background；
Window, which returns layer and exports each candidate region in each position and correspond to window, should translate the parameter of scaling.
Further, above-mentioned steps 3-3, specific training process are as follows: suggest the Suggestion box that network generates using region,
By Fast R-CNN training, one is individually detected network, which initialized by ResNet101, is trained first
Suggest that network, fixed shared convolutional layer only finely tune region and suggest the exclusive layer of network, then keep shared convolutional layer in region
It is fixed, finely tune the full articulamentum of Fast R-CNN；The identical convolutional layer of two network shares constitutes a unified network.
The above-mentioned cardiovascular OCT image based on deep learning easily loses plaque detection system, including sequentially connected OCT shadow
As acquiring unit, digital signal processing unit, data storage cell, testing result display unit；
OCT image capturing unit: image is transferred to system from OCT image documentation equipment by connection OCT image documentation equipment and this system
In；
Digital signal processing unit: implementing entire detection algorithm, denoises including (1) to the original image that image documentation equipment is shot
Extract the easy loss patch in area-of-interest and (2) positioning, identification OCT image；
Data storage cell: the original image of result and the image documentation equipment shooting to OCT Image detection is stored；
Testing result display unit: output test result, result figure show OCT image in easily lose patch position and
Its probability.
Beneficial effects of the present invention are the cardiovascular easily loss spot of OCT marked using deep learning network and specialist
The image of block data set magnanimity trains the algorithm that can accurately detect easily loss patch, realizes the whole deep learning of diagnosis
Change, using the large volume data sets training deep learning network after expansion, realizes that picture is accurately divided, lesion accurately detects.
The Faster R-CNN target detection frame detection speed that the present invention uses is exceedingly fast, the average picture inspection on GPU server
Survey only needs 200ms.Solve the problems, such as that the time-consuming that patch is easily lost by doctor Artificial Diagnosis OCT image and misdiagnosis rate are higher.
Detailed description of the invention
Fig. 1 is the original OCT image schematic diagram under polar coordinate system.
Fig. 2 is the easy loss plaque detection method flow diagram based on deep learning.
Fig. 3 is that the present invention carries out saying for data augmentation to original OCT image data collection in such a way that scalloping deforms
It is bright, wherein (a) original OCT image, (b) carries out the distortion of flag shape to original image (a), local expansion (c) is carried out to original image (a)
Distortion, (d) carries out extruding distortion to original image (a).
Fig. 4 is the label figure (a) and original image (b) example of the training set production to full convolutional network used in step 2.
Fig. 5 is a kind of architecture diagram for full convolutional network U-net that detection method step 2 uses.
Fig. 6 is using the OCT striograph for only retaining area-of-interest after trained full convolutional network segmentation.
Fig. 7 is the target detection frame Faster R-CNN architecture diagram that the present invention uses.
Fig. 8 is the plaque location coordinate text examples marked by specialist.
Fig. 9 is the composition of the loss function used in Faster R-CNN detection framework figure of the invention.
Figure 10 is that detection block of the present invention merges and deletes schematic diagram.
Figure 11 is the detection final result figure of the easy loss plaque detection method and system based on deep learning.
Figure 12 is that the easy loss plaque detection system based on deep learning forms figure.
Figure 13 is the easy loss plaque detection method flow diagram based on deep learning.
Specific embodiment
Next specific embodiments of the present invention are described with reference to the accompanying drawings, make technology path of the invention, feature and excellent
Point is more understandable.
Step 1: obtaining angiocarpy OCT image from OCT image documentation equipment, data augmentation is carried out to image data set.
Step 1-1 obtains 2000 cardiovascular OCT images from equipment, establishes image data set, wherein 1000 for containing
The easily positive sample of loss patch, 1000 are the negative sample without easy loss patch.OCT raw video is as shown in Figure 1.
Step 1-2 carries out data augmentation to image data set referring to Fig. 2, is carried out using the method for scalloping to original image
The distortions such as part expansion, contraction, torsion, corolla and the flag shape of appropriateness increase image data to 4N~10N.Specifically, originally
Invention expands image data amount to original 4 times, obtains 8000 images.(a) in Fig. 2 (c), is respectively (d): one (b)
Original OCT image is opened, the distortion of flag shape is carried out to original image, part expansion is carried out to original image and is distorted, extruding distortion is carried out to original image.
The algorithm flow chart of step 2 and step 3 deep learning method, in brief, step 2 are given referring to Fig. 3, Fig. 3
Area-of-interest, step 3 are denoised and extracted to original cardiovascular OCT image using a kind of full convolutional neural networks (U-net)
It uses the OCT image handled through step 2 as input, is positioned using Faster R-CNN frame and identify the rapid wear in OCT image
Lose patch.
