Hierarchical Classifier For Microaneurysm Detection

7 downloads 102 Views 532KB Size Report
To design modules like walter klein enhancement, histogram equalization, ..... Olson, 2010, “Automated grading for diabetic retinopathy: A large-scale.
International Journal of Applied Engineering Research ISSN 0973-4562 Volume 10, Number 1 (2015) pp. 1449-1458 © Research India Publications http://www.ripublication.com

Hierarchical Classifier For Microaneurysm Detection M. Kavitha Associate Professor, Department of ECE, Mookambigai College of Engineering, Kalamavur Pudukottai District E-mail: [email protected] Dr. S. Palani Director & Professor, ECE, Sudharsan Engineering College, Pudukottai District E-mail:[email protected]

Abstract Diabetic-related eye diseases are the commonest cause of vision defects and blindness in the world. Monitoring the health of the retina is important for those people with signs of diabetic retinopathy. People affected by diabetic retinopathy may get microaneurysms in the retina, a condition in which blood vessels of the retina are damaged by diabetes, which can lead to blindness. Microaneurysm[MA] is defined as a small swelling that forms in the wall of tiny blood vessels. From the abnormal retinal image, we have detected the microaneurysms using Klein contrast enhancement, tophat filtering and median filtering. Then the features are extracted with the aid of variance, crosscorrelation and co-variance features. Once the features are computed, training of Levenberg-Marquard based neural network is done to classify microaneurysm or nonmicroaneurysm. Accuracy is calculated to find out effectiveness of the Levenberg-Marquard based neural network and comparision is done with the other existing classifiers. Here the experimentation is done using messidor database. The proposed new method clearly shows better accuracy compared with other existing methods. Keywords:- Diabetic, Retinal Image, Microaneurysm, Neural Network, Segmentation, Feature Extraction, Median filtering

1450

M. Kavitha, Dr. S. Palani

1.Introduction Diabetic Retinopathy (DR) is a severe eye disease that occurs because of diabetes which plays a major role in making people blind and to have vision defects in the modern world. With reference to some investigations done in the past, it has identified microaneurysms in fundus images based on their gray level [7], their high contrast or their color [8]. The brightness, contrast and color of microaneurysms vary a lot among different patients in different photographs and therefore, these methods would not work in all the images used in clinical environment. The main purpose of feature selection is to select a number of features used in the classification and at the same time to maintain acceptable classification accuracy [9]. This technique will emphasize on the steps like contrast enhancement, image enhancement, median filtering, ensemble creation, tophat transform and image subtraction. The main contribution of the proposed technique is:  To detect microaneurysms in a person from retinal image.  To design modules like walter klein enhancement, histogram equalization, image enhancement, image subtraction and tophat filtering.  To train the Neural Network for the final classification.  To conduct experiment using the Messidor database to prove the significance of the proposed technique.  To calculate accuracy to find out effectiveness of the proposed technique and compare with other existing classifiers such as neural network with backpropagation and K nearest neighbor classifier(KNN). The rest of this paper is organized as follows. Section 2 gives a brief description of the related works. The details of the proposed method is discussed in section 3. We summarize our experimental results in section 4. Finally, conclusion is drawn in section 5.

2.Related Works Spencer et al.reported 49% sensitivity with 0.2 false positives per image. However the test database involved only a few images, all with a low grade of retinopathy [4]. Cree et al.improved the region growing, added to the measured feature set, redesigned the classifier and achieved 82% sensitivity at 5.7 false positives per image with a much more realistic suite of test images [1]. The above two studies were performed on angiographic images which are unsuitable for screening purposes. Prasad et al.[2] treated MAs and hemorrhages as holes and morphological filling is performed on the green channel to identify them. The unfilled green channel image is then subtracted from the filled one and fix the threshold value to yield a resultant image with MA patches. To remove noise vessel segments, the full blood vessel network skeleton is dilated and subtracted from the resultant image. The remaining patches are further classified using intensity properties and a color model method based on the detected blood vessels.

