Professional Documents
Culture Documents
Predictive Diagnosis of Glaucoma
Predictive Diagnosis of Glaucoma
Abstract—Automatic evaluation of the retinal fundus image is regarded as one of the most important future
tools for early detection and treatment of progressive eye diseases like glaucoma. Glaucoma leads to progres-
sive degeneration of vision which is characterized by shape deformation of the optic cup associated with focal
notching, wherein the degeneration of the blood vessels results in the formation of a notch along the neuroret-
inal rim. In this study, we have developed a methodology for automated prediction of glaucoma based on fea-
ture analysis of the focal notching along the neuroretinal rim and cup to disc ratio values. This procedure has
three phases: the first phase segments the optic disc and cup by suppressing the blood vessels with dynamic
thresholding; the second phase computes the neuroretinal rim width to detect the presence and direction of
notching by the conventional ISNT rule apart from calculating the cup-to-disc ratio from the color fundus
image (CFI); the third phase uses linear support vector based machine learning algorithm by integrating
extracted parameters as features for classification of CFIs into glaucomatous or normal. The algorithm out-
puts have been evaluated on a freely available database of 101 images, each marked with decision of five glau-
coma expert ophthalmologists, thereby returning an accuracy rate of 87.128%.
Keywords: glaucoma, optic-disc segmentation, cup segmentation, blood vessel suppression, cup-to-disc
ratio, focal notching, retinal fundus image, computer aided diagnoses, ISNT rule, machine learning, support
vector machine
DOI: 10.1134/S1054661819030155
ISSN 1054-6618, Pattern Recognition and Image Analysis, 2019, Vol. 29, No. 3, pp. 523–532. © Pleiades Publishing, Ltd., 2019.
524 MUKHERJEE et al.
Nauroretinal rim
Main blood vessel
S
Optics disc
T N
Pallor region
Cup I
correlated. The contour of the cup is influenced by increases due to abnormal elongation of cup towards
this correlation. Along with this, intra ocular pressure, inferior (I) or superior (S) direction. The difference
retinal nerve fiber layer (RNFL) characteristics, focal between CDR and ISNT rule is the area of focus.
notching, neuroretinal rim width and peri-papillary While the CDR focuses on the optic cup size with
atrophy (PPA) serves as important indicators for pro- respect to the disc, the ISNT focuses on the neuroret-
gression of glaucoma. The neural rim area appears to inal rim width i.e., the area between the boundary of
decline with age and increase in intraocular pressure. optic cup and disc. Our work is focused in developing
As a note of caution, it has been observed that patients an automated pre-screening system to analyze the CFI
with diabetes mellitus may have an increase in the to compute the classification parameters, which on
neural rim over time, which could be secondary to integration with machine learning algorithm will help
nerve swelling. in providing an insight to the ophthalmologists to clas-
Our algorithm uses cup-to-disc ratio and neuroret- sify CFIs as glaucomatous and normal.
inal rim width based ISNT rule as classification
parameters for detection of glaucoma. The most com- This paper proposes a pixel based feature classifi-
mon clinical indicator used by ophthalmologists is the cation which analyzes multiple pixel features including
cup-to-disc ratio found out by the visual assessment of numeric properties of a pixel and it’s surrounding. Our
the optic disc through ophthalmoscopy. Higher CDR proposed algorithm has two phases: a training phase
values signify greater chances of glaucoma. However and a test phase. During the training phase, the algo-
the cup-to-disc ratio often fails to address problems rithm learns to classify pixels from known classifica-
with genetically large optic cup cases and myopic eye tions, and in the test phase the algorithm is tested on
(where the optic cup is inherently large). As a result of unknown images using machine learning which uti-
this fallacy, we additionally introduce notching estab- lizes classification factors as parameters. The optic
lished by the ISNT rule which is a technique used to disc is segmented using a statistical model, where the
measure the width of the neuroretinal rim. disc pixels are separated based on their intensity val-
ISNT stands for the four regions in which the opti- ues. Repetitive edge detection and correction tech-
cal nerve head can be divided, as shown in Fig. 2. nique is employed to suppress the blood vessels. Cup
• I = Inferior, i.e., the bottommost region. segmentation is based on average distance vector cal-
• S = Superior, i.e., the topmost region. culation on the masked disc region with respect to the
• N = Nasal, i.e., the near nose region. maximum energy pixels, for intensities greater than
• T = Temporal, i.e., the opposite of nasal region. the RMS value of the pixels. The segmentation of
For a normal eye, the inferior region of the neu- optic disc and cup is the primary requirement in
roretinal rim has the maximum width, followed by the extracting the different parameters from the fundus
superior, nasal and temporal region, following the images. The segmentation process has been elabo-
order rately discussed in our previous work [6]. The ISNT
rule is used to determine the presence and direction of
I > S > N > T. the focal notch from the width of the neuroretinal rim.
