CN109993757A - A kind of retinal images lesion region automatic division method and system - Google Patents

A kind of retinal images lesion region automatic division method and system Download PDF

Info

Publication number
CN109993757A
CN109993757A CN201910309970.2A CN201910309970A CN109993757A CN 109993757 A CN109993757 A CN 109993757A CN 201910309970 A CN201910309970 A CN 201910309970A CN 109993757 A CN109993757 A CN 109993757A
Authority
CN
China
Prior art keywords
region
csc
retinal images
lesion region
segmentation
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
Application number
CN201910309970.2A
Other languages
Chinese (zh)
Other versions
CN109993757B (en
Inventor
李登旺
王卓
牛四杰
孔问问
吴敬红
薛洁
陈美荣
刘婷婷
黄浦
赵睿
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shandong Normal University
Original Assignee
Shandong Normal University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shandong Normal University filed Critical Shandong Normal University
Priority to CN201910309970.2A priority Critical patent/CN109993757B/en
Publication of CN109993757A publication Critical patent/CN109993757A/en
Application granted granted Critical
Publication of CN109993757B publication Critical patent/CN109993757B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30041—Eye; Retina; Ophthalmic

Abstract

The invention discloses a kind of retinal images lesion region automatic division method and systems, realize while dividing center slurries choroidopathy region and ellipsoid band-Bruch's membrane region.Method includes the following steps: acquisition retinal images, configure training set and test set；Full convolutional neural networks model is configured, loss function is modified, utilizes the full convolutional neural networks model of image data training in training set；Primary segmentation is carried out to image data in test set using trained full convolutional neural networks model, and two step compensation are carried out to primary segmentation result, obtains probability graph；Probability graph is divided again, obtains the region EZ-BM and CSC lesion region；Ellipsoid band-Bruch's membrane region up-and-down boundary that the region EZ-BM is extracted using bianry image edge detection method, obtains the segmentation result of ellipsoid belt Yu Bruch film layer.