Hierarchical Classifier For Microaneurysm Detection

1451

Ege et. al. [3] described a tool for the automatic analysis of digital retinal images, and evaluated classifiers such as Bayes classifier, Mahalanobis classifier and KNN on an array of features such as log(area), minor axis variance, major axis variance, etc. Among them, the Bayes classifier and KNN classifier showed promising results, but their approach requires a pre-processing step that involves mean filtering and thresholding. Small microaneurysms are removed during the filtering step, while thresholding is not effective given the variations in background intensity. Spencer, Cree, Frame, and co-workers [5] added a shade-correction preprocessing step and a matched filtering post-processing step to the basic tophat transform based detection technique. After detection and segmentation of candidate microaneurysms, various shapes and intensity based features were extracted and a classifier was used to separate the real microaneurysms from spurious responses. The advantage of using fluorescein angiography images of the fundus is that the contrast between the image and the retinal background is larger than in digital color photographs. Niemeijer et al. [6] presented a hybrid scheme that used both the tophat based method as well as a supervised pixel classification based method to detect the microaneurysm candidates in color fundus photographs. The pixel classification method was trained using example pixels from both the vasculature and “red lesions” (i.e. microaneurysms and hemorrhages). After training, the detector would detect all retinal vessel and red lesion pixels in an image. After eliminating all connected, elongated structures the remaining objects were considered as candidate microaneurysms. This method allowed for the detection of larger “red lesions” (i.e. hemorrhages) in addition to the microaneurysms using the same system. A classifier, trained with example candidates from a training set, distinguished between real and spurious candidate lesions.

3.Proposed Technique For Detection Of Microaneurysm Through Retinal Image In this section, we describe the proposed technique for effectively detecting the presence of microaneurysms in a patient by examining the input retinal images. Proposed technique makes use of klein, tophat transform and neural network to accomplish the task and bring out with better results when compared to the other algorithms. The technique consists of four phases namely, image enhancement, tophat transform, feature extraction and microaneurysm detection using neural network. Block diagram representation of the proposed method is shown in Figure 1.

1452

M. Kavitha, Dr. S. Palani

ORIGINAL IMAGE

APPLY TOP HAT FILTERING AND IMAGE SUBTRACTION

FIX THE THRESHOLD VALUE

KLEIN CONTRAST ENHANCEMEN T ENSEMBLE CREATOR

HISTOGRAM EQUALISATION IMAGE ENHANCEMENT

&

IMAGE SUBTRACTION( FILTERED – GRAY)

GRAY SCALE IMAGE

PERFORM MEDIAN FILTER

MICROANEURYS M DETECTION & CLASSIFICATION

Figure 1. Block diagram representation of the proposed method The functions of blocks of Figure1 are described below. To enhance the contrast of the fundus image, walter klein method of enhancement is used. The image is split into disjoint regions, and in each region histogram equalization is applied. Histogram equalisation computes the probability of occurrence of each of the 256 gray levels in the image. Complete vessel system is being removed based on image enhancement so that MAs appearing near vessels become more easily detectable. The enhanced image is converted into grayscale image. The grayscale image is then applied to the median filtering such that it simultaneously reduces noise and preserves edges. Next the filtered image is subtracted with the gray scale image. Ensemble creation framework is provided to select the best combination followed by the tophat filtering. It performs morphological operation i.e erosion followed by dilation. Threshold value is fixed in order to detect the microaneurysm. The features are extracted from the detected microaneurysm. Classification of microaneurysm and non microaneurysm is done using neural network. 3.1 Walter–Klein Contrast Enhancement: This preprocessing method aims to enhance the contrast of fundus images by applying a gray level transformation using the following operator: Steps to be followed: 1. Determine the mean value for the retinal image. 2. Find maximum and minimum frequency component for the image. 3. To find the intensity value use the following conditions. m=mean value; if pixm

(1)

Hierarchical Classifier For Microaneurysm Detection pixE = ((1/2*(fmax-fmin))/(m-fmax)^2)*(pix-fmax)+fmax;

1453 (2)

3.2 Histogram equalisation: This method usually increases the global contrast of many images, especially when the usable data of the image is represented by close contrast values. Through this adjustment, the intensities can be better distributed on the histogram. This allows for areas of lower local contrast to gain a higher contrast. Histogram equalization accomplishes this by effectively spreading out the most frequent intensity values. The method is useful in images with backgrounds and foregrounds that are both bright or both dark. It is very effective in making the usually interesting salient parts more visible.. For vessel extraction image enhancement is used. It highlight the certain characteristics of the image. The result is more suitable than the original image for a specific application. 3.3 Median filter: The median filter is a nonlinear digital filtering technique, often used to remove noise. Such noise reduction is a typical pre-processing step to improve the results of later processing (for example, edge detection on an image). Median filtering is very widely used in digital image processing because, under certain conditions, it preserves edges while removing noise. Specifically, the median filter replaces a pixel by the median, instead of the average, of all pixels in a neighborhood w. The output of the image is described by the following equation[13 ]

(3) where w represents a neighborhood defined by the user, centered around location (m,n) in the image. y[m,n] is the output image and x[i,j] is the input image. 3.4 Ensemble creation: Ensemble creation is a process where all ensembles E from an ensemble pool E is evaluated and the best performing one Ebest ∈ E regarding an evaluation function on a training set is selected. To evaluate such a large number of combinations, we used simulated annealing [20] as a search algorithm to find the final ensemble, which is proven to be effective in such large search spaces. This ensemble Ebest can be used to detect MAs on retinal images. 3.5 Tophat filtering: Tophat filtering computes the morphological opening of the image (using imopen) and then subtracts the result from the original image. Imtophat uses the structuring element SE. SE must be a single structuring element object, not an array containing multiple structuring element objects.

