Review on Retinal Image Segmentation Techniques

Review on Retinal Image Segmentation Techniques

Mr. Madhusudhan S

Dept of ECE AIeMS

Ms. Namratha S Rao

6th sem ECE AIeMS

Ms. Dhanyashree S Navilekar

6th sem ECE AIeMS

Abstractnow days eye diseases are common problem in all age group people starts form infant to old age. The detection and extraction of these diseases in a difficult and time consuming task, for this fundus and retinal digital images are considered, the first and foremost method is to extract vessel in fundus images. For this three methods are used supervised methods, here training set are applied to extract vessel information by pre-defined algorithms and this method is manually handled using gold std, Vessel extraction is done first before pathology algorithms is applied. In unsupervised the detection and extraction is made automatic means The training set and ground truth labeling which wont directly applicable to algorithm . Retinal vessels extraction are improving because of noninvasive imaging the retinal images also the data obtained from the structure of the vasculature and this information is necessary in the detection and, diagnosis of a fundus image retinal diseases and pathologies which includes glaucoma, hypertension, Diabetics Retina graph, and Age based Macula De- generation. These can be detected by fast segmentation algorithms.

Keywords: Age-based Macula De-generation; Diabetics Retina graph ; Diabetics macula edema ; Glaucoma; Hyper-tension

1. INTRODUCTION

   Eye related vessels detection and extraction are the major pathological problem in the field of ophthalmology. The etiology of the vessel diseases are still a research area where lot of research and development are in progress. For the identification, extraction and localization of retinal fundus and retinal

   images in other format aims to divide the different retina curvature and vasculature tissues, which includes widened and narrow structures with in the retina fundus and retinal background images and other optical structures like optic-disc, exudates, macular, fovea and lesion. Research and Development of more than two decades, the techniques with innovative and fast approaches are applied with computerized methods applicable to localizing and segmentation in retinal eye vessels like structure are important and its becoming more crucial and much needed to real time clinical applications. The fundus image eye retinal vessel segment methods have some common processing stages starts with preprocessing stage, midprocessing stage and postprocessing stage. The review categorizes on the basics of method used , algorithm flow and technical methods applied in the each stages. The supervised algorithm methods are to build on pre classify data, the efficiency of supervised method is superior than un-supervised methods which produces a accurate results to good fundus retina eye images. These supervised and unsupervised methods are most commonly used image segmentation methods. Mainly describes the following 5 algorithms for analysis. The first one is threshold segmentation method. this method is most commonly used in region-based segmentation algorithms [2]. Its use is to automatically extract the normalized threshold value on the bases of certain criterion, and apply these pixels values

   according to the gray scale level to achieve required clustering. The idea of the general regional grow technique is to merge the pixel values with same properties to get the region. The Edge detection segmentation algorithm is the use of various regions of the picture element gray level or color level discontinuity in detection area of the edge to achieve fundus image segmentation [3]. In clustering segmentation the problem of setting a label to every image element in the retinal image which consists of three regions . The first is image which contains objects. Secondly give the object border. Thirdly object area in the image is marked with a partial picture element [5].

   Figure 1: General vessel extraction algorithm

   The potential and the ability of retinal and fundus image segment or decompose methods are to excerpt or remove the eye vascular and which is endorse by so many of cadent. The widely

   applied parameters are: mean Truly- Positive-Rate (TPR), mediocre Falsely Positive Rate (FPR), moderate-Sensitivity, mean-Specificity, mean-Accuracy, and exactness. Sensitive and specific parameters are the most commonly used grades in ophthalmology. If values are higher for these parameters, then the treatment is better. The sensitive parameter ruminates ability of these method to perceive eye vessels picture elements, the specific parameters contemplate the aptness of the method to extract no-vessel elements . Specificity and specific parameters reflects headlining of breakthrough which is related to accurate value in some of allopathic eye retina fundus imaging.

   The vessel extraction algorithm generally applies following three parameters for exact extraction of vessels.

   Sensitivity =+ Specificity =+

   Accuracy =++++

   Where, TP:True Positives: vessel pixels which are correctly classified.

   FP, False Positives: non-vessel pixels which are wrongly classified as vessel pixels.

