CN110189320A - Segmentation Method of Retinal Blood Vessels based on middle layer block space structure - Google Patents

Segmentation Method of Retinal Blood Vessels based on middle layer block space structure Download PDF

Info

Publication number
CN110189320A
CN110189320A CN201910471171.5A CN201910471171A CN110189320A CN 110189320 A CN110189320 A CN 110189320A CN 201910471171 A CN201910471171 A CN 201910471171A CN 110189320 A CN110189320 A CN 110189320A
Authority
CN
China
Prior art keywords
block
blood vessel
random forest
feature
label value
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
CN201910471171.5A
Other languages
Chinese (zh)
Other versions
CN110189320B (en
Inventor
赵荣昌
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.)
Central South University
Original Assignee
Central South 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 Central South University filed Critical Central South University
Priority to CN201910471171.5A priority Critical patent/CN110189320B/en
Publication of CN110189320A publication Critical patent/CN110189320A/en
Application granted granted Critical
Publication of CN110189320B publication Critical patent/CN110189320B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30041—Eye; Retina; Ophthalmic

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Quality & Reliability (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of Segmentation Method of Retinal Blood Vessels based on middle layer block space structure, including construction sample set；The corresponding blood vessel structure label value of feature representation is extracted from the block structure of color image sample set；Random forest grader is constructed, is classified to block structure, indicates the blood vessel structure on new images using the content linear combination in feature integrated model；Input colored eyeground figure to be analyzed, blood vessel structure classification belonging to random forest grader detection structure block is used after extracting feature, color image medium vessels structure is indicated using the content sparse linear in model, and match blood vessel label value, by calculating the probability of overlay structure block label value on single pixel point, the segmentation to image is completed.The method of the present invention can quick and precisely divide retinal vessel, and high reliablity, Riming time of algorithm are short.