1454

M. Kavitha, Dr. S. Palani

3.6 Neural network(classification): Once the features are computed, training of Levenberg-Marquardt-based neural network is done to classify microaneurysm or nonmicroaneurysm. Now, LevenbergMarquardt-based neural network is employed to do hierarchical classification and it is trained based on the features extracted from the microaneurysm. Training: In the training phase, the feature is extracted from the retinal image and this feature vector is specified as the input to the neural network. At first, the nodes are specified as random weights. As the output is previously known in the training phase, the output attained from the neural network is compared to the original and weights are differed by means of Levenberg-Marquardt algorithm so as to decrease the fault. This process is carried for a large number of images so as to give up a steady system having weights allocated in the nodes. Testing: In the testing stage, the input image is fed to the trained neural network having particular weights in the nodes and the output is computed so as to discover if the abnormal retinal image is microaneurysm or non microneurysm..

4. Results and Discussion This section presents the experimental results of the proposed technique with the images of Messidor database. The retinal images obtained from Messidor database are sufficient but variations in color, illumination and quality. Our analysis was implemented in Matlab 7.12 running on an Intel (R) Core i5 processor, 3.20 GHz with 4 GB RAM. 4.1 Dataset Description: The images are digitized slides captured by a Top Con TRV-50 fundus camera with 35 degree field of view. Each slide was digitized to create a 605 × 700 pixel image with 24 bits per pixel. All the images were vigilantly labeled by hand to develop ground truth vessel segmentation by a specialist. 4.2 Experimental Results This section shows the experimental results of the proposed technique. Figure.2 shows the experimental result of the abnormal retinal image which is the input. Walter klein method for image enhancement, histogram equalisation for intensity adjustment, image enhancement for vessel extraction and median filtered image to reduce the noise and preserve the edges followed by image subtraction are the outputs. Finally microaneurysm detection is done. Our analysis was implemented in Matlab 7.12 running on an Intel (R) Core i5 processor, 3.20 GHz with 4 GB RAM.

Hierarchical Classifier For Microaneurysm Detection

1455

Sample digital fundus image with microaneurysm

Water klein enhancement

Histogram equalisation

Image enhancement

filtered image

Image subtraction

Microaneurysm detection Figure 2. Experimental result of micraneurysm detection

1456

M. Kavitha, Dr. S. Palani

4.3 Evaluation metrics of classification phase: In this section, the microneurysm and non microneurysm classification is done with the help of three measures like sensitivity, specificity and accuracy. Here, our proposed classifier of Levenberg-Marquardt (LM) neural network is compared against two other classifiers in this section. For comparison, neural network with back propagation and K-nearest neighbor are used. Accuracy: The common classification criterion or accuracy is defined as [12] TF  TN Accuracy  TF  TN  FP  FN (4) Sensitivity: The measure of the sensitivity is the proportion of actual positives which are properly classified. It relates to the capacity of test to classify positive results and is expressed as [ 12 ] TP Sensitivity   100 TP  TN (5) Specificity: The measure of the specificity is the proportion of negatives which are properly classified. It relates to the capacity of test to classify negative results and is given by [12 ] TN Specificit y  100 TN  FP (6) where, TF Total number of normal classes correctly classified as positive (true positive) TN Total number of abnormal classes truly classified as negative (True negative) FP Total number of abnormal classes falsely classified as positive (false positive) FN Total number of normal classes falsely classified as negative (false negative) 4.4 Comparative Analysis: Our analysis was implemented in Matlab 7.12 running on an Intel (R) Core i5 processor, 3.20 GHz with 4 GB RAM. Once the segmentation is carried out and features like variance, crosscorrelation and co-variance are extracted, classification of microaneurysm is made with the aid of neural networks. The comparative analysis of the proposed technique with the existing methods is described in this Table 1. It is compared with neural network with back propagation, where back propagation algorithm is used for neural network. It is also compared with KNN which stands for K nearest neighbour method which is a nonparametric technique for classifying objects based on closed training examples in the feature space. Accuracy of microaneurysm detection is found out using Matlab 7.12 version and for the proposed method it is found to be 90% compared to 80.3% and 86.79% obtained for neural network with backpropagation and KNN classifier respectively. From the results, we can infer that the proposed technique achieved high accuracy value when compared to the accuracy of the existing classifiers. This shows the effectiveness of the proposed technique.