Description

A kind of retinal images lesion region automatic division method and system
Technical field
This disclosure relates to image segmentation field, and in particular to a kind of OCT retinal images based on full convolutional neural networks Ellipsoid band and Bruch's membrane region automatic division method and system.
Background technique
Central serous chorioretinopathy (CSC) is a kind of common retinopathy, in Lipid on Young-middle Male In it is especially common so that the secondary neural sensation retina slurries of retinal pigment epithelium (RPE) are detached from.Ellipsoid band often occurs (EZ) slurries between Bruch's membrane (BM) are detached from, and cause ellipsoid band extremely prominent.
Optical coherence tomography (OCT) is a kind of common retinal morphology observational technique, is observation disease variation A kind of good method.The technology has many advantages, such as noninvasive, depth resolution, volume imagery.Inventor has found in R&D process, existing Have and go out segmentation CSC lesion region in OCT retinal images by the way of dividing by hand, but needs a large amount of time, and Artificial experience is relied on, causes segmentation result inaccurate.
Summary of the invention
In order to overcome the above-mentioned deficiencies of the prior art, the OCT view based on full convolutional neural networks that present disclose provides a kind of Nethike embrane image lesion region automatic division method and system are realized while dividing center slurries choroidopathy region and ellipsoid Band-Bruch's membrane region.
A kind of technical solution of retinal images lesion region automatic division method of the first aspect of the disclosure is:
A kind of retinal images lesion region automatic division method, method includes the following steps:
Retinal images are acquired, training set and test set are configured；
Full convolutional neural networks model is configured, loss function is modified, utilizes the full convolution mind of image data training in training set Through network model；
Primary segmentation is carried out to image data in test set using trained full convolutional neural networks model, and to preliminary Segmentation result carries out two step compensation, obtains probability graph；
Threshold segmentation is carried out to probability graph, obtains the region EZ-BM and CSC lesion region；
Ellipsoid band-Bruch's membrane region the up-and-down boundary in the region EZ-BM is extracted using bianry image edge detection method, Obtain the segmentation result of ellipsoid belt Yu Bruch film layer.
A kind of technical solution of automatic segmenting system of retinal images lesion region of the second aspect of the disclosure is:
A kind of automatic segmenting system of retinal images lesion region, the system include:
Image capture module configures training set and test set for acquiring retinal images；
Model training module modifies loss function, utilizes image in training set for configuring full convolutional neural networks model The full convolutional neural networks model of data training；
Image compensation divides module, for utilizing trained full convolutional neural networks model to image data in test set Primary segmentation is carried out, and two step compensation are carried out to primary segmentation result, obtains probability graph；
Carrying out image threshold segmentation module obtains the region EZ-BM and CSC lesion region for carrying out Threshold segmentation to probability graph；
Ellipsoid belt and Bruch film layer divide module, for extracting EZ-BM using bianry image edge detection method The ellipsoid band in region-Bruch's membrane region up-and-down boundary, obtains the segmentation result of ellipsoid belt Yu Bruch film layer.
Through the above technical solutions, the beneficial effect of the disclosure is:
(1) disclosure using the full convolutional neural networks structure of FCN-8S simultaneously divide center slurries choroidopathy region and Ellipsoid band-Bruch's membrane region improves full convolutional neural networks model using the method that two steps compensate and is differing greatly Segmentation accuracy in data set, and Jie Kade range loss function expansion to three is classified.
(2) disclosure uses relatively advanced full convolutional neural networks algorithm, accuracy with higher, and utilizes two steps Compensation method improves the generalization ability of training pattern, and the requirement to hardware is lower, and system cost is low, reusable.
Detailed description of the invention
The Figure of description for constituting a part of this disclosure is used to provide further understanding of the disclosure, and the disclosure is shown Meaning property embodiment and its explanation do not constitute the improper restriction to the disclosure for explaining the application.
Fig. 1 is the flow chart of one retinal images lesion region automatic division method of embodiment；
Fig. 2 is one manual sort result of embodiment and final label figure；
Fig. 3 be embodiment one normally and anomaly classification result and adjacent with the 8 of one, corresponding CSC lesion edge pixel respectively Domain figure；
Fig. 4 is the flow chart for the specific iterative process that embodiment one finds optimum a-value；
Fig. 5 is one retinal images segmentation result figure of embodiment；
Fig. 6 is the structure chart of the two automatic segmenting system of retinal images lesion region of embodiment；
Specific embodiment