Description

Segmentation Method of Retinal Blood Vessels based on middle layer block space structure
Technical field
The invention belongs to field of image processings, and in particular to a kind of retinal vessel based on middle layer block space structure point Segmentation method.
Background technique
With the raising of economic level, people also become more concerned with the physical condition of oneself.And with intelligent hand The universal and a large amount of of machine etc. uses, and ophthalmology disease has seriously endangered and influenced daily life.Related data shows, Glaucoma, congenital and Genetic eye diseases, eyeground pathological changes etc. have accounted for 8.8%, 5.1%, 8.4% blinding ratio respectively.Glycosuria The fundus oculi diseases such as sick retinopathy, glaucoma have the characteristics that again irreversible, high incidence, to these diseases The life harmfulness of patient is very big.
Colored eyeground figure is the important evidence of medical clinic applications ophthalmology disease, can be used as and sentences to the analysis of its structure Disconnected such as hypertension, diabetes, the important evidence of cardio-cerebrovascular diseases.Meanwhile retinal vessel can also be other systemic diseases Disease provides effective diagnosis basis.For example, showing body, there may be certain circulations when retinal microvascular changes Property disease risk.Retinal vessel is as the deep layer that can be uniquely observed directly by image technology, atraumatic in human body Capillary, the severity of disease such as its structure change and diabetes and recovery situation, which exist, closely to be contacted.Glycosuria Sick retinopathy is capilary changes since human body.The structure of retinal vessel is more stable, even if With the aging of human body, biggish variation will not be generated, in addition to diabetes, hypertension, cardio-cerebrovascular diseases or outer masterpiece Other than influence, its structure change is also little affected by the influence of Other diseases.
Currently, the research for retinal vessel cutting techniques, it is other can be substantially divided into two major classes: being based on supervised learning The retinal vessel segmentation of method, the retinal vessel segmentation based on unsupervised learning.The wherein retinal blood of unsupervised learning Pipe segmentation contains blood vessel method for tracing, the method for matched filtering, morphologic method, method based on model etc. again.
In existing Segmentation Method of Retinal Blood Vessels, the method for unsupervised learning is generally without supervised learning method point Precision height is cut, the segmentation precision of supervised learning method and timeliness etc. all show more preferable.It now more commonly will be non-supervisory Learning method, which incorporates in supervised learning method, to be used, in the hope of better retinal vessel segmentation effect.But existing prison Educational inspector's learning method is not still able to satisfy the requirement of Real-time segmentation, and there is biggish delay in sliced time, has seriously affected retinal blood The application of pipe cutting techniques.
Summary of the invention
The purpose of the present invention is to provide a kind of high reliablity and the operation time short views based on middle layer block space structure Retinal vasculature dividing method.
This Segmentation Method of Retinal Blood Vessels based on middle layer block space structure provided by the invention, includes the following steps:
S1. sample set is constructed；
S2. the corresponding blood vessel structure label value of feature representation is extracted from the block structure in sample set color image；
S3. random forest grader is constructed, is classified using the random forest grader of construction to block structure, then make The blood vessel structure in input picture is indicated with the content linear combination in feature integrated model；
S4. colored eyeground figure to be analyzed is inputted, is examined after extracting feature using the random forest grader of step S3 building Blood vessel structure classification belonging to the block structure in colored eyeground figure to be analyzed is surveyed, and using the sparse line of content in training pattern Property indicate color image medium vessels structure, provide corresponding blood vessel label value, pass through the probability calculation of overlay structure block label value It is final as a result, completing the segmentation to image.
Construction sample set described in step S1 specially constructs sample set using following steps:
A. original image is indicated with Z, is the blood vessel label construction block of Num × Num, A is the model that structure block feature is constituted And size is K × Num × Num, X is the K dimensional vector of rarefaction representation, and N is the quantity of block structure, and Z ≈ A*X；
B. the image that several sizes are Num1 × Num1 is chosen in the database, each pixel of image is chosen in scanning, And after scanning is to puncta vasculosa, the block structure of Num × Num size is extracted centered on the puncta vasculosa, records the block structure of extraction Central pixel point position；
C. the block structure extracted using in step B as model training content, to construct sample set.
Block structure in color image in slave sample set described in step S2 extracts the corresponding blood vessel structure of feature representation Label value specially extracts blood vessel structure label value using following steps:
A. the multiple features channel information of each block structure is calculated；
B. the multiple features channel information obtained based on step a is carried out self-similarity characteristics calculating, and matches blood vessel structure mark Label value.
Construction random forest grader described in step S3 classifies to block structure using random forest grader, tool Body is to construct random forest grader using following steps and carry out block structure classification:
(1) decision tree: the label of its left subtree of decision tree F recursion cycle or right subtree, each node in tree are constructed There is a division function, determines that node is just to stop in left branch or right branch until it reaches leaf node by division function Division, realizes the classification to blood vessel structure block, and final output is stored on leaf node；
(2) it constructs one group of decision tree and forms random forest, it is ensured that the training sample and feature selected at random are various enough Change, prevents from generating overfitting in random forest training process；
(3) random forest predicts the final output of one group of decision tree voting results, to complete block structure point Class.
Segmentation Method of Retinal Blood Vessels provided by the invention based on middle layer block space structure, according to colored eye fundus image The sparsity structure feature of retinal vessel, it is assumed that eye fundus image shares the same or similar tiny middle layer figure of blood vessel structure As block, propose to realize that retinal vessel is divided using the method for middle layer block space structure.Different from having method, the present invention is not Based on classifying again with pixel, but blood vessel segmentation is realized by more classification to blood vessel structure block, can accelerate blood vessel detection Time；There is no complicated feature in the feature of selection, is the combination of color and Gauss feature, such feature is not required to There are complicated calculating and parameter setting；In order to realize more classification of blood vessel structure, random forest grader is selected, its energy It realizes the classification to whole knot building block, feature is trained to model, the effect not only classified is good, but also speed is fast；In detection eye When base map is as blood vessel structure, use every two pixel for the method for step-length, so that can obtain on a pixel multiple The prediction label value of block structure can improve the accuracy of segmentation by the probability calculation to single pixel point label；Therefore The high reliablity of the method for the present invention, and sliced time is quick.