Hierarchical Classifier For Microaneurysm Detection

1457 Messidor database

Classification Methods Proposed Method Neural network with Back propagation KNN

Accuracy (%) 90.0 80.3 86.79

Table 1 Comparative analysis

5. Conclusion In this paper, a new efficient technique is presented to detect the presence of microaneurysm in a person from the retinal images automatically. The proposed technique comprises of four phases of pre-processing, segmentation, feature extraction and final classification using neural network. The proposed technique makes use of messidor database and features like variance, crosscorrelation and covariance. Employment of Levenberg-Marquardt based neural network results in having an improved system for detecting microaneurysm. Accuracy is found out to evaluate the proposed technique and compared with neural network with back propagation and KNN method. The proposed method clearly shows better accuracy compared with other existing methods.

6. References [1]

[2]

[3]

[4]

[5]

A. D. Fleming, K. A. Goatman, S. Philip, G. J. Prescott, P. F. Sharp, and J. A. Olson, 2010, “Automated grading for diabetic retinopathy: A large-scale audit using arbitration by clinical experts,” Br. J. Ophthalmol., vol. 94, no. 12, pp. 1606–1610.. S. Prasad, A. Jain, and A. Mittal, 2009,“Automated feature extraction for early detection of diabetic retinopathy in fundus images,” in Proc. IEEE Conf. Comput. Vision Pattern Recog., pp. 210–217. Ege BM, Hejlesen OK, Larsen OV, Moller K, Jenning B, Kerr D., 2000," Screening for diabetic retinopathy using computer based image analysis and statistical classification." conference on Computer Methods Programm Biomedical,vol.165,pp.62-75. T. Spencer, J. A. Olson, K. C. McHardy, P. F. Sharp, and J. V. Forrester, 1996, “An image-processing strategy for the segmentation and quantification of microaneurysms in fluorescein angiograms of the ocular fundus,” Comput.Biomed. Res., vol. 29, pp. 284–302. M. Niemeijer, B. van Ginneken, M. Cree, A. Mizutani, G. Quellec, C. Sanchez, B. Zhang, R. Hornero, M. Lamard, C. Muramatsu, X. Wu,G. Cazuguel, J. You, A. Mayo, Q. Li, Y. Hatanaka, B. Cochener, C. Roux,F. Karray, M. Garcia, H. Fujita, and M. Abramoff, 2010, “Retinopathy

1458

[6]

[7]

[8]

[9]

[10] [11] [12] [13]

M. Kavitha, Dr. S. Palani onlinechallenge: Automatic detection of microaneurysms in digital color fundus photographs,” IEEE Trans. Med. Imag., vol. 29, no. 1, pp. 185–195.. M. Abramoff, J. Reinhardt, S. Russell, J. Folk, V. Mahajan, M. Niemeijer, and G. Quellec, 2010 ,“Automated early detection of diabetic retinopathy,”Ophthalmology, vol. 117, no. 6, pp. 1147–1154. M. Niemeijer, M. Loog, M. D. Abramoff, M. A. Viergever, M. Prokop,and B. van Ginneken, 2011 ,“On combining computer-aided detection systems,”IEEE Trans. Med. Imag., vol. 30, no. 2, pp. 215–223. N. P. Ward, S. Tomlinson, and C. J. Taylor, 1989, “Image analysis of fundus photographs – The detection and measurement associated with diabetic retinopathy,” Ophthalmol., vol.96, pp. 80–86. R. Philips, J. Forrester, and P. Sharp, 1982 ,“Automated detection and quantification of retinal microneurysms,” Graefe’s Arch. Clin. Exp.Ophthalmol., vol. 231, pp. 90–94, 1993. K. Akita and H. Kuga, “A computer method of understandingocular fundus images,” Pattern Recogn., vol. 21, no. 6, pp. 431–443. H. Li and O. Chutatape, 2000, “Fundus image features extraction,” IEEE transaction on Engineer and Medical Biol. Soc., pp. 3071–3073.. L. I. Kuncheva, 2004 “Combining Pattern Classifiers. Methods and Algorithms.” Hoboken, NJ: Wiley. K. Zuiderveld, 1994 “Contrast limited adaptive histogram equalization,” Graphics Gems, vol. 4, pp. 474–485.