The disclosure is described further with embodiment with reference to the accompanying drawing.
It is noted that following detailed description is all illustrative, it is intended to provide further instruction to the disclosure.Unless another It indicates, all technical and scientific terms that the disclosure uses have logical with disclosure person of an ordinary skill in the technical field The identical meanings understood.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
Explanation of nouns:
(1) B-scan image, transversal scanning image.
(2) region CSC, central serous chorioretinopathy region；
(3) region EZ-BM, ellipsoid band-Bruch's membrane region.
Embodiment one
The present embodiment provides a kind of the retinal images lesion region automatic division method based on full convolutional neural networks, benefit Divide two area-of-interests simultaneously with the full convolutional neural networks structure of FCN-8S, proposes the method for two steps compensation, improve mould Classification accuracy of the type in the data set to differ greatly, and Jie Kade range loss function expansion to three is classified；The party Method accuracy with higher, and improve using two step compensation methodes the generalization ability of training pattern, the requirement to hardware compared with Low, system cost is low, reusable.
Please refer to attached drawing 1, the retinal images lesion region automatic division method the following steps are included:
S101 acquires OCT retinal images, and classifies to it.
Specifically, in order to provide reliably trained classification standard to neural network, include in acquisition retinopathy eyes 757 retina B-scan images of lesion region, using ImageJ software by retina B-scan image segmentation at the area EZ-BM Domain and CSC lesion region, please refer to attached drawing 2, and (b) is the region EZ-BM in Fig. 2, (c) are the region CSC.Then, two regions are folded Add integration, with season background area as the first kind, removes the region EZ-BM of CSC lesion as the second class, CSC lesion region As third class, in Fig. 2 shown in (d).
In the present embodiment, classify to collected retina B-scan image, can be provided to neural network reliable Trained segmentation standard.
S102 configures training dataset and test set.
Specifically, randomly select all retina B-scan images 70% is used as neural metwork training collection；It will be remaining 30% test set as neural network of retina B-scan image.
The present embodiment configures training dataset and test set, for full convolutional neural networks model training and test.
S103 configures full convolutional neural networks model, modifies loss function, inputs the full volume of image pattern training in training set Product neural network model.
Specifically, the full convolutional neural networks model trained uses the FCN-8s network structure based on VGG-16, by 13 convolutional layers, 5 maximum pond layers, activation primitive layer, 3 warp laminations, two dropout layers of compositions, finally by classifying Layer is classified.
Improved loss function are as follows:
L=JEZ-BM+JCSC
Wherein, JEZ-BMFor cross entropy loss function, expression formula are as follows:
JCSCFor refined card range loss function, expression formula are as follows:
Wherein, y2Belong to the true distribution of the second class, y for retinal images pixel3Belong to third for retinal images pixel The true distribution of class, if retinal images pixel belongs to the second class, y2=1, it is not belonging to the second class then y2=0.p2+p3To belong to Probability in the region EZ-BM, p3For the probability for belonging to CSC lesion region.
Image data in test set is carried out primary segmentation to trained full convolutional neural networks, obtained automatic by S104 Divide binary result, and two step compensation are carried out to segmentation binary result, obtains EZ-BM area probability figure and the lesion CSC probability graph.
Specifically, using p2+p3> 0.5 is as the criterion for extracting the region EZ-BM, p3> 0.5 is sick as CSC is extracted The criterion for becoming region, using trained full convolutional neural networks to retinal images primary segmentation, obtain it is preliminary from Dynamic segmentation binary result, including EZ-BM binary map and CSC binary map.
Specifically, in the step 104, the specific implementation of two step compensation is carried out to segmentation result are as follows:
For the abnormal external plexiform layer detection error being free in outside the region EZ-BM, by only retaining in EZ-BM binary map The method in maximum area region is removed, and EZ-BM area probability figure is obtained.
Classification error caused by too low for CSC lesion region contrast, by by retina B-scan in test set The method that original image pixel value is greater than 1 coefficient multiplied by one improves image data contrast in test set, using trained complete Convolutional neural networks are again split image data in the test set after raising contrast, obtain the best lesion CSC probability Figure, pixel value range 0-255:
F (i, j)=α B (i, j)
Wherein, F is retinal images in the test set after improving contrast, and B is original retinal images in test set, α is the constant more than or equal to 1.