Step 2: being denoised and extracted area-of-interest to cardiovascular OCT image with a kind of full convolutional neural networks.
Step 2-1 concentrates the training for choosing 5000 pictures as full convolutional network from the OCT image data by expanding
Collection chooses 500 pictures as verifying collection from training set, randomly selects 500 conducts from remaining 3000 OCT images
Test set.Referring to Fig. 4 (a), label figure is opened for the OCT making video one in every training set, only retains area-of-interest, removes
Unrelated conduit, background, corner, Fig. 4 (b) are original image corresponding with Fig. 4 (a).
Referring to Fig. 5, Fig. 5 is a kind of framework for full convolutional network U-net that the present invention uses.The full convolutional neural networks by
The constricted path in left side and the extensions path composition on right side, it is U-shaped integral into one.Specifically, which carries out 4 to input picture
Secondary down-sampling constitutes left side constricted path, carries out 4 up-samplings to image and constitutes right extension path.Wherein, constricted path by
The convolutional layer that the convolution kernel size of two repeated applications is 3*3 forms, and each convolutional layer follows an amendment linear unit
(rectified linear unit, ReLU).Specifically, the present invention takes A=2, carries out a down-sampling after every convolution 2 times
(down sampling).The maximum pond layer (max pooling) for the 2*2 that down-sampling operation is 2 by step-length is completed, under each
Sampling step doubles feature number of active lanes.Repeated application two convolution kernel sizes in right extension path are the convolutional layer of 3*3, often
One all follows an amendment linear unit, is once up-sampled after every convolution 2 times, and up-sampling (up sampling) makes spy
The quantity in sign channel halves.The convolution kernel size of right extension path the last layer is 1*1, converts certain depth for characteristic pattern
As a result, the pixel by original OCT image specifically, in the present embodiment is needed to be divided into 2 classifications, then certain depth is equal to B=
2。
Particularly, be illustrated to the full convolutional network U-net of present invention training: setting momentum=0.90 in this way may be used
To use GPU memory to large extent, so that a large amount of former network sample for training determines in current optimizer more
Newly.The last layer introduces w using intersection entropy function and softmax in order to keep certain pixels more important in formula
(x).It is estimated to each mark image to have calculated a weight map, carry out the different frequency that compensation training concentrates every class pixel, makes net
Network more focuses on learning the partitioning boundary of area-of-interest (RoI).
Step 2-2, after step 2-1 trains full convolutional network U-net, data set 8000 opens image and uses U-net
It carries out image denoising and extracts area-of-interest, Fig. 6 is provided two anticipatory remarks invention and handled using trained full convolutional network U-net
The result figure obtained after original OCT image.
Step 3: being positioned and identified in cardiovascular OCT image with Faster R-CNN target detection frame and easily lose patch.
Referring to Fig. 7, Fig. 7 is the target detection frame Faster R-CNN flow chart that the present invention uses.FasterR-CNN is calculated
Method is made of two big modules: (1) network (Region Proposal Network, RPN) is suggested in region, (2) Fast R-CNN inspection
Survey grid network.
Step 3-1: data preparation first is carried out for the training of Faster R-CNN.The input picture of Faster R-CNN network
The result figure obtained from second stage.As shown in figure 8, the corresponding patch region marked by specialist of each figure
Coordinate position label.Specifically, be illustrated to each data concrete meaning of label: first data indicates picture name, the
Two data indicate that the figure belongs to positive sample there are two selectable value (0,1), 1, contain patch region at least one, 0 represents the figure
For negative sample, patch region is free of.Begin from third data are split as the column pixel of the starting and ending in patch region in image.
By taking 0006.png as an example, which is positive sample, contains patch region at two, first patch region starting column pixel is 1, knot
Beam column pixel is 164；Second patch region starting column pixel is 688, and end column pixel is 720.Patch region column pixel is sat
It is denoted as the reference standard (ground truth) for Faster R-CNN network detection block.
Step 3-2: suggesting that network extracts detection zone with region, it can be special with the convolution of entire detection network share full figure
Sign, so that region suggests hardly taking time.Region suggests that network carries out feature extraction first, then generates candidate region, most
Window classification and position refine are carried out afterwards.Feature extraction is carried out using deep learning network of network common on ImageNet, special
Not, the present invention deep learning network ResNet101 best using current performance.Candidate region is generated specifically to this
Each position of image considers that x kind ratio, y kind size amount to the candidate window of x*y size.Particularly, the present invention takes
X=3, y=3, taking three kinds of sizes is { 1282, 2562, 5122, three kinds of ratios are { 1:1,1:2,2:1 }, carry out window classification and
When the refine of position, classification layer exports the probability window recurrence that 3*3=9 candidate region on each position belongs to foreground and background
The layer output each candidate region in each position, which corresponds to window, should translate the parameter of scaling.
