18-06-2014, 12:11 PM
Evolutionary algorithm based classifier parameter tuning for automatic
diabetic retinopathy grading: A hybrid feature extraction approach
Evolutionary algorithm based classifier parameter.pdf (Size: 1.4 MB / Downloads: 20)
a b s t r a c t
Human eye is one of the most sophisticated organ, with retina, pupil, iris cornea, lens and optic nerve.
Automatic retinal image analysis is emerging as an important screening tool for early detection of eye
diseases. Uncontrolled diabetes retinopathy (DR) and glaucoma may lead to blindness. DR is caused by
damage to the small blood vessels of the retina in the posterior part of the eye of the diabetic patient.
The main stages of DR are non-proliferate diabetes retinopathy (NPDR) and proliferate diabetes retinopathy
(PDR). The retinal fundus photographs are widely used in the diagnosis and treatment of various eye
diseases in clinics. It is also one of the main resources used for mass screening of DR. We present an automatic
screening system for the detection of normal and DR stages (NPDR and PDR). The proposed systems
involves processing of fundus images for extraction of abnormal signs, such as area of hard exudates, area
of blood vessels, bifurcation points, texture and entropies. Our protocol uses total of 156 subjects consisting
of two stages of DR and normal. In this work, we have fed thirteen statistically significant (p < 0.0001)
features for Probabilistic Neural Network (PNN), Decision Tree (DT) C4.5, and Support Vector Machine
(SVM) to select the best classifier. The best model parameter ® for which the PNN classifier performed
best was identified using global optimization techniques such as Genetic Algorithm (GA) and Particle
Swarm Optimization (PSO). We demonstrated an average classification accuracy of 96.15%, sensitivity
of 96.27% and specificity of 96.08% for r = 0.0104 using threefold cross validation using PNN classifier.
The computer-aided diagnosis (CAD) results were validated by comparing with expert ophthalmologists.
The proposed automated system can aid clinicians to make a faster DR diagnosis during the mass screening
of normal/DR images
Introduction
Digital photography of the retina is widely used for screening of
patients suffering from sight threatening diseases such as DR and
glaucoma [11,27,3,40]. Increase in the number of aging population,
physical inactivity and obesity are contributing factors for diabetes.
The global prevalence of diabetes is expected to rise from
2.8% in 2000 to 4.4% of the global population by 2030 [39]. Complication
of diabetes leads to DR and a leading cause of blindness in
the world. It is estimated that by the year 2010, the number of diabetic
patients worldwide will be more than 221 million [66,55].
Diabetes Retinopathy (DR) is a silent disease and may only be
recognized by the patients when changes in the retina have progressed
to a level, where the treatment becomes complicated or
nearly impossible [72]. It can be broadly classified as NPDR and
PDR depending on the presence of clinical features (microaneurysms,
haemorrhages, hard exudates, cotton wool spots or venous
loops) on the retina [14,56]. A normal retina of the eye does not
have any of the above cited features and is shown in Fig. 1a. In
the NPDR stage (shown in Fig. 1b), the disease can advance from
mild, moderate to severe stage with various levels of above said
features except less growth of new blood vessels [56]. Fig. 1c is
the typical PDR image, where the fluids sent by the retina for nourishment
trigger the growth of new blood vessels. They grow along
the retina and over the surface of the clear, vitreous gel that fills
the
Results
The features such as blood vessels area, exudates area, bifurcation
point count, LBP energy, LBP entropy, Laws mask energy and
entropies (Shannon, Kapur and Renyi) were extracted for three
classes. We have extracted twenty-five features using above mentioned
methods and only thirteen were found to be clinically significant
(shown in Table 1]. The values of the blood vessel area and
bifurcation point counts are increasing gradually from normal to
PDR stage. It can also be seen from the table that, the exudates area
is zero for all the normal class indicating absence of exudates in
them. The LBP, LTE, and Entropy features are mostly high in PDR
compared to NPDR due to higher variability in the pixel intensities.
The statistical significance was analyzed using ANOVA test. The
results of the ANOVA test of extracted features are shown in Table
1 and Fig. 13. It can be seen from Table 1 that, our features are statistically
significant (p < 0.0001). The graphical representation of
the features of three classes is shown in Fig. 13a. We can clearly
understand from the box plot that, median values are distinct for
each group for the feature Bifurcation point. The 3D scatter plot
shown in Fig. 13b depicts the variation of LTE9, LTE14 and Shannon
entropy for normal, NPDR and PDR classes.
Discussion
Several studies have been reported in the automated classification
of DR images. In this section, we have summarized the classification
of two, three, four and five classes of digital fundus images
(normal and DR classes).
5.1. Two-class (normal and DR) analysis
The performance of lesion detection algorithm was evaluated
using an automated fundus photographic image-analysis algorithm
subjects with DR and subjects without DR [28]. The automated
lesion detection correctly identified 90.1% retinopathy
patients and 81.3% subjects without retinopathy.
The retinal thickness analyzer (RTA) was found to be suitable
for application in tele-screening of DR. It yielded a mean sensitivity
Conclusion
Diabetes retinopathy and glaucoma are the leading cause of
blindness in the world. In this study, advanced image processing
and machine learning algorithms were used to pre-process retinal
digital images, extract the features and classify the retinal images
to Normal, NPDR and PDR classes. Our experimental results show
that the proposed method yields an average accuracy of 96.15%,
sensitivity of 96.27%, and specificity of 96.08%. The higher accuracy
obtained by the PNN classifier is due to the selection of best model
parameter (r = 0.0104) using the evolutionary algorithms (GA and
PSO) in addition to the combination of morphological and texture
features. The proposed classification model was compared with
state-of-art classification models such as decision tree C4.5 and
SVM, which yielded an accuracy of 88.46% and 77.56% respectively.