Attached drawing 3 is please referred to, normally and anomaly classification result and one, CSC lesion edge corresponding with them respectively is illustrated 8 Neighborhood Graphs of pixel.The FCN three classes output that the present embodiment defines is background respectively, the region EZ-BM in addition to CSC lesion, CSC lesion region.Wherein CSC lesion region this kind is surrounded by the second class.Belong to first kind picture in segmentation result The value of element is 0, and the value of the second class pixel is 1, and the value of third class pixel is 2.So CSC lesion region is every in correct classification The minimum value of one 8 neighborhood of edge pixel is all 1.But the classification error as caused by contrast can make the second class excalation Or third class breaks through the second class and extends out, this will make the minimum of a part of 8 neighborhood of edge pixel of CSC lesion region Value is 0.
It is found by the method for maximizing the mean value of each 8 neighborhood minimum of pixel of CSC lesion region edge come iteration Most suitable α value, the testing classification since α=1 is as a result, if the mean value of 8 neighborhood minimum of CSC edges of regions is equal to 1, at this time α is optimum value；If the mean value of 8 neighborhood minimum of the edge CSC increases by 0.1 less than 1, α and continues to test the upper limit until α=3, Compare all classification results, then takes the α value of the mean value maximum of 8 neighborhood minimum of the edge CSC as optimum value.
Attached drawing 4 is please referred to, is found by maximizing the method for mean value of CSC lesion region edge neighborhood minimum value come iteration Most suitable α value, the specific implementation process is as follows:
(1) α=1 is enabled to start testing classification result；
(2) the retinal images F after raising contrast trained full convolutional neural networks are inputted to classify；
(3) judge whether the mean value of each 8 neighborhood minimum of pixel of CSC lesion region edge is equal to 1；
The mean value of each 8 neighborhood minimum of pixel of CSC lesion region edge is by obtaining CSC lesion region side Minimum value in 8 neighborhoods of each pixel of edge, then takes mean value to obtain.
(4) if each 8 neighborhood minimum of pixel of CSC lesion region edge is equal to 1, α and corresponding classification are exported As a result, terminating iteration；If the mean value of each 8 neighborhood minimum of pixel of CSC lesion region edge is not equal to 1, then follow the steps (5)；
(5) judge whether α reaches the upper limit of α=3, if reaching the upper limit, find each pixel of CSC lesion region edge α corresponding to the mean value of 8 neighborhood minimums and classification results, and α and corresponding classification results are exported, terminate iteration；If α does not have There is the upper limit of arrival α=3, then so that α=α+0.1 is then return to step (2) and continue iteration.
S105 carries out Threshold segmentation to EZ-BM area probability figure and the lesion CSC probability graph, obtains the region EZ-BM and CSC Lesion region.
S106 extracts ellipsoid band-Bruch's membrane region up-and-down boundary as ellipsoid belt and cloth from the region EZ-BM The segmentation result of the conspicuous film layer in Shandong.
Specifically, ellipsoid is extracted using bianry image edge detection method from the region EZ-BM that step 105 obtains Band-Bruch's membrane region up-and-down boundary, using obtained ellipsoid band-Bruch's membrane region up-and-down boundary as ellipsoid belt EZ With the segmentation result of Bruch film layer BM.
Please refer to attached drawing 5, it is shown that the layer segmentation result of CSC lesion segmentation and EZ and BM that the present embodiment obtains, wherein (a), (e) are original retina b scan image in Fig. 5；(b), (f) is EZ-BM area probability figure；(c), (g) is that the lesion CSC is general Rate figure；(d), (h) is segmentation result, including the lesion CSC, EZ layer segmentation result, BM layers of segmentation result.
Retinal images lesion region automatic division method provided in this embodiment based on full convolutional neural networks is realized Center slurries choroidopathy region and the automatic segmentation in ellipsoid band-Bruch's membrane region utilize the full convolution mind of FCN-8S Divide center slurries choroidopathy region and ellipsoid band-Bruch's membrane region simultaneously through network structure, proposes two steps benefit The method repaid improves segmentation accuracy of the full convolutional neural networks model in the data set to differ greatly, and by Jie Kade Range loss function expansion is classified to three.This method accuracy with higher, and training is improved using two step compensation methodes The generalization ability of model, the requirement to hardware is lower, and system cost is low, reusable.
Embodiment two
The present embodiment provides a kind of automatic segmenting system of retinal images lesion region based on full convolutional neural networks leads to Cross image compensation segmentation module simultaneously divide two area-of-interests, using two steps compensation method, improve model difference compared with Classification accuracy in big data set, and by Jie Kade range loss function expansion to three classification, requirement to hardware compared with Low, system cost is low, reusable.
Attached drawing 6 is please referred to, the automatic segmenting system of retinal images lesion region includes image capture module, classification mould Block, model training module, image compensation segmentation module, carrying out image threshold segmentation module and ellipsoid belt and Bruch film layer are divided Module.