   TN:True Negatives: non-vessel pixels which are correctly classified.

   FN:False Negatives: vessel pixels which are wrongly classified as non-vessel pixels.

   Sensitivity: Ratio of correctly classified vessel pixels and total number of vessel pixels.

   Specificity: Ratio of correctly classified non-vessel pixels and total number of non- vessel pixels.

   Accuracy: Extent to which algorithm works correctly.

2. Supervised methods

   Line Strength, Multi-scale Gabor and Morphological Features: in this method a seven-dimensional attribute-vector made with evaluating output comes off the linear morphological operators, line brawn with the oriented-Gabor-filters in different measures. This method encodes by a attribute-vector using time intense values a with in vessels location, size and area at many measures . A Gaussian Mixture Model and the Bayesian Classifier are applied to classification of the fundus image into vessel and no-vessel picture elements. Improved matched filtering method : this method applies a newly generated High-re- solution eye retina fundus imaging database. The matched filtering and novel method gives accurate segmenting technique for retinal vessels. It provides an accuracy of

   0.95 percent when applied to a newly obtained imaging datas.

   Major Vessel Extraction & Subimage Classification method: in this method green based plane of the retinal image is applied in pre-processing which generates binaries image after HPF also on more binaries image from the morphological re obtained and intensify image related to the vessel and no-vessel regions. The regions that are repeated in both the binaries images when compared are extracted as the major vessels, next all residue picture element contained in two binaries images are classed using a GMM-classifier by a group of 8 features obtained based on neighbor hood and the ordered gradients images. Next is post-processing stage. This method of segmentation gives accuracy of 95.20%, 95.150%, and 95.30% for 100 with

   the mean of time to execute of 3.1s, 6.7s, and 11.7sec applied to datasets DRIVEs, STAREs, and CHASE_DB1database.

   Ridge-Based method: A method is presented for the automatic segmenta of vessels using 2-D colour image of retinal fundus . The methodology gives an extract of image ridges, that are coincided approx. on to the vessels centre lines For each picture element, fatured vector are to be performed which makes use of property from the patches, line element. The featured vector are classed by NN-c and sequenced forwarded-feature selection. This method is tested on 40 label pictures . This algorithm provides an area under the receiver operating curve of 0.9520.

   In the combinational radial projected, semi- supervised self-training using SVM method

   ,The vesseled center lines, narrower low contrasted vessel are to be pointed by radial projection. The modify steerable complexes wavelet are applied for the vessel enhancement, line strength are considered for the enhancement of vessel to get the featured vectors. This method is very nice in extracting narrower and very lowered contrasted vessel but subjected to noise. the parameters like accuracy, sensitive and specificities on the DRIVE database is 94.34, 74.10, and 97.51, and for

   the STARE database 94.97, 72.60, and

   97.56 for 100 respectively.

   Neural network based method using supervised methodology: This uses a seven- Dimensional featured vectors which is made up of level gray values. A multicore FFNN is applied to get training-set and to classify. The layer of input contains of 7 neurons, the hidden three layers contains 15 Neurons x 3 and 1 for output layer . The mean of accuracy, A.U.C, sensitive and specificity when applied to DRIVEs DB is 94.52, 95.88, 70.67, and 98.01%,

   respectively and for the STARE database 95.26, 97.69, 69.44, and 98.19%,

   respectively.

3. Unsupervised Methods

   Trimline and np window estimation method: this is a unsupervised vascular segmentation algorithm concerned to eye fundus images, this can be generalized for 2-D vascular image. This method is for intensities distributed estimate applied to localized imaging-patches and to treat smaller vessel by performing transform to the 1-D and get enhance detection and high accuracy of

   93.420 percent vs the active-contour based method of accuracy 92.580% on the public available DRIVEs db Without manual labeled training set method: in this two methods for detection of exudates in the fundus image are presented which do not use a supervised learning step instead uses unsupervised method and not require any labeling of lesion training sets which are not fast and time consuming to create, which is prone to manual error. This method evaluate algorithm with different dataset and compare results with 2 recent exudate segmentation methods as compared there is a magnitude reduction in computational time.