Particularly, it in order to train region to suggest network, needs to distribute class label { target, non-mesh to each candidate region
Mark }.For positive and negative label, target is equal to positive label, non-targeted to be equal to negative label；
The present invention provides following regulation:
Referring to Fig. 9, the multitask loss in Faster R-CNN is described.To the function of an image in FasterR-CNN
Is defined as:
piThe probability of target is predicted as anchor；Ground truth label:
ti={ tx,ty,tw,thIt is a vector, indicate that 4 parametrizations of the bounding box bounding box of prediction are sat
Mark；It is the coordinate vector of ground truth bounding box corresponding with positive anchor；It is two classes
Other logarithm loss: It is
Loss is returned, is usedIt calculates, R is smooth L1 function.This means only
There is prospectJust there is recurrence to lose, other situations just do not have.Cls layers and reg layers of output is respectively by { pi}
{ uiComposition, this two respectively by NclsAnd NregAn and balance weight λ normalization.
Bounding box in Faster R-CNN is returned:
tx=(x-xa)/wa,ty=(y-ya)/ha,
tw=log (w/wa),th=log (h/ha),
Wherein, x, y, w, h are centre coordinate, width, the height of box:
Referring to Fig.1 0 (a), since feature extraction network extracts large number of candidate frame, have between each candidate frame very much
Overlapping region realizes that detection block merges and deletion using non-maxima suppression method (NMS, non-maximum suppression).
Specifically, the operation of non-maxima suppression method is as follows: the score of all candidate frames being sorted, best result and its corresponding frame are chosen.
The candidate frame for being greater than 0.7 (IoU > 0.7) with the registration of top score candidate frame is all deleted, this top score time is only left
Frame is selected, as shown in Figure 10 (b).Similarly, settable by registration about certain threshold value x or probability about certain threshold at final test side
The prediction block of value P is merged using non-maxima suppression method.Non-maxima suppression method will not influence final Detection accuracy,
But significantly reduce the quantity of Suggestion box.
After region suggests that network (RPN) extraction obtains candidate region, final detection is realized using Fast R-CNN module
And identification, the convolutional layer that network is suggested in region and Fast R-CNN shares ResNet101 is made by training.Specific training process
It is as follows:
Suggest the Suggestion box that network generates using region, one is individually detected network, the inspection by Fast R-CNN training
Survey grid network is initialized by ResNet101, and region is trained to suggest that network, fixed shared convolutional layer only finely tune area first
The exclusive layer of network is suggested in domain, and shared convolutional layer is then kept to fix, and finely tunes the full articulamentum of Fast R-CNN.In this way, two
The convolutional layer of a network share ResNet101 constitutes a unified network.
1 (a) and Figure 11 (b) referring to Fig.1, provides two the present embodiment testing result figures, it can be seen that result figure is clearly marked
It is good to outpour the easily position of loss patch and probability, detection effect.
Referring to Fig.1 2, provide the framework of the easy loss plaque detection device based on deep learning.The detection device is by OCT shadow
As obtaining module, digital signal processing module, data memory module and testing result display module totally four module compositions.
OCT image acquiring module includes three units: probe unit, probe interface unit and optical signal processing unit, OCT
Image acquiring module is used to obtain the image of patient vessel's inner wall.
Digital signal processing module: the digital signal processing module is used to detect the easy loss in the OCT image obtained
Patch.OCT image to be detected is handled by above-mentioned second stage and the deep learning method of phase III, original image is gone in realization
It makes an uproar, extract area-of-interest and detect the position and the probability that easily lose patch.
Data memory module: the memory module is used to store the OCT image and OCT of OCT image acquiring module acquisition
The testing result of image.
Testing result display module: the testing result display module is for showing that patient OCT image easily loses patch
Testing result.
The present embodiment carries out data augmentation to the OCT image obtained from OCT equipment, devises the calculation of two stages deep learning
Method, and the easy loss plaque location label of specialist mark is combined, so that trained deep learning network exact divides OCT
Equipment shooting Cardiovascular imaging detects whether to will test containing easy loss patch and its position and probability, last detection system
Data storage simultaneously detects output result.This embodiment achieves the deep learning of diagnosis process whole process, accuracy in detection is high, fast
Degree is fast, and an OCT image is detected on GPU server averagely only needs 200ms, is truly realized real-time detection, has preferable
Clinical value.