Specifically, described image acquisition module, for acquiring 757 views in retinopathy eyes comprising lesion region Nethike embrane B-scan image, randomly select all retina B-scan images 70% are used as neural metwork training collection；It will be remaining 30% test set as neural network of retina B-scan image.
The categorization module, for using ImageJ software by retina B-scan image segmentation at the region EZ-BM and CSC Then two regions are superimposed and integrate by lesion region, with season background area as the first kind, remove the area EZ-BM of CSC lesion Domain is as the second class, and CSC lesion region is as third class.
The model training module is modified loss function, is inputted in training set for configuring full convolutional neural networks model The full convolutional neural networks model of image pattern training.
Specifically, the full convolutional neural networks model trained uses the FCN-8s network structure based on VGG-16, by 13 convolutional layers, 5 maximum pond layers, activation primitive layer, 3 warp laminations, two dropout layers of compositions, finally by classifying Layer is classified.
Improved loss function are as follows:
L=JEZ-BM+JCSC
Wherein, JEZ-BMFor cross entropy loss function, expression formula are as follows:
JCSCFor refined card range loss function, expression formula are as follows:
Wherein, y2Belong to the true distribution of the second class, y for retinal images pixel3Belong to third for retinal images pixel The true distribution of class, if retinal images pixel belongs to the second class, y2=1, it is not belonging to the second class then y2=0.p2+p3To belong to Probability in the region EZ-BM, p3For the probability for belonging to CSC lesion region.
Described image compensation segmentation module, for by image data in test set to trained full convolutional neural networks into Row primary segmentation obtains dividing binary result automatically, and carries out two step compensation to segmentation binary result, and it is general to obtain the region EZ-BM Rate figure and the lesion CSC probability graph.
In the present embodiment, described image compensation segmentation module includes image primary segmentation module, the compensation mould of image first The second compensating module of block and image；Wherein:
Described image primary segmentation module, for using p2+p3> 0.5 is as the criterion for extracting the region EZ-BM, p3 > 0.5 is preliminary to retinal images using trained full convolutional neural networks as the criterion for extracting CSC lesion region Segmentation, obtains preliminary automatic segmentation binary result, including EZ-BM binary map and CSC binary map.
The first compensating module of described image, for being missed for the abnormal external plexiform layer detection being free in outside the region EZ-BM Difference, the method by only retaining maximum area region in EZ-BM binary map are removed, and EZ-BM area probability figure is obtained.
The second compensating module of described image leads to for classification error caused by too low for CSC lesion region contrast It crosses and retina B-can image pixel value in test set is greater than the method for 1 coefficient multiplied by one to improve retina in test set B-can picture contrast, using trained full convolutional neural networks again to picture number in the test set after raising contrast According to being split, the best lesion CSC probability graph is obtained, pixel value range 0-255:
F (i, j)=α B (i, j)
Wherein, F is retinal images in the test set after improving contrast, and B is original retinal images in test set, α is the constant more than or equal to 1.
Described image Threshold segmentation module, for carrying out threshold value point to EZ-BM area probability figure and the lesion CSC probability graph It cuts, obtains the region EZ-BM and CSC lesion region.
The ellipsoid belt and Bruch film layer divide module, for using bianry image edge detection method from EZ- Segmentation knot of the ellipsoid band-Bruch's membrane region up-and-down boundary as ellipsoid belt and Bruch film layer is extracted in the region BM Fruit.
Retinal images lesion region automatic segmenting system provided in this embodiment based on full convolutional neural networks is realized Center slurries choroidopathy region and the automatic segmentation in ellipsoid band-Bruch's membrane region divide mould using image compensation Block divides center slurries choroidopathy region and ellipsoid band-Bruch's membrane region simultaneously, proposes the method for two steps compensation Improve segmentation accuracy of the full convolutional neural networks model in the data set to differ greatly, and by Jie Kade range loss Function expansion is classified to three.System accuracy with higher, and the general of training pattern is improved using two step compensation methodes Change ability, the requirement to hardware is lower, and system cost is low, reusable.
Although above-mentioned be described in conjunction with specific embodiment of the attached drawing to the disclosure, model not is protected to the disclosure The limitation enclosed, those skilled in the art should understand that, on the basis of the technical solution of the disclosure, those skilled in the art are not Need to make the creative labor the various modifications or changes that can be made still within the protection scope of the disclosure.