   In infinite Peri-meter Actived Contour Hybrid Regional method : the neighbor hood of boundary are applied for effective detecting smaller oscillatory function or repeating branch For general segmentation, the local-phase method is applied because of its superiority in preserve edges of the vessels information and the picture intensity parameter data for segmentation. The sensitivity, specificity and accuracy of 0.742 0.982 0.954 respectively when applied in DRIVEds.

   Divergence of Vector Fields method: this methodology is constructed to detect vessels in retinal image fundus. This are executed by applying Laplacian operator and noise data are altered related to lines in center, and detected by normalize gradient field. This method can avoid detect falser vessels in error regions and gives better outputs to health congested areas.

   Hessian matrix and clustering algorithm method: this is a novel unsupervised one. The new vessel is based on detecting vessel center lines and orientation in scale space. This vesselness marks a generated ground truth image is generated by thres-holding and removing smaller segments. (TPR) of

   1. at (FPR) of 0.45 is achieved compared with TPR of 0.76 at the same FPR . Unsupervised fuzzy vessel segment : this approach applies image intensity data from red band, green band channels fundus image to analyze and modify nonuniform illuminate in images. The Match filter is used to get enhanced and contrasted vessels with vs back ground. the timely weight FC-means cluster and connect component label is considered in the identifing the tree like structure. The ROC curve 95.18 and 96.02, for DRIVE and STARE databases.

      GLCM: local entropy rate in combined gray scale level co-occurrence is used for vessel segmentation. Match filter applied to enhancement vessel then computation of the GLCM, then from this the statistic feature is evaluated the stranded parameters obtained are 0.9648, 0.9480 and 0.9759.

      Local and region-based probing thresholded along a match filter approached method : described local, region-base property related to retina vessel for segment by applying a threshold-probing applied to match filters output pictures. the image analyzes MFR

      images, after this iterative threshold probing for each picture element as vessel or non- vessel. The hand-labeled images are tested on basic TH of the MFR. A fifteen times reduce in the calculation of false positives for MFR and seventy five percent for true positive is observed.

4. Advance methods

   Optimal multi threshold method in this modus operandi Firstly, uneven enhancements is applied used to steadiness the uneven back ground reflected by the light source. Secondly the image is decomposed into grids. Thirdly , optimum threshold is assigned. Then the vessel are segment in gray level scale images. By this method a good continuity and accurate center-line is obtained.

   Local entropy thresholding method : This method uses matched filtering by 2 different numericals of variance parameter. Pre processing are applied to enhance the quality of the image with the eliminating image noise Local average data TH then the length-filtering applied to get last autopsy of vessel. Tested on 40 pictures from DRIVEs while tested and obtained 0.8063, 0.9716 and 0.9684 respectively .

   MH and FCMD Cluster Method: this uses automated technique for disease etiology, first divide the gore vein and elucidate the parameters that produces the medical state of the resigned. labouring of gore vein of the retina fundus is very slow and long process.

   Line-Operators with Support-Vector Classify method: a line detector, is first given to the green-band channel Which is root on the calculation of mean color less level on cord of fix length passing via the earmark elements at various inclination. The un-supervised picture element classify and

   employed to 2 boxy cord locator with the grey level of the target picture element to construct a attribute vector .

   Optimal multi-threshold method: in this modus operandi the retinal blood vein are segment decivise. A uneven enhance is applied to equalize back ground reflected by light source. Next image is decomposed like mesh. the optimum limit is elected for each mesh. At the end bvs are segment in no color scale images. The with drawn bvs will have well progression and precise center-line which provides useful data for the ophthalmologist.

   Figure 1 the schematic outline of general method applied to different technique.

5. Conclusions

   The segmentation is the first and foremost requirement for the retinal blood vessels in recent days and fast growing research area. The accurate and effective extract of retinal vasculature bvs provides very necessary of many computerized systems in detection, extraction and identification of oculist illness. As lots of hundreds of promising methods currently available, need for improve and advance the segment ways. the review algorithms contribute for etiology and pathological and noise contained retinal

   eye images and are necessary for scrutiny of retinal fundus color images. above mentioned ways exists in this review are examined with datasets or database with at least 10 images each from the DRIVEs and STAREs dbs. The measures for performance in most of the papers are calculated on a small number of images of particular parameters. The images in the DRIVE and STARE databases.The above reviewed methods can be applied to different fundus imagestaken by different resolution fundus cameras under many other

   abode conditions which is an current area for fact finding in vein decimation algorithms.