Claims (6)
1. the cardiovascular OCT image based on deep learning easily loses plaque detection method, which is characterized in that comprise the following steps that
Step 1: obtaining angiocarpy OCT image from OCT image documentation equipment, data augmentation is carried out to image data set；
Step 1-1 obtains the data set being made of m OCT angiocardiograms, including the positive sample containing easy loss patch
This and the negative sample without easy loss patch, positive sample quantity are greater than negative sample quantity, and all positive samples are by specialist mark
Outpour the specific location of easily loss patch；
Step 1-2 carries out data augmentation to the data set of m OCT angiocardiogram compositions, to obtain 4m~10m samples；
The data augmentation mode includes the part expansion that appropriateness is carried out to original image, contraction, torsion, corolla and the distortion of flag shape；
Step 2: being denoised and extracted area-of-interest to cardiovascular OCT image with full convolutional neural networks；
Data set is divided into training set, verifying collection and test set, training set and collects each image corresponding one with verifying by step 2-1
Label figure is opened, the label figure is to remove background, conduit redundance, only retains interested region (RoI)；Test set does not have
Label figure；By training set, the verifying collection input full convolutional network training, with the performance of test set test network；
After the full convolutional neural networks of step 2-2, step 2-1 show well on test set, all images of data set are made
Area-of-interest is denoised and extracted with the full convolutional neural networks；
Step 3: being positioned and identified in cardiovascular OCT image with Faster R-CNN target detection frame and easily lose patch；
Training dataset image passes through full convolutional network processing described in step 2, the corresponding spot marked by specialist of every figure
Block area coordinate location information；FasterR-CNN algorithm is made of two big modules: (1) network (Region is suggested in region
Proposal Network, RPN), (2) Fast R-CNN detects network；
Step 3-1 suggests that network extracts detection zone using region, and region is suggested that network carries out feature extraction first, then given birth to
At candidate region (anchor), window classification and position refine are finally carried out；
Step 3-2 suggests network for subsequent trained region, gives each candidate region distribution class label { target, non-targeted }；To area
Suggest that the label setting in network candidates region is described as follows in domain: target is equal to positive label, non-targeted to be equal to negative label；
Positive label: with the registration of any reference standard (ground truth) bounding box (bounding box) be greater than x (IoU >
X) candidate region；
Negative label: it is less than the candidate region of (1-x) with the registration of all reference standard bounding boxs；
Realize that detection block merges and deletion using non-maxima suppression method；Specifically, the score of all candidate frames is sorted, is chosen
Best result and its corresponding frame；Candidate frame with the registration of top score candidate frame greater than x (IoU > x) is all deleted, only
It is left this top score candidate frame；After non-maxima suppression method, candidate region is ranked up, the top n after sorting is taken to examine
It surveys；
Step 3-3 is realized finally after region suggests that network (RPN) extraction obtains candidate region using Fast R-CNN module
Detection and identification make the convolutional layer that network is suggested in region and Fast R-CNN shares ResNet101 by training；
Step 4: system stores detection data, output test result；
To each OCT image of OCT image documentation equipment acquisition, mark can be obtained after step 2, the processing of 3 deep learning methods
Outpour the output result that plaque region domain and its probability are easily lost on image.
2. the cardiovascular OCT image according to claim 1 based on deep learning easily loses plaque detection method, feature
It is, the full convolutional network model framework that the step 2 uses: constricted path and right side of the full convolutional neural networks by left side
Extensions path composition, it is U-shaped integral into one；To input picture carry out n times down-sampling constitute left side constricted path, to image into
Row n times up-sampling constitutes right extension path；Wherein, the volume that constricted path is 3*3 by the convolution kernel size of two repeated applications
Lamination composition, each convolutional layer follow an amendment linear unit (rectified linear unit, ReLU), every convolution A
A down-sampling (down sampling), the maximum pond layer (max for the 2*2 that down-sampling operation is 2 by step-length are carried out after secondary
Pooling it) completes, each down-sampling step doubles feature number of active lanes；Two convolution kernels of right extension path repeated application
Size is the convolutional layer of 3*3, each follows an amendment linear unit, is once up-sampled after every convolution A times, above adopted
Sample (up sampling) halves the quantity in feature channel；The convolution kernel size of right extension path the last layer is 1*1, will
Characteristic pattern be converted into certain depth as a result, if the pixel of an image is divided into B classification, certain depth is equal to B.