Claims (10)

1. a kind of retinal images lesion region automatic division method, characterized in that the following steps are included:
Retinal images are acquired, training set and test set are configured；
Full convolutional neural networks model is configured, loss function is modified, utilizes the full convolutional Neural net of image data training in training set Network model；
Primary segmentation is carried out to image data in test set using trained full convolutional neural networks model, and to primary segmentation As a result two step compensation are carried out, probability graph is obtained；
Threshold segmentation is carried out to probability graph, obtains the region EZ-BM and CSC lesion region；
Ellipsoid band-Bruch's membrane region up-and-down boundary that the region EZ-BM is extracted using bianry image edge detection method, is obtained The segmentation result of ellipsoid belt and Bruch film layer.
2. retinal images lesion region automatic division method according to claim 1, characterized in that further include:
After collecting retinal images, retinal images are divided into the region EZ-BM and CSC lesion region, and two regions are folded Add integration, with season background area as the first kind, removes the region EZ-BM of CSC lesion region as the second class, CSC lesion Region is used as full convolutional neural networks model segmentation standard as third class, by these three types.
3. retinal images lesion region automatic division method according to claim 1, characterized in that the full convolution mind The FCN-8s network structure based on VGG-16 is used through network model.
4. retinal images lesion region automatic division method according to claim 1, characterized in that described to test set The method that middle image data carries out primary segmentation are as follows:
It is greater than 0.5 as the criterion for extracting the region EZ-BM, retinal map using the probability that retinal images belong to the second class Probability as belonging to third class is greater than 0.5 as the criterion for extracting CSC lesion region, utilizes trained full convolutional Neural Network obtains preliminary automatic segmentation binary result, including EZ-BM binary map and CSC two-value to retinal images primary segmentation Figure.
5. retinal images lesion region automatic division method according to claim 4, characterized in that described to preliminary point Cut the method that result carries out two step compensation are as follows:
By retaining maximum area region in EZ-BM binary map, removal is located at the abnormal external plexiform layer inspection outside the region EZ-BM Error is surveyed, EZ-BM area probability figure is obtained；
Optimum a-value is found using the method iteration for the mean value for maximizing CSC lesion region edge neighborhood minimum value, it will be in test set The pixel value of retinal images is multiplied with obtained α value, obtains new retinal images, utilizes trained full convolutional Neural net Network model is again split image data new in test set, obtains the best lesion CSC probability graph.
6. retinal images lesion region automatic division method according to claim 5, characterized in that described using maximum Change CSC lesion region edge neighborhood minimum value mean value method iteration find optimum a-value the step of include:
Judge whether the mean value of each 8 neighborhood minimum of pixel of CSC binary map edge is equal to 1；
If each 8 neighborhood minimum of pixel of CSC binary map edge is equal to 1, α value and corresponding CSC binary map are exported, is tied Beam iteration；
If the mean value of each 8 neighborhood minimum of pixel of CSC binary map edge is not equal to 1, judge whether α reaches the upper of α=3 Limit；
If reaching the upper limit, α and classification corresponding to the mean value of each 8 neighborhood minimum of pixel of CSC binary map edge are found As a result, and exporting α and corresponding CSC binary map, end iteration；
If α does not reach the upper limit of α=3, α=α+0.1 is made to restart to continue iteration.
7. a kind of automatic segmenting system of retinal images lesion region, characterized in that the system includes:
Image capture module configures training set and test set for acquiring retinal images；
Model training module modifies loss function, utilizes image data in training set for configuring full convolutional neural networks model The full convolutional neural networks model of training；
Image compensation divides module, for being carried out using trained full convolutional neural networks model to image data in test set Primary segmentation, and two step compensation are carried out to primary segmentation result, obtain probability graph；
Carrying out image threshold segmentation module obtains the region EZ-BM and CSC lesion region for carrying out Threshold segmentation to probability graph；
Ellipsoid belt and Bruch film layer divide module, for extracting the region EZ-BM using bianry image edge detection method Ellipsoid band-Bruch's membrane region up-and-down boundary, obtain the segmentation result of ellipsoid belt Yu Bruch film layer.
8. the automatic segmenting system of retinal images lesion region according to claim 7, characterized in that further include classification mould Block, the categorization module are used for:
After collecting retinal images, retinal images are divided into the region EZ-BM and CSC lesion region, and two regions are folded Add integration, with season background area as the first kind, removes the region EZ-BM of CSC lesion region as the second class, CSC lesion Region is used as full convolutional neural networks model segmentation standard as third class, by these three types.
9. the automatic segmenting system of retinal images lesion region according to claim 7, characterized in that described image compensation Dividing module includes the second compensating module of image primary segmentation module, the first compensating module of image and image；Wherein:
Described image primary segmentation module, the probability for belonging to the second class using retinal images are greater than 0.5 as extraction EZ- The criterion in the region BM, the probability that retinal images belong to third class are greater than 0.5 as the judgement mark for extracting CSC lesion region Standard obtains preliminary automatic segmentation binary result using trained full convolutional neural networks to retinal images primary segmentation, Including EZ-BM binary map and CSC binary map；
The first compensating module of described image, for by retaining maximum area region in EZ-BM binary map, removal to be located in EZ- Abnormal external plexiform layer detection error outside the region BM, obtains EZ-BM area probability figure；
The second compensating module of described image, for the method using the mean value for maximizing CSC lesion region edge neighborhood minimum value Iteration finds optimum a-value, and the pixel value of retinal images in test set is multiplied with obtained α value, obtains new retinal map Picture is again split image data new in test set using trained full convolutional neural networks model, obtains best The lesion CSC probability graph.
10. the automatic segmenting system of retinal images lesion region according to claim 7, characterized in that described image Two compensating modules are also used to:
Judge whether the mean value of each 8 neighborhood minimum of pixel of CSC binary map edge is equal to 1；
If each 8 neighborhood minimum of pixel of CSC binary map edge is equal to 1, α value and corresponding CSC binary map are exported, is tied Beam iteration；
If the mean value of each 8 neighborhood minimum of pixel of CSC binary map edge is not equal to 1, judge whether α reaches the upper of α=3 Limit；
If reaching the upper limit, α and classification corresponding to the mean value of each 8 neighborhood minimum of pixel of CSC binary map edge are found As a result, and exporting α and corresponding CSC binary map, end iteration；
If α does not reach the upper limit of α=3, α=α+0.1 is made to restart to continue iteration.
CN201910309970.2A 2019-04-17 2019-04-17 Automatic segmentation method and system for retina image pathological change area Expired - Fee Related CN109993757B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910309970.2A CN109993757B (en) 2019-04-17 2019-04-17 Automatic segmentation method and system for retina image pathological change area