Detailed description of the invention
Fig. 1 is the flow diagram of the method for the present invention.
Fig. 2 is the local vascular block structure schematic diagram of various shapes of the method for the present invention.
Fig. 3 is colored eyeground figure block structure component part schematic diagram corresponding with its of the method for the present invention.
Fig. 4 is a peacekeeping two-dimensional Gaussian function image schematic diagram of the method for the present invention.
Fig. 5 is the result schematic diagram after the colored eye fundus image gaussian filtering of the method for the present invention.
Fig. 6 is the blood vessel structure classification schematic diagram of the method for the present invention.
Fig. 7 is the method schematic diagram that the method for the present invention extracts block structure in colored eyeground figure.
Specific embodiment
It is as shown in Figure 1 the method flow schematic diagram of the method for the present invention: provided by the invention this based on middle layer block space The Segmentation Method of Retinal Blood Vessels of structure, includes the following steps:
S1. sample set is constructed；Specially sample set is constructed using following steps:
A. original image is indicated with Z, is the blood vessel label construction block of Num × Num, A is the model that structure block feature is constituted And size is K × Num × Num, X is the K dimensional vector of rarefaction representation, and N is the quantity of block structure, and Z ≈ A*X；
B. the image that several sizes are Num1 × Num1 is chosen in the database, each pixel of image is chosen in scanning, And after scanning is to puncta vasculosa, the block structure of Num × Num size is extracted centered on the puncta vasculosa, records the block structure of extraction Central pixel point position；
C. the block structure extracted using in step B as model training content, to construct sample set；
When it is implemented, indicate original image using Z, i.e., size be 16 × 16 blood vessel label construction block, A then indicates The model that structure block feature is constituted, size are the K dimensional vectors that K × 16 × 16, X are rarefaction representation, and N is the number of block structure Amount, calculation formula are as follows:
Z≈A*X
Training set in the database includes colored eyeground figure two-value vessel graph corresponding with its, wherein each image Size normalization is 584 × 565, and selection represents the label construction block construction of different blood vessel structure from these two-value vessel graphs Initial matrix in model, wherein each block structure can indicate vessel information different in colored eyeground figure, such as Fig. 2 It show different blood vessel structure；
Each pixel in two-value blood-vessel image is scanned, after scanning is to puncta vasculosa, as center extraction size For 16 × 16 block structure, blood vessel label construction block, and the central pixel point position D (x, y) of interrecord structure block are obtained, so as to Late feature extracts, and a part of sample is chosen from the block structure extracted as the training content in model, as shown in Figure 3 The block structure etc. for color image and wherein taken out；
S2. the corresponding blood vessel structure label of feature representation is extracted from the block structure in color image corresponding in sample set Value；Specially blood vessel structure label value is extracted using following steps:
A. the multiple features channel information of each block structure is calculated；
B. the multiple features channel information obtained based on step a is carried out self-similarity characteristics calculating, and matches blood vessel structure Label value；
In the specific implementation, the feature of higher level is expressed after being combined by low-level features, and the feature in color image can To be expressed by combining some essential characteristics, when use low-level feature Multiple Combination expresses corresponding blood in the figure of colored eyeground It is also more advantageous to the result of blood vessel segmentation when pipe label value.Two kinds of feature is selected in the invention, is channel spy respectively It seeks peace self-similarity characteristics, wherein channel characteristics include the features such as color, gaussian filtering, directed access；
Three components in CIE-LUV color space can be for indicating a point, can be color therein with one Point is to indicate, then the distance between point is calculated using Euclid's formula, it is assumed that two colors are respectively (L1,U1,V1) and (L2,U2,V2), range formula is as follows, by calculating the color characteristic in the available block structure of range formula:
Gaussian filter is used to extract local feature, and in image detection, noise belongs to high-frequency part, gaussian filtering The maximum feature of device is can to do smoothing processing, and the removal noise jamming of high degree is true using the second order derivation of gaussian filtering Determine the position of blood vessel.Gaussian function is separable function, in two-dimensional Gaussian function, can be rolled up respectively to row and column Product operation, can be greatly reduced computation complexity in this way, and the two-dimensional Gaussian function of a peacekeeping is as follows, x, and y is that point is sat Mark, σ are standard deviations, and Fig. 3 is the distribution map of a peacekeeping two-dimensional Gaussian function
When using one-dimensional Gaussian function, same position in 16 × 16 corresponding colored eyeground figure of block structure is selected Block structure is filtered.The selection of standard deviation sigma value be it is very crucial, if standard deviation sigma is excessive, closer to mean value filter Wave device, it is relatively significant to the smooth effect of image, but the coefficient difference of each template is reduced；Standard deviation sigma is too small, each template Center coefficient is big, also big with surrounding formation difference, will be deteriorated to the smooth effect of image, two that σ is chosen in this chapter method Value, respectively 0,1.5 calculate, and Fig. 5 is the result for using whole picture eye fundus image 2-d gaussian filters；