3. the cardiovascular OCT image according to claim 1 or 2 based on deep learning easily loses plaque detection method,
It is characterized in that, feature extraction is using deep learning network ResNet101 in step 3-1；Generating candidate region is to the figure
Each position of picture considers that x kind ratio, y kind size amount to the candidate window of x*y size；Carry out window classification and position
When setting refine, classification layer exports the probability that x*y candidate region on each position belongs to foreground and background；It is defeated that window returns layer
The each candidate region in each position, which corresponds to window, out should translate the parameter of scaling.
4. the cardiovascular OCT image according to claim 1 or 2 based on deep learning easily loses plaque detection method,
It is characterized in that, step 3-3, specific training process is as follows: suggesting the Suggestion box that network generates using region, by Fast R-CNN
Training one individually detects network, which initialized by ResNet101, trains region to suggest net first
Network, fixed shared convolutional layer only finely tune region and suggest the exclusive layer of network, then keep shared convolutional layer to fix, finely tune
The full articulamentum of Fast R-CNN；The identical convolutional layer of two network shares constitutes a unified network.
5. the cardiovascular OCT image according to claim 3 based on deep learning easily loses plaque detection method, feature
It is, step 3-3, specific training process is as follows: suggesting the Suggestion box that network generates using region, is trained by Fast R-CNN
One is individually detected network, which initialized by ResNet101, trains region to suggest network first, Gu
Surely shared convolutional layer only finely tunes region and suggests the exclusive layer of network, then keeps shared convolutional layer to fix, finely tune Fast
The full articulamentum of R-CNN；The identical convolutional layer of two network shares constitutes a unified network.
6. using the detection system of any detection method of claim 1-5, which is characterized in that including sequentially connected OCT
Image capturing unit, digital signal processing unit, data storage cell, testing result display unit；
OCT image capturing unit: image is transferred in system by connection OCT image documentation equipment and this system from OCT image documentation equipment；
Digital signal processing unit: implementing entire detection algorithm, extracts including (1) to the original image denoising that image documentation equipment is shot
Easy loss patch in area-of-interest and (2) positioning, identification OCT image；
Data storage cell: the original image of result and the image documentation equipment shooting to OCT Image detection is stored；
Testing result display unit: output test result, result figure shows the position that patch is easily lost in OCT image and it is general
Rate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810678850.5A CN108961229A (en) | 2018-06-27 | 2018-06-27 | Cardiovascular OCT image based on deep learning easily loses plaque detection method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810678850.5A CN108961229A (en) | 2018-06-27 | 2018-06-27 | Cardiovascular OCT image based on deep learning easily loses plaque detection method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108961229A true CN108961229A (en) | 2018-12-07 |
Family
ID=64487159
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810678850.5A Pending CN108961229A (en) | 2018-06-27 | 2018-06-27 | Cardiovascular OCT image based on deep learning easily loses plaque detection method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108961229A (en) |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109685776A (en) * | 2018-12-12 | 2019-04-26 | 华中科技大学 | A kind of pulmonary nodule detection method based on ct images and system |
| CN109712140A (en) * | 2019-01-02 | 2019-05-03 | 中楹青创科技有限公司 | Method and device of the training for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection |
| CN109829892A (en) * | 2019-01-03 | 2019-05-31 | 众安信息技术服务有限公司 | A kind of training method of prediction model, prediction technique and device using the model |
| 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 |
| CN109886245A (en) * | 2019-03-02 | 2019-06-14 | 山东大学 | A kind of pedestrian detection recognition methods based on deep learning cascade neural network |