Any retinal image which violates this order is sub- The cup-to-disc ratio and the notching characteristics
jected to scrutiny, as the probability of glaucoma are the learning parameters which are integrated with
the machine learning algorithm to classify the CFIs for obtained the cup pixels via the thresholding of green
prediction of glaucoma. color plane.
The paper is organized as follows. Section 2 reviews The small bends in the blood vessels, i.e., the kinks,
previous related works on retinal feature extraction automatically mark the cup boundary. Wong [21] et al.
from CFIs and classification of the same into glauco- provided a way for detecting the cup boundary [22]
matous and normal. Section 3 presents our proposed from the kinks in blood vessels by extracting image
methods of segmentation followed by extraction of patches from the estimated cup boundary obtained in
different retinal image parameters and the classifica- [20] and the vessel pixels are identified using the edge
tion method based on them. In Section 4, the experi- and wavelet transform information. The vessel bends
ment results and the segmentation and classification identified by points of direction change in vessel pixels
accuracy is presented. Section 5 presents the discus- are detected and used to obtain the cup boundary.
sions of our works and the conclusion and future Edge detection and hough transform combined with
scopes of the paper are presented in final section. statistical deformable model can be used to extract
optic cup from vessel removed disc [23]. Gopal et al.
[24] proposed a method of cup segmentation by using
2. BACKGROUND r-bends information and by producing a local interpo-
The optic disc segmentation is the primary step for lating spline to naturally approximate cup boundaries
automatic processing of the fundus image, hence a in region where r-bends are absent. The ISNT rule is
considerable amount of research work is available. The widely used by the ophthalmologists to measure the
earliest attempt of optic disc segmentation was made neuroretinal rim width and deduce the diagnosis and
by shape based template matching in which the disc is progression of glaucoma. Tan et al. [25] proposed an
modeled as either circular [7–9] or elliptical [10]. This algorithm to classify notching based on ISNT rule and
method was unable to characterize the irregularities in also detect the feature points from the vessel bends and
shape arising from pathological changes. For this local image gradients.
problem, the approach has shifted to thresholding
based techniques. Nagakawa et al. [11] developed a P-
tile thresholding technique value in gray scale images. 3. MATERIALS AND METHODS
Gradient based active contour model are developed in 3.1. Data
which Mendels [12] approaches contour initialization
manually while Lee [13] approaches it automatically. Proposed algorithm has been trained and tested on
Recent works on active contours are more focused on a free database of colored fundus images provided by
region based approaches [14, 15] inspired by Mum- Arvind Eye Hospital in Madurai, India in collabora-
ford-Shah model [16]. The disc segmentation has tion with IIIT Hyderabad [26]. The dataset comprises
been succeeded by blood vessel extraction and removal of both normal and glaucomatous images that are of a
[11], because the presence of blood vessels makes single demographic population, as all the subjects
detection of optic cup difficult. A fuzzy algorithm was whose eye images are a part of this database are all
developed by Tolias [17] to identify the blood vessel Indians. In this dataset, the images are classified pri-
regions. Chaudhari et al. [18] presented a method of marily by four along with an additional expert oph-
detection of blood vessels by designing 12 different thalmologist. The glaucomatous images are denoted
templates intended to be used in searching for blood by +1 and the normal images as–1. The final decision
vessel segments in all possible directions. Vermeer et was taken on the basis of majority count.
al. [19] employed a thresholding segmentation based
on Laplace transform to segment retinal vessels. 3.2. Preprocessing
The optic cup segmentation poses more challenge
than disc segmentation because of the decreased visi- The preprocessing involves three steps: disc seg-
bility of the boundary and high density of blood vessels mentation, blood vessel detection, and removal and
in the boundary of optic cup and disc region. The oph- cup segmentation.