Firstly, calculating each block structure multiple features channel information, each channel characteristics represent the different letters of same structure block Breath, the feature quantity of each picture structure block are 16 × 16 × C, and wherein C is the quantity of channel characteristics, channel characteristics by color, Gaussian sum directed access information composition, three Color Channel features are calculated by CIE-LUV color space, wherein standardization ladder It spends channel characteristics and is based on two scales realizations, extract this feature using two Gaussian filters are fuzzy, be 0,1.5 two by σ The gradient amplitude channel of value is split as four channels according to direction respectively, forms eight directed access, therefore channel characteristics include 3 Color Channel features, 2 gaussian filtering features and 8 directed access features, totally 13 channel characteristics.Based on content Image segmentation, which refers to, utilizes the color of image, texture, semanteme, shape and other features, it is intended to according to the division to picture material Carry out segmented image.The calculating of self-similarity characteristics is that 5 × 5 resolution ratio are used in block structure based on color and gradient channel information Cell, all block structures are sampled, in each channel layer be arranged 5 × 5 sizes cell, cell can Mutually to cover, a channel layer is extractedA self-similarity characteristics.
The feature that a block structure is calculated in the present invention includes: 3 color characteristics, 2 gaussian filtering features, 8 orientations Channel characteristics, totally 13 channel characteristics include 300 self-similarity characteristics, one 16 × 16 structure in each channel characteristics again Block includes 13 channel characteristics and 13 × 300 self-similarity characteristics in total；
S3. random forest grader is constructed.Classified as classifier to block structure using random forest, is reused Content linear combination in feature integrated model indicates the blood vessel structure on new images；Specially using following steps construction with Machine forest classified device is simultaneously classified:
(1) decision tree: the label of its left subtree of decision tree F recursion cycle or right subtree, each node in tree are constructed There is a division function, determines that node is just to stop in left branch or right branch until it reaches leaf node by division function Division, realizes the classification to blood vessel structure block, and final output is stored on leaf node；
(2) it constructs one group of decision tree and constitutes random forest, it is ensured that the training sample and feature selected at random are various enough Change, prevents from generating overfitting in random forest training process；
(3) random forest predicts the final output of one group of decision tree voting results, to complete point of middle layer block Class；
In the specific implementation, classified using random forest grader to block structure, use feature integrated model.From Its feature is chosen in the block structure of extraction for random forest training process, randomly selects blood vessel structure as in training process Label；
Decision tree is constructed, decision tree F passes through the label of its left subtree of recursion cycle or right subtree, each node in tree There is a division function, determines that node is just to stop in left branch or right branch until it reaches leaf node by division function It only divides, realizes the classification to blood vessel structure block, final output is stored on leaf node, as Fig. 6 shows that blood vessel structure is classified Example, the training process of every tree are that the independent and parallel depth capacity or sample for reaching tree when training is divided into certain threshold When value, stop the division of tree；
Construct one group of decision tree composition random forest, it is ensured that the training sample and feature selected at random are diversified enough, prevent Overfitting only is generated in random forest training process, although the phenomenon that single tree may also can generate over-fitting, its meeting Over-fitting is eliminated by increasing width；
8 decision trees of training form random forest, are trained each tree until the structure on each leaf node is pure Ability deconditioning when structure number of blocks on net or leaf node is less than 2, the depth capacity of each tree are 64；
Random forest predicts the final output of one group of decision tree voting results, it is assumed that one value x of input passes through Multiple decision tree predicted value ft(x) aggregation model is formed as its output valve, uses top-down recursion method
In order to ensure the diversity of tree, each tree is trained in random sampling from sample, for each node, characteristic attribute Be from all features random secondary sample F andFeature, by each layer of randomness injection tree, so that the dictionary of training Model is more complete.Gini coefficient is used for the feature selecting and decision of each node, as follows to be the expression of gini coefficient Formula:
PiIt is the frequency of the classification i occurred in sample T, NiIt is the quantity of the classification j in sample T, S is the sample number in T Amount, S1, S2Similar to T1And T2；
S4. colored eyeground figure to be analyzed is inputted, is examined after extracting feature using the random forest grader of step S3 building Blood vessel structure classification belonging to the block structure in colored eyeground figure to be analyzed is surveyed, and using the sparse line of content in training pattern Property indicate color image medium vessels structure, and give blood vessel label value, pass through calculate single pixel point on overlay structure block label value Probability, complete segmentation to image；
When it is implemented, the colored eyeground figure that input is new, pixel in scan image, every two pixel up and down Point is step-length, centered on the pixel that Current Scan arrives, extracts the block structure of size 16 × 16, as shown in Figure 7；
The feature in the corresponding colored eyeground figure of block structure of size 16 × 16 is calculated, is recorded with its central pixel point X special Vector is levied, expression formula is X ∈ F16*16*C, wherein C is characterized the quantity in channel, and feature includes four category features in model training；
Feature feeding random forest grader is trained, indicates this feature using the content sparse linear in model, And predict its corresponding blood vessel label value, to the structure block sort in color image, other than trained model, at random Forest classified device stores corresponding structure feature on each leaf node, so that more rapidly and easily prediction is colored Corresponding blood vessel structure label value in the figure of eyeground.
When information in the figure of sense colors eyeground, using two pixels as step-length therefrom selecting structure block for random Classify in forest classified device, each pixel can obtain the block structure label value of multiple overlappings, therefore in reality output, greatly Each pixel in small 16 × 16 block structure can averagely be obtained more than 256 secondary label predicted values.
Through the above steps, each pixel obtains the multiple prediction of adjacent bonds building block in a width colour eye fundus image Value carries out probability calculation to the predicted value of pixel in block structure, has obtained blood vessel probability graph, formula 3- to realize segmentation 12 be calculation expression, and wherein P represents probability, and x indicates the pixel in color image, piIt represents and passes through the i-th of pixel x A block structure, I are the probability graph of prediction, piIn I
The information of whole image is predicted due to using block structure, so that this chapter method shows very surprising effect Rate, and greatly reduced calculation amount, the prediction of each tree is all incoherent during detection, selects secondary pumping The method of sample, each decision tree have output and input overlapping, and this method also greatlys improve the accuracy of segmentation.