| CN110136115A (en) * | 2019-05-14 | 2019-08-16 | 重庆大学 | IVOCT image vulnerable plaque detects Approach For Neural Network Ensemble automatically |
| CN110136107A (en) * | 2019-05-07 | 2019-08-16 | 上海交通大学 | Based on DSSD and time-domain constraints X-ray coronary angiography sequence automatic analysis method |
| CN110148111A (en) * | 2019-04-01 | 2019-08-20 | 江西比格威医疗科技有限公司 | The automatic testing method of a variety of retina lesions in a kind of retina OCT image |
| CN110189292A (en) * | 2019-04-15 | 2019-08-30 | 浙江工业大学 | A kind of cancer cell detection method based on Faster R-CNN and density estimation |
| CN110223280A (en) * | 2019-06-03 | 2019-09-10 | Oppo广东移动通信有限公司 | Phlebothrombosis detection method and phlebothrombosis detection device |
| CN110222759A (en) * | 2019-06-03 | 2019-09-10 | 中国医科大学附属第一医院 | A kind of Plaque Vulnerability in Coronary Artery automatic recognition system |
| CN110223781A (en) * | 2019-06-03 | 2019-09-10 | 中国医科大学附属第一医院 | A kind of various dimensions plaque rupture Warning System |
| CN110321844A (en) * | 2019-07-04 | 2019-10-11 | 北京万里红科技股份有限公司 | A kind of quick iris detection method based on convolutional neural networks |
| CN110706246A (en) * | 2019-10-15 | 2020-01-17 | 上海微创医疗器械（集团）有限公司 | Blood vessel image segmentation method and device, electronic equipment and storage medium |
| CN111160479A (en) * | 2019-12-31 | 2020-05-15 | 华南理工大学 | Welding spot identification method |
| CN111341420A (en) * | 2020-02-21 | 2020-06-26 | 四川大学 | Cardiovascular image recognition system and method based on whole-heart seven-dimensional model |
| CN111507335A (en) * | 2019-01-31 | 2020-08-07 | 斯特拉德视觉公司 | Method and device for automatically labeling training images for deep learning network |
| WO2020222004A1 (en) | 2019-05-01 | 2020-11-05 | Cambridge Enterprise Limited | Method and apparatus for analysing intracoronary images |
| CN112767350A (en) * | 2021-01-19 | 2021-05-07 | 深圳麦科田生物医疗技术股份有限公司 | Method, device, equipment and storage medium for predicting maximum interval of thromboelastogram |
| CN112767375A (en) * | 2021-01-27 | 2021-05-07 | 深圳技术大学 | OCT image classification method, system and equipment based on computer vision characteristics |
| CN112819705A (en) * | 2021-01-13 | 2021-05-18 | 西安交通大学 | Real image denoising method based on mesh structure and long-distance correlation |
| CN112861849A (en) * | 2020-12-29 | 2021-05-28 | 北京航空航天大学 | Tissue identification method in spinal deformity correction surgery |
| CN112950546A (en) * | 2021-02-03 | 2021-06-11 | 中南民族大学 | Esophagus cancer detection method and system of barium meal radiography image |
| CN113256531A (en) * | 2021-06-11 | 2021-08-13 | 云南电网有限责任公司电力科学研究院 | Noise reduction method for hyperspectral image of power equipment |
| CN114140453A (en) * | 2021-12-08 | 2022-03-04 | 广州医科大学附属第一医院（广州呼吸中心） | Neural network-based silhouette angiography segmentation method |
| CN114693622A (en) * | 2022-03-22 | 2022-07-01 | 电子科技大学 | Plaque erosion automatic detection system based on artificial intelligence |
| CN114757847A (en) * | 2022-04-24 | 2022-07-15 | 汕头市超声仪器研究所股份有限公司 | Multi-information extraction extended U-Net and application method thereof in low-dose X-ray imaging |
| WO2023284055A1 (en) * | 2021-07-13 | 2023-01-19 | 深圳市中科微光医疗器械技术有限公司 | Method and device for calculating ipa of intraluminal oct image |
| CN112767350B (en) * | 2021-01-19 | 2024-04-26 | 深圳麦科田生物医疗技术股份有限公司 | Method, device, equipment and storage medium for predicting maximum section of thromboelastography |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107507201A (en) * | 2017-09-22 | 2017-12-22 | 深圳天琴医疗科技有限公司 | A kind of medical image cutting method and device |
| CN107993221A (en) * | 2017-11-16 | 2018-05-04 | 济南大学 | cardiovascular optical coherence tomography OCT image vulnerable plaque automatic identifying method |
| US20190325605A1 (en) * | 2016-12-29 | 2019-10-24 | Zhejiang Dahua Technology Co., Ltd. | Systems and methods for detecting objects in images |
-
2018
- 2018-06-27 CN CN201810678850.5A patent/CN108961229A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190325605A1 (en) * | 2016-12-29 | 2019-10-24 | Zhejiang Dahua Technology Co., Ltd. | Systems and methods for detecting objects in images |
| CN107507201A (en) * | 2017-09-22 | 2017-12-22 | 深圳天琴医疗科技有限公司 | A kind of medical image cutting method and device |
| CN107993221A (en) * | 2017-11-16 | 2018-05-04 | 济南大学 | cardiovascular optical coherence tomography OCT image vulnerable plaque automatic identifying method |