thalmologists mainly use three visual aids to detect the 3.2.1. Optic disc segmentation. The optic disc
cup, i.e., (i) change in color, that is, the cup appears to region appears brighter than the other parts of the ret-
be much paler as compared to neural rim, (ii) bend in ina in the retinal fundus image and can be segmented
small blood vessels, and (iii) notching of inner margin with a threshold obtained from a selectively chosen
of neuroretinal rim. Compared to the disc segmenta- histogram of single row profile across the fundus disc
tion very few methods have been proposed for the cup image. Our algorithm is based on statistical RGB
segmentation. Liu et al. [20] proposed a method in model using pixel based feature classification by
which a potential set of pixels from the cup region is examining relevant pixels and their surrounding,
derived based on the reference color obtained from a which is used in developing an automatic threshold
manually selected point. Then an ellipse fit to this set based on image characteristics. The row with maxi-
of pixels to obtain cup boundary, but the cup boundary mum mean pixel intensity value Rmax will pass through
thus obtained is coarse. A variant of this method the optic disc.
Fig. 3. (a) Binary disc mask, (b) red channel for disc, (c) Fig. 4. (a) Blood vessel profile in RGB, (b) blood vessel
disc mask in RGB suppressed in red layer, (c) blood vessel suppressed disc
mask.
(a) (b)
S
T N S-distance
I-distance
I
TP + FP
TP Fig. 8. Qualitative measure of segmentation accuracy in
Recall Factor = , optic disc.
TP + FN
TP + FP
Specificity (SPEC)= TN ,
TN + FP
Sensitivity (SENS)= TP , 82
TP + FN
TP + TN
Accuracy (ACC) = .
TP + FP + TN + FN 97
Table 1. Precision factor, recall factor and Fscore of the images shown in Fig. 8
Image Disc Cup
number precision factor recall factor Fscore precision factor recall factor Fscore
05 0.9556 0.9904 0.9727 0.9306 0.9359 0.9332
58 0.8799 0.9960 0.9343 0.8492 0.9349 0.8900
72 0.9728 0.9726 0.9727 0.9613 0.6988 0.8099
82 0.9162 0.9983 0.9555 0.9069 0.7366 0.8129
97 0.9495 0.9548 0.9521 0.9662 0.6726 0.7931
Mean 0.9348 0.9824 0.9575 0.9228 0.7957 0.8478
Range 0.0929 0.0435 0.0384 0.1170 0.2637 0.1401
Std. Dev. 0.0369 0.0184 0.0161 0.0477 0.1295 0.0607
area CDR, diameter CDR, notch factor, S-distance, eters along with the traditional CDR parameters helps
I-distance and notch prediction, achieving sensitivity to improve the accuracy of the algorithm for detection
of 86.42% and specificity of 90%, with an accuracy of of glaucoma.
87.128%. The integration of the focal notching param-
5. DISCUSSIONS
Table 2. Confusion matrix representation for images with
only CDR The dataset contains certain images where the
True
glaucoma phenomenon is not so vivid, and the deci-
sion of whether that CFI can be called as glaucoma-
condition
tous or normal varies among the experts. In such
total cases, the algorithms decision is validated against the
glaucoma normal opinions of at least two of the five expert ophthalmol-
population
ogists, so that the algorithms classification can fetch a
Predicted Predicted TP FP PREC six opinion going in line with those two varying expert
condition condition 62 7 89.85% decisions. Thus, our algorithm successfully evaluates
Glaucoma 76.54% 35% 88 of 101 images present in the database, hereby etch-
ing out a cumulative success rate of 87.128%. Also,
Predicted FN TN NPV there are observer variations in evaluation of CDR.
condition 19 13 40.62% The difference between the estimated cup to disc ratio
Normal 23.46% 65% for the same eye at different times seldom exceeds 0.2.
SENS. SPEC. ACC Hence, any difference greater than 0.2 over time
should be viewed with suspicion.
76.54% 65% 74.25%
The technique developed in our work has the fol-
lowing advantages over the existing techniques. Pro-
Table 3. Confusion matrix representation for images with posed method considered all the RGB layers for the
CDR and notching processing of CFIs, instead of the vastly used grayscale
True approximation which involves loss of information.
condition Furthermore, we have applied an image property
based thresholding approach to correctly extract the
total retinal features. Other works have used shape based
glaucoma normal
population template matching technique which is mostly database
specific, and also image specific. Also, we have used
Predicted Predicted TP FP PREC linear SVM based machine learning algorithm which
condition condition 70 2 97.22% makes our method adaptive, instead of being database
Glaucoma 86.4% 10% specific.
Predicted FN TN NPV Finally, we have used a unique method of detecting
condition 11 18 62.06% a focal notch to corroborate the presence of glaucoma,
Normal 13.6% 90% along with the traditional CDR evaluation techniques.