6. Future scope

The techniques reviewed can be extended to advanced techniques which are applied to new data bases including HD and local data bases and also can be reviewed based on different parameters considered for evaluation of results.

Method	Source	Sensitivity(%)	Specificity(%)	Accuracy(%)
observer	DRIVEdatabase	77.63	97.23	94.7
observer	STAREdatabase	89.50	93.84	93.4
Sinthanayothin[10]	Local database	83.30	91.00	——-
Abramoff[11]	DRIVEdatabase	71.45	94.16	92.9
Staal[12]	DRIVEdatabase	94.42	95.2	—
Staal[12]	STARE database	95.16	96.14	—-
Soares[13]	DRIVEdatabase	94.66	96.14	—-
Soares[13]	STARE database	94.80	96.71	—-
Ricci, Perfetti[14]	DRIVEdatabase	95.63	95.58	—-
Ricci, Perfetti[14]	STARE database	95.84	96.02	—-

Table 1: Comparison between supervised methods

Method	source	Sensitivity(%)	Specificity(%)	Accuracy(%)
observer	DRIVEdatabase	77.63	97.23	94.70
observer	STAREdatabase	89.51	93.84	93.48
Ng .	STAREdatabase	70.00	95.30	—-
Kande .	DRIVEdatabase	89.11	95.18	—-
Kande	STAREdatabase	89.76	92.98	—–
Salem .	STAREdatabase	82.15	97.50	—–
Villalobos-Castaldi	DRIVEdatabase	96.48	94.80	97.59

Table 2: Comparison between unsupervised methods

Method	Sensitivity(%)	Specificity(%)	Accuracy(%)
Martinez-Perez. [19]	72.460	96.550	93.440
Perez. [20]	66.000	96.120	92.200
Anzalone. [21]	————-	—————	94.19
Vlachos and Dermatas. [22]	74.70	95.50	92.90
Fraz, Rudnicka and Barman. [18]	74.00	98.00	96.00
Yin and Bourennane. [17]	72.48	96.66	—————–

References

Table 3: Comparison between other methods

1. Retinal Image Segmentation using Texture,Thresholding, and Morphological Operations Aqeel F AQEEL and Subra Ganesan, International Conference on Electro/Information Technology, May 15-17, 2011

2. Application of Intuitionistic Fuzzy Histon Segmentation for theAutomated Detection of Optic Disc in Digital Fundus ImagesMuthu Rama Krishnan M (BHI 2012)Hong Kong and Shenzhen, China, 2-7

3. SEGMENTATION OF BLOOD VESSELS IN COLOR FUNDUS IMAGES BASED ON OPTIMAL MULTI-THRESHOLD METHOD FA-LING YIt, WEN-HUAXU2 1 Sch. of Med. Inf. Eng., Guangdong Pharm. Univ., Guangdong Pharm. Univ.,

   GuangZhou, , 15-17 July, 2012

4. Evaluation of Algorithms for Segmentation of Retinal Blood Vessels Ramish B. Kawadiwale Dept. of Electronics and Telecommunication Vishwakarma Institute of Technology Pune,

5. Vessel Segmentation of Retinal Images using

   Marr- Hildreth and FCM Divide Clustering Method Sukhpreet Kaur1 Ms Isha Vats2 M.Tech (Scholar) 1 Landran(Punjab) International Conference on Image Information Processing (ICIIP)

6. Blood Vessels Segmentation in Retina: Preliminary Assessment of the MathematicalMorphology and of the Wavelet Transform Techniques JORGE J. G. LEANDRO1 ROBERTO M CESAR JR1 HERBERT F. JELINEK2, Rua do Matão, 1010, São Paulo, SP, 05508-900 Brazil [7] Retinal Vessel Segmentation Using MinimumSpanning Superpixel Tree Detector Bin Sheng , Ping Li, Shuangjia Mo, Huating Li, Xuhong Hou, Qiang Wu, Jing Qin , Ruogu Fang, and David Dagan Feng , Fellow, IEEE IEEE TRANSACTIONS ON CYBERNETICS 1