Non-Patent Citations (1)
| Title |
|---|
| 王黎 等: "基于深度学习的心血管OCT影像易损失斑块检测方法", 《中国计量大学学报》 * |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109685776B (en) * | 2018-12-12 | 2021-01-19 | 华中科技大学 | Pulmonary nodule detection method and system based on CT image |
| CN109685776A (en) * | 2018-12-12 | 2019-04-26 | 华中科技大学 | A kind of pulmonary nodule detection method based on ct images and system |
| CN109712140A (en) * | 2019-01-02 | 2019-05-03 | 中楹青创科技有限公司 | Method and device of the training for the full link sort network of evaporating, emitting, dripping or leaking of liquid or gas detection |
| CN109829892A (en) * | 2019-01-03 | 2019-05-31 | 众安信息技术服务有限公司 | A kind of training method of prediction model, prediction technique and device using the model |
| CN111507335B (en) * | 2019-01-31 | 2023-10-13 | 斯特拉德视觉公司 | Method and device for automatically labeling training images used for deep learning network |
| CN111507335A (en) * | 2019-01-31 | 2020-08-07 | 斯特拉德视觉公司 | Method and device for automatically labeling training images for deep learning network |
| 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 |
| CN109886245A (en) * | 2019-03-02 | 2019-06-14 | 山东大学 | A kind of pedestrian detection recognition methods based on deep learning cascade neural network |
| CN110148111A (en) * | 2019-04-01 | 2019-08-20 | 江西比格威医疗科技有限公司 | The automatic testing method of a variety of retina lesions in a kind of retina OCT image |
| CN110189292A (en) * | 2019-04-15 | 2019-08-30 | 浙江工业大学 | A kind of cancer cell detection method based on Faster R-CNN and density estimation |
| WO2020222004A1 (en) | 2019-05-01 | 2020-11-05 | Cambridge Enterprise Limited | Method and apparatus for analysing intracoronary images |
| CN110136107A (en) * | 2019-05-07 | 2019-08-16 | 上海交通大学 | Based on DSSD and time-domain constraints X-ray coronary angiography sequence automatic analysis method |
| CN110136107B (en) * | 2019-05-07 | 2023-09-05 | 上海交通大学 | Automatic analysis method based on DSSD and time domain constraint X-ray coronary angiography sequence |
| CN110136115A (en) * | 2019-05-14 | 2019-08-16 | 重庆大学 | IVOCT image vulnerable plaque detects Approach For Neural Network Ensemble automatically |
| CN110223280B (en) * | 2019-06-03 | 2021-04-13 | Oppo广东移动通信有限公司 | Venous thrombosis detection method and venous thrombosis detection device |
| CN110222759A (en) * | 2019-06-03 | 2019-09-10 | 中国医科大学附属第一医院 | A kind of Plaque Vulnerability in Coronary Artery automatic recognition system |
| CN110222759B (en) * | 2019-06-03 | 2021-03-30 | 中国医科大学附属第一医院 | Automatic identification system for vulnerable plaque of coronary artery |
| CN110223781A (en) * | 2019-06-03 | 2019-09-10 | 中国医科大学附属第一医院 | A kind of various dimensions plaque rupture Warning System |
| CN110223280A (en) * | 2019-06-03 | 2019-09-10 | Oppo广东移动通信有限公司 | Phlebothrombosis detection method and phlebothrombosis detection device |
| CN110321844A (en) * | 2019-07-04 | 2019-10-11 | 北京万里红科技股份有限公司 | A kind of quick iris detection method based on convolutional neural networks |
| CN110321844B (en) * | 2019-07-04 | 2021-09-03 | 北京万里红科技股份有限公司 | Fast iris detection method based on convolutional neural network |
| CN110706246A (en) * | 2019-10-15 | 2020-01-17 | 上海微创医疗器械（集团）有限公司 | Blood vessel image segmentation method and device, electronic equipment and storage medium |
| CN111160479A (en) * | 2019-12-31 | 2020-05-15 | 华南理工大学 | Welding spot identification method |
| CN111341420A (en) * | 2020-02-21 | 2020-06-26 | 四川大学 | Cardiovascular image recognition system and method based on whole-heart seven-dimensional model |
| CN111341420B (en) * | 2020-02-21 | 2022-08-30 | 四川大学 | Cardiovascular image recognition system and method based on whole-heart seven-dimensional model |
| CN112861849A (en) * | 2020-12-29 | 2021-05-28 | 北京航空航天大学 | Tissue identification method in spinal deformity correction surgery |
| CN112819705A (en) * | 2021-01-13 | 2021-05-18 | 西安交通大学 | Real image denoising method based on mesh structure and long-distance correlation |
| CN112819705B (en) * | 2021-01-13 | 2023-04-18 | 西安交通大学 | Real image denoising method based on mesh structure and long-distance correlation |
| CN112767350A (en) * | 2021-01-19 | 2021-05-07 | 深圳麦科田生物医疗技术股份有限公司 | Method, device, equipment and storage medium for predicting maximum interval of thromboelastogram |
| CN112767350B (en) * | 2021-01-19 | 2024-04-26 | 深圳麦科田生物医疗技术股份有限公司 | Method, device, equipment and storage medium for predicting maximum section of thromboelastography |