SENS. SPEC. ACC
This addition of notching parameters increases our
algorithm classification accuracy, which is clearly
86.42% 90% 87.128% depicted by the Tables 2 and 3.
23. F. Yin, J. Liu, D. W. K. Wong, N. M. Tan, C. Cheung, from the optic nerve head analysis,” JSM Biomed. Im-
M. Baskaran, T. Aung, and T. Y. Wong, “Automated aging Data Pap. 2 (1), 1004 (2015).
segmentation of optic disc and optic cup in fundus im- 27. H. Scharr, PhD Thesis (Ruprecht-Karls-Universität,
ages for glaucoma diagnosis,” in Proc. 2012 25th Inter- Heidelberg, 2000).
national Symposium on Computer-Based Medical Sys-
tems (CBMS), Rome, Italy (IEEE, Piscataway, 2012), 28. N. Harizman, C. Oliveira, A. Chiang, C. Tello,
pp. 1–6. M. Marmor, R. Ritch, and J. M. Liebmann, “The
ISNT rule and differentiation of normal from glauco-
24. G. D. Joshi, J. Sivaswamy, and S. Krishnadas, “Optic matous eyes,” Arch. Ophthalmol. 124 (11), 1579–1583
disk and cup segmentation from monocular color reti- (2006).
nal images for glaucoma assessment,” IEEE Trans. 29. M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and
Med. Imaging 30 (6), 1192–1205 (2011). B. Scholkopf, “Support vector machines,” IEEE Intell.
Syst. Their Appl. 13 (4), 18–28 (1998).
25. M. H. Tan, Y. Sun, S. H. Ong, J. Liu, M. Baskaran,
T. Aung, and T. Y. Wong, “Automatic notch detection
in retinal images,” in Proc. 2013 IEEE 10th Internation- Kaushik Dutta received his B.Tech
al Symposium on Biomedical Imaging (ISBI), San Fran- degree in Electronics and Commu-
cisco, CA, USA (IEEE, Piscataway, 2013), pp. 1440– nications Engineering from Heritage
1443. Institute of Technology affiliated to
26. J. Sivaswamy, S. R. Krishnadas, A. Chakravarty, Maulana Abul Kalam Azad Univer-
G. D. Joshi, Ujjwal, and T. A. Syed, “A comprehensive sity of Technology in 2018. He has
been the joint first author to a previ-
retinal image dataset for the assessment of glaucoma ous research paper on Glaucoma
detection from CFIs. His research
interests include biomedical image
processing, computer vision, and
cognitive radio networks.
Rishav Mukherjee received the
B.Tech degree from Heritage Insti-
tute of Technology affiliated to Anindya Sen is a Professor at the
Maulana Abul Kalam Azad Univer- department of Electronics and
sity of Technology in 2018 in Elec- Communication Engineering, Heri-
tronics and Communication Engi- tage Institute of Technology, a pri-
neering. He has successfully quali- vate autonomous engineering col-
fied GATE, examination held by lege in Anandapur, Kolkata, India.
IIT. He has been the joint first He received his B.E. from Jadavpur
author to a previous research paper University, India in the year 1980,
on Glaucoma detection from CFIs. PhD from University of Minnesota,
His research interests include image Twin Cities in 1996, and got his
processing, biomedical image analy- Post-Doctoral training from Uni-
sis, networking and cognitive radio networks (CRN). versity of Chicago from 1996 to
2000. He currently holds one US
patent and thirty research paper publications. His research
interests include, medical image processing, internet of
things, artificial intelligence and VLSI design.
Shamik Kundu received his
B.Tech degree in Electronics and
Communications Engineering Somnath Majumdar is a senior con-
from Heritage Institute of Tech- sultant ophthalmologist in Kolkata,
nology affiliated to Maulana India for last 20 years. He is cur-
Abul Kalam Azad University of rently a consultant ophthalmologist
Technology in 2018. He has for Apollo Hospitals, Kolkata and
been the joint first author to a Fortis Hospitals. He had done post
previous research paper on graduation from Dr. R.P. Centre for
Glaucoma detection from CFIs. Ophthalmic Sciences, AIIMS, New
His research interests include Delhi and completed his FRCS
signal and image processing, (Edin. and Glasgow) in 2000. He is
communications, and networks. an expert in fields of Glaucoma,
Cataract, and Retinal surgery.