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910309970.2A CN109993757B (en) 2019-04-17 2019-04-17 Automatic segmentation method and system for retina image pathological change area

Publications (2)

Publication Number Publication Date
CN109993757A true CN109993757A (en) 2019-07-09
CN109993757B CN109993757B (en) 2021-01-08

Family

ID=67134010

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910309970.2A Expired - Fee Related CN109993757B (en) 2019-04-17 2019-04-17 Automatic segmentation method and system for retina image pathological change area

Country Status (1)

Country Link
CN (1) CN109993757B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113554593A (en) * 2021-06-15 2021-10-26 汕头大学 Choroid thickness estimation method, apparatus, device and medium

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101667289A (en) * 2008-11-19 2010-03-10 西安电子科技大学 Retinal image segmentation method based on NSCT feature extraction and supervised classification
CN107767380A (en) * 2017-12-06 2018-03-06 电子科技大学 A kind of compound visual field skin lens image dividing method of high-resolution based on global empty convolution
US20180089505A1 (en) * 2016-09-23 2018-03-29 Samsung Electronics Co., Ltd. System and method for deep network fusion for fast and robust object detection
CN108095683A (en) * 2016-11-11 2018-06-01 北京羽医甘蓝信息技术有限公司 The method and apparatus of processing eye fundus image based on deep learning
CN108520522A (en) * 2017-12-31 2018-09-11 南京航空航天大学 Retinal fundus images dividing method based on the full convolutional neural networks of depth
CN109493330A (en) * 2018-11-06 2019-03-19 电子科技大学 A kind of nucleus example dividing method based on multi-task learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101667289A (en) * 2008-11-19 2010-03-10 西安电子科技大学 Retinal image segmentation method based on NSCT feature extraction and supervised classification
US20180089505A1 (en) * 2016-09-23 2018-03-29 Samsung Electronics Co., Ltd. System and method for deep network fusion for fast and robust object detection
CN108095683A (en) * 2016-11-11 2018-06-01 北京羽医甘蓝信息技术有限公司 The method and apparatus of processing eye fundus image based on deep learning
CN107767380A (en) * 2017-12-06 2018-03-06 电子科技大学 A kind of compound visual field skin lens image dividing method of high-resolution based on global empty convolution
CN108520522A (en) * 2017-12-31 2018-09-11 南京航空航天大学 Retinal fundus images dividing method based on the full convolutional neural networks of depth
CN109493330A (en) * 2018-11-06 2019-03-19 电子科技大学 A kind of nucleus example dividing method based on multi-task learning

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
J LONG 等: "Fully convolutional networks for semantic segmentation", 《COMPUTER VISION&PATTERN RECOGNITION》 *
ZHUO WANG 等: "Automated detection of photoreceptor disruption in mild diabetic retinopathy on volumetric optical coherence tomography", 《BIOMED. OPT. EXPRESS》 *
王卓: "基于OCT图像的视网膜椭圆体带损伤检测技术研究与应用", 《中国优秀硕士学位论文全文数据库信息科技辑》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113554593A (en) * 2021-06-15 2021-10-26 汕头大学 Choroid thickness estimation method, apparatus, device and medium
CN113554593B (en) * 2021-06-15 2023-07-25 汕头大学 Choroid thickness estimation method, apparatus, equipment and medium

Also Published As

Publication number Publication date
CN109993757B (en) 2021-01-08

Similar Documents

Publication Publication Date Title
CN105787927B (en) Automatic identification method is oozed out in a kind of color fundus photograph image
CN102842136B (en) A kind of optic disk projective iteration method of comprehensive vascular distribution and optic disk appearance characteristics
CN109087302A (en) A kind of eye fundus image blood vessel segmentation method and apparatus
CN104318546B (en) Multi-scale analysis-based greenhouse field plant leaf margin extraction method and system
CN108095683A (en) The method and apparatus of processing eye fundus image based on deep learning
CN106846293B (en) Image processing method and device
CN111797712B (en) Remote sensing image cloud and cloud shadow detection method based on multi-scale feature fusion network
CN107330876A (en) A kind of image automatic diagnosis method based on convolutional neural networks
CN110503610B (en) GAN network-based image rain and snow trace removing method
CN107122597B (en) Intelligent diagnosis system for corneal damage
CN110428421A (en) Macula lutea image region segmentation method and apparatus
CN109447945A (en) Wheat Basic Seedling rapid counting method based on machine vision and graphics process
CN103996023B (en) Light field face recognition method based on depth belief network
CN108961301A (en) It is a kind of based on the unsupervised Chaetoceros image partition method classified pixel-by-pixel
CN109993757A (en) A kind of retinal images lesion region automatic division method and system
CN109325498A (en) The Vein extraction algorithm of Canny operator is improved based on window dynamic threshold
CN109872337B (en) Eye fundus image optic disc segmentation method based on rapid mean shift
CN114596274A (en) Natural background citrus greening disease detection method based on improved Cascade RCNN network
CN110309808A (en) A kind of adaptive smog root node detection method under a wide range of scale space
CN106846314A (en) A kind of image partition method based on post-operative cornea OCT image datas
Frucci et al. Using contrast and directional information for retinal vessels segmentation
CN109410191A (en) Optical fundus blood vessel localization method and its anaemia screening method based on OCT image
Foracchia et al. Cell contour detection in corneal endothelium in-vivo microscopy
CN108573267A (en) A kind of method and device of liver organization textural classification
Bedke et al. Secondary glaucoma diagnosis technique using retinal nerve fiber layer arteries

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
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20210108

CF01 Termination of patent right due to non-payment of annual fee