| CN112767375B (en) * | 2021-01-27 | 2022-03-08 | 深圳技术大学 | OCT image classification method, system and equipment based on computer vision characteristics |
| CN112767375A (en) * | 2021-01-27 | 2021-05-07 | 深圳技术大学 | OCT image classification method, system and equipment based on computer vision characteristics |
| CN112950546A (en) * | 2021-02-03 | 2021-06-11 | 中南民族大学 | Esophagus cancer detection method and system of barium meal radiography image |
| CN112950546B (en) * | 2021-02-03 | 2023-10-31 | 中南民族大学 | Method and system for detecting esophagus cancer by barium meal radiography image |
| CN113256531A (en) * | 2021-06-11 | 2021-08-13 | 云南电网有限责任公司电力科学研究院 | Noise reduction method for hyperspectral image of power equipment |
| WO2023284055A1 (en) * | 2021-07-13 | 2023-01-19 | 深圳市中科微光医疗器械技术有限公司 | Method and device for calculating ipa of intraluminal oct image |
| CN114140453A (en) * | 2021-12-08 | 2022-03-04 | 广州医科大学附属第一医院（广州呼吸中心） | Neural network-based silhouette angiography segmentation method |
| CN114140453B (en) * | 2021-12-08 | 2024-04-12 | 广州医科大学附属第一医院（广州呼吸中心） | Neural network-based segmentation method for silhouette angiography |
| CN114693622A (en) * | 2022-03-22 | 2022-07-01 | 电子科技大学 | Plaque erosion automatic detection system based on artificial intelligence |
| CN114693622B (en) * | 2022-03-22 | 2023-04-07 | 电子科技大学 | Plaque erosion automatic detection system based on artificial intelligence |
| CN114757847A (en) * | 2022-04-24 | 2022-07-15 | 汕头市超声仪器研究所股份有限公司 | Multi-information extraction extended U-Net and application method thereof in low-dose X-ray imaging |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108961229A (en) | Cardiovascular OCT image based on deep learning easily loses plaque detection method and system | |
| CN110599448B (en) | Migratory learning lung lesion tissue detection system based on MaskScoring R-CNN network | |
| Amiri et al. | Two-stage ultrasound image segmentation using U-Net and test time augmentation | |
| CN109087327B (en) | Thyroid nodule ultrasonic image segmentation method of cascaded full convolution neural network | |
| CN108171698B (en) | Method for automatically detecting human heart coronary calcified plaque | |
| CN107492071B (en) | Medical image processing method and equipment | |
| CN110428432B (en) | Deep neural network algorithm for automatically segmenting colon gland image | |
| CN107194872B (en) | Remote sensed image super-resolution reconstruction method based on perception of content deep learning network | |
| CN110288597B (en) | Attention mechanism-based wireless capsule endoscope video saliency detection method | |
| CN111539930A (en) | Dynamic ultrasonic breast nodule real-time segmentation and identification method based on deep learning | |
| Morris | A pyramid CNN for dense-leaves segmentation | |
| CN112017198A (en) | Right ventricle segmentation method and device based on self-attention mechanism multi-scale features | |
| CN111445478A (en) | Intracranial aneurysm region automatic detection system and detection method for CTA image | |
| CN110992377B (en) | Image segmentation method, device, computer-readable storage medium and equipment | |
| CN112862824A (en) | Novel coronavirus pneumonia focus detection method, system, device and storage medium | |
| CN111784721A (en) | Ultrasonic endoscopic image intelligent segmentation and quantification method and system based on deep learning | |
| CN112270667B (en) | TI-RADS-based integrated deep learning multi-tag identification method | |
| CN112926652A (en) | Fish fine-grained image identification method based on deep learning | |
| CN110047075A (en) | A kind of CT image partition method based on confrontation network | |
| CN114359629A (en) | Pneumonia X chest radiography classification and identification method based on deep migration learning | |
| CN103169506A (en) | Ultrasonic diagnosis device and method capable of recognizing liver cancer automatically | |
| Kitrungrotsakul et al. | Interactive deep refinement network for medical image segmentation | |
| CN110895815A (en) | Chest X-ray pneumothorax segmentation method based on deep learning | |
| CN104732230A (en) | Pathology image local-feature extracting method based on cell nucleus statistical information | |
| JP2023517058A (en) | Automatic detection of tumors based on image processing |
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: 20181207 |