1. SEGMENTATION OF BLOOD VESSELS IN COLOR FUNDUS IMAGES BASED ON OPTIMAL MULTI-THRESHOLD METHOD FA-LING YIt,

   WEN-HUAXU2 1 Sch. of Med. Inf. Eng., Guangdong Pharm. Univ., gzwhxu@163.com Proceedings of the 2012 International Conference on Machine Learning and Cybernetics, Xian, 15-17 July, 2012

2. Segmentation of Blood Vessels in Fundus Color Images by Radon Transform and Morphological Reconstruction Reza Pourreza, Hamidreza Pourreza, Touka Banaee on Advanced Computational Intelligence August 25-27, 2010 – Suzhou, Jiangsu,

3. Segmentation of Retinal Vasculature in Colour Fundus Images Kee Yong Pang, Lila Iznita I, Ahmad Fadzil M H, Hanung A. N., Hermawan N., Vijanth S (CITISIA 2009)Monash University, Sunway campus, Malaysia, 25th & 26th July 2009.

4. Segmentation of Vessel-Like Patterns Using Mathematical Morphology and Curvature Evaluation Frédéric Zana and Jean-Claude Klein IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 10, NO.7, JULY 2001

5. Retinal Image Segmentation using Texture, Thresholding, and Morphological Operations Aqeel F AQEEL and Subra Ganesanl Conference on Electro/Information Technology, May 15-17, 2011

6. Blood vessel segmentation methodologies in retinal images A survey M.M. Fraza,, P. Remagninoa, A. Hoppea, B. Uyyanonvarab, A.R. Rudnickac, C.G. Owenc, S.A. Barmana St. Georges,

   University of London

7. Retinal Vessels Segmentation Techniques and

   Algorithms: A Survey Jasem Almotiri 1,*, Khaled Elleithy 1 and Abdelrahman Elleithy 2 USA; * Correspondence: jalmotir@my.bridgeport.edu; Tel.:

   +1-203-606-0918;

8. Leveraging Multiscale Hessian-Based Enhancement With a Novel Exudate Inpainting Alessandro Mecocci, and Emanuele Trucco IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 00, NO. 0, 2015 1

[16]M. Moazam Fraz, Alicja R. Rudnicka,Christopher G. Owen and Sarah A. Barman, Delineation of blood vessels in pediatric retinal images using decision trees-based ensemble lassification, in International journal of computer assisted radiology and surgery, pp. 1-17, (2013).

[17] Y. Yin, M. Adel, and S. Bourennane, Automatic Segmen- tation and Measurement of Vasculature in Retinal Fundus Images Using Probabilistic Formulation, ,vol.2013, pp. 16, (2013).

[18 ]S. Roychowdhury, D. D. Koozekanani , Keshab K and Parhi, Blood Vessel Segmentation of Fundus Images by Major Vessel Extraction and Sub-Image Classification, , pp. 2168-2194, (2014).

[19]. D. Marn, , A new supervised method for blod vessel segmentation inretinal in Medical Imaging, IEEE Transactions onl,vol.30,no.1, pp. 146-158, (2011). J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. van Ginneken , [20] M.E. Martinez-Perez, A.D. Hughes, S.A. Thom, A.A. Bharath, K.H. Parker , Segmentation of blood vessels from red-free and fluorescein retinal images,

Medical Image Analysis, pp. 47-61, (2007).

1. M.E.M. Perez, Improvement of a retinal blood vessel segmentation method using the Insight Segmentation and Registration Toolkit (ITK), 29th IEEE Transactions on Medical Imaging, vol. 23, no. 3, pp. 892-895, (2007).

2. M. Vlachos, E. Dermatas, Multi-scale retinal vessel segmen- tation using line tracking, Computerized Medical Imaging and Graphics, vol. 34, no. 3, pp. 213-227, (2010)