Claims (4)

1. a kind of Segmentation Method of Retinal Blood Vessels based on middle layer block space structure, includes the following steps:
S1. sample set is constructed；
S2. the corresponding blood vessel structure label value of feature representation is extracted from the block structure of color image sample set；
S3. random forest grader is constructed, is classified using the random forest grader of construction to block structure, reuses feature Content linear combination in integrated model indicates the blood vessel structure on new images；
S4. input colored eyeground figure to be analyzed, extract after feature using the random forest grader detection of step S3 building to Blood vessel structure classification belonging to the block structure in the figure of colored eyeground is analyzed, and is indicated using the content sparse linear in training pattern Color image medium vessels structure, and blood vessel label value is given, by the probability calculation of overlay structure block label value, complete to image Segmentation.
2. the Segmentation Method of Retinal Blood Vessels according to claim 1 based on middle layer block space structure, it is characterised in that step Construction sample set described in rapid S1, specially constructs sample set using following steps:
A. original image is indicated with Z, is the blood vessel label construction block of Num × Num, A is the model that structure block feature is constituted and big Small is K × Num × Num, and X is the K dimensional vector of rarefaction representation, and N is the quantity of block structure, and Z ≈ A*X；
B. the image that several sizes are Num1 × Num1 is chosen in the database, and each pixel of image is chosen in scanning, and works as After scanning puncta vasculosa, the block structure of Num × Num size is extracted centered on the puncta vasculosa, is recorded in the block structure of extraction Imago vegetarian refreshments position；
C. the block structure extracted using in step B as model training content, to construct sample set.
3. the Segmentation Method of Retinal Blood Vessels according to claim 2 based on middle layer block space structure, it is characterised in that step The block structure in color image in slave sample set described in rapid S2 extracts the corresponding blood vessel structure label value of feature representation, specifically To extract blood vessel structure label value using following steps:
A. the multiple features channel information of each block structure is calculated；
B. the multiple features channel information obtained based on step a carries out self-similarity characteristics calculating, to obtain final blood vessel structure Label value.
4. the Segmentation Method of Retinal Blood Vessels according to claim 3 based on middle layer block space structure, it is characterised in that step Construction random forest grader described in rapid S3, classifies to block structure using the random forest grader of construction, specially Random forest grader is constructed using following steps and is classified:
(1) decision tree: the label of its left subtree of decision tree F recursion cycle or right subtree is constructed, each node in tree has one A division function determines that node is just to stop division until it reaches leaf node in left branch or right branch by division function, Realize the classification to blood vessel structure block, final output is stored on leaf node；
(2) it constructs one group of decision tree and constitutes random forest, it is ensured that the training sample and feature selected at random are diversified enough, prevent Overfitting is generated in random forest training process；
(3) random forest predicts the final output of one group of decision tree voting results, to complete to classify.
CN201910471171.5A 2019-05-31 2019-05-31 Retina blood vessel segmentation method based on middle layer block space structure Active CN110189320B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910471171.5A CN110189320B (en) 2019-05-31 2019-05-31 Retina blood vessel segmentation method based on middle layer block space structure

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910471171.5A CN110189320B (en) 2019-05-31 2019-05-31 Retina blood vessel segmentation method based on middle layer block space structure

Publications (2)

Publication Number Publication Date
CN110189320A true CN110189320A (en) 2019-08-30
CN110189320B CN110189320B (en) 2023-04-07

Family

ID=67719511

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910471171.5A Active CN110189320B (en) 2019-05-31 2019-05-31 Retina blood vessel segmentation method based on middle layer block space structure

Country Status (1)

Country Link
CN (1) CN110189320B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110448267A (en) * 2019-09-06 2019-11-15 重庆贝奥新视野医疗设备有限公司 A kind of multimode eyeground dynamic imaging analysis system and its method

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010138645A2 (en) * 2009-05-29 2010-12-02 University Of Pittsburgh - Of The Commonwealth System Of Higher Education Blood vessel segmentation with three-dimensional spectral domain optical coherence tomography
CN106408562A (en) * 2016-09-22 2017-02-15 华南理工大学 Fundus image retinal vessel segmentation method and system based on deep learning
CN109523524A (en) * 2018-11-07 2019-03-26 电子科技大学 A kind of eye fundus image hard exudate detection method based on integrated study

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010138645A2 (en) * 2009-05-29 2010-12-02 University Of Pittsburgh - Of The Commonwealth System Of Higher Education Blood vessel segmentation with three-dimensional spectral domain optical coherence tomography
CN106408562A (en) * 2016-09-22 2017-02-15 华南理工大学 Fundus image retinal vessel segmentation method and system based on deep learning
CN109523524A (en) * 2018-11-07 2019-03-26 电子科技大学 A kind of eye fundus image hard exudate detection method based on integrated study

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110448267A (en) * 2019-09-06 2019-11-15 重庆贝奥新视野医疗设备有限公司 A kind of multimode eyeground dynamic imaging analysis system and its method
CN110448267B (en) * 2019-09-06 2021-05-25 重庆贝奥新视野医疗设备有限公司 Multimode fundus dynamic imaging analysis system and method

Also Published As

Publication number Publication date
CN110189320B (en) 2023-04-07

Similar Documents

Publication Publication Date Title
Chouhan et al. Soft computing approaches for image segmentation: a survey
CN107437092B (en) The classification method of retina OCT image based on Three dimensional convolution neural network
Gouda et al. Skin cancer classification using ResNet
Akram et al. Detection of neovascularization in retinal images using multivariate m-Mediods based classifier
CN109635846A (en) A kind of multiclass medical image judgment method and system
CN101667289B (en) Retinal image segmentation method based on NSCT feature extraction and supervised classification
CN109345538A (en) A kind of Segmentation Method of Retinal Blood Vessels based on convolutional neural networks
CN107977671A (en) A kind of tongue picture sorting technique based on multitask convolutional neural networks
CN110069958A (en) A kind of EEG signals method for quickly identifying of dense depth convolutional neural networks
CN104809480A (en) Retinal vessel segmentation method of fundus image based on classification and regression tree and AdaBoost
CN108256452A (en) A kind of method of the ECG signal classification of feature based fusion
Dey et al. Automatic detection of whitefly pest using statistical feature extraction and image classification methods
CN109410204A (en) A kind of processing of cortex cataract image and Enhancement Method based on CAM
CN108256453A (en) A kind of method based on one-dimensional ECG signal extraction two dimension CNN features
CN107748889A (en) A kind of Ultrasound Image of Breast Tumor automatic classification method
Moura et al. Combining abcd rule, texture features and transfer learning in automatic diagnosis of melanoma
Manjramkar Survey of diabetic retinopathy screening methods
CN110189320A (en) Segmentation Method of Retinal Blood Vessels based on middle layer block space structure
CN113343975A (en) Deep learning-based white blood cell classification system and method
Fan et al. Automated blood vessel segmentation in fundus image based on integral channel features and random forests
Zabihi et al. Retinal vessel segmentation using color image morphology and local binary patterns
Rao et al. A Methodology for Identification of Ayurvedic Plant Based on Machine Learning Algorithms
Rammy et al. Conditional patch-based generative adversarial network for retinal vessel segmentation
Mehidi et al. Retinal image segmentation u sing clustering methods: Performance analysis
Aggarwal et al. Morphological based optimized random forest classification for Indian oxygen plants

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