Retinal Image Preprocessing: Background and Noise Segmentation

Medical imaging is very popular research area these days and includes computer aided diagnosis of different diseases by taking digital images as input. Digital retinal images are used for the screening and diagnosis of diabetic retinopathy, an eye disease. An automated system for the diagnosis of diabetic retinopathy should highlight all signs of disease present in the image and in order to improve the accuracy of the system, the retinal image quality must be improved. In this article, we present a method to improve the quality of input retinal image and we consider this method as a preprocessing step in automated diagnosis of diabetic retinopathy. The preprocessing consists of background estimation and noise removal from retinal image by applying coarse and fine segmentation. We perform extensive results to check the validity of proposed preprocessing technique using standard fundus image database.


Introduction
One out of ten is affected by diabetes, which result in vision loss, stroke and heart failure.Group of eye problems that people with diabetes may face is known as diabetic eye disease.People affected by diabetic are more facing eye problems like contracts and glaucoma and the main reason for vision loss is the disease affecting on retina [1].The complications of diabetes affect the vascular structure of human retina and cause the leakage of blood on surface of retina which leads to blindness and it is known as diabetic retinopathy [1].Diabetic retinopathy is a progressive disease as there are no such signs of disease at its early stages but as the time passes the disease turns into severe [2].The common symptoms of diabetic retinopathy are blurred vision and even loss of vision if not treated in time [2].
By using retinal image we can check whether a patient suffers from diabetic retinopathy or not.The regular screening of diabetic retinopathy results in a large number of retinal images that need to be examined by the ophthalmologists.Since the cost of manual examination is high so the best solution is an automated screening system for retinal images [1].This system can be differentiating between normal and abnormal retinal images and will result in reducing the workload for the ophthalmologists.Since the ophthalmologists will examine only the abnormal images which is diagnosed by the system [3].
Healthy retina contains blood vessels, optic disc, macula and fovea as main features.An automated system for screening and diagnosis of diabetic retinopathy should detect all these features automatically [5], [6], [7] and all signs of diabetic retinopathy such as microaneurysms [4], [8], [9], hemorrhages, edema [4], hard exudates and cotton wool spots [10][11].The accuracy of all these depends on the quality of acquired retinal image as some times it contains uneven illumination, blurry and noisy areas.The center region of a retinal image is usually highly illuminated while the noise increases closer to the edge of the retina [17].So, Illumination equalization and noise removal are required to enhance the image quality.
In the beginning, before the detection of abnormalities and features in retinal image we must remove the noise and background from retinal image which will increase the quality of the image.This is done in preprocessing step and without this step the automated system will give poor result for feature extraction and abnormality detection.
The aim of preprocessing is to increase the quality of an image by reducing the amount of noise appearing in the image and highlighting features that are used in image segmentation.Two typical techniques used in preprocessing are filtering and contrast enhancing.Standard contrast stretching techniques have been applied by [4], [13] and [23] for segmentation and noise reduction.In [14], [15] and [16] the local contrast enhancement method is used for equalizing uneven illumination in the intensity channel of retinal images.A large mean filter, large median filter and collectively are used for retinal image have used intensity channel values to detect the dark regions from retinal image.Thresholding is also an important and widely used technique in image segmentation [19], because thresholding is effective and simple to implement.In thresholding, pixels within a defined range are selected as belonging to the foreground whereas gray-levels outside the range are rejected to the background [19].
In this paper, we present the retinal image preprocessing technique that detects the background using local mean and variance and removes noise using HSI color space.The proposed preprocessing method consists of two steps i.e. coarse and fine segmentation.In the first step, it does coarse segmentation that creates binary masks for background estimation and noise removal.In the second step, it does fine segmentation that combines background segmented mask and noise segmented mask and applies morphological operations to remove single pixel noise and edge pixels.
This paper is organized in four sections.Section 2 presents the step by step techniques required for color retinal image segmentation.Experimental results are discussed in section 3 followed by conclusion in section 4.

System Overview
Computer assisted diagnosis for various diseases is very common now a days and medical imaging is playing a vital role in such computer assisted diagnosis.We present an automated preprocessing system to improve the quality of retinal images so that the accuracy of automated screening of diabetic retinopathy can be improved.The proposed preprocessing method is used to improve the quality of image by extracting background and removing the noisy area from the retinal image.In automatic diagnosis of diabetic retinopathy, the processing of the surrounding background and noisy areas in retinal image is not necessary and consumes more processing time in all stages.Cutting or cropping out the region that contains the retinal image feature minimizes the number of operations on the retinal image.Moreover the noisy pixels may cause false features and reduce the accuracy of the automated classification so it is necessary to remove the noisy pixels.Figure 1 shows the flow diagram of our preprocessing technique.It shows the step by step outputs of proposed preprocessing technique.

Coarse Segmentation
Coarse segmentation creates mask for background and noise estimation using mean and variance method and HSI (Hue, Saturation, Intensity) color model respectively. .

TELKOMNIKA
Retinal Image Preprocessing: Background and Noise Segmentation Figure 1.

Background Estimation
A color retinal image consists of a (semi) circular region of interest background.This dark background is initially never really black.feature extraction and lesion detection algorithms on this area and it consumes more time so it is important to remove the background from input retin We present a local mean creates a binary background segmentation mask greater than threshold value, the block is considered as belongs to background.
The algorithm for the background extraction mask is as below: Step 1: Divide the acquired retinal image into non Step 2: Compute the local mean value Step 3: Use local mean value computed in step 2 to compute the local standard deviation value std(I) using equation 2 Step 4: Select threshold value empirically Step 5: for each pixel, Calculate std(I)> if true, add pixel in original retinal image area pixels if false, add pixel in background area pixels end_for

Noise Segmentation Mask
Noise in color retinal image is normally due to noise pixels distorted.Both seem to exist in regions where illumination has been inadequate.illumination is usually adequate in the center of the image, poor image quality regions are located near the edge of the retinal image.Region abnormality detection.That is why they should be detected and removed before detection of abnormalities.
In our technique, we create binary noise segmentation mask which includes the noisy area and it is applied on retinal image to ensure not to process the noisy area in upcoming steps i.e. feature extraction and abnormality detection.In this segmentation technique, we convert RGB (Red, Green, Blue) retinal image into HSI color space because firstly it is the way a human experiences colors and secondly noise can be easily removed in HSI color space [19] and pixels whose color is distorted.Both seem to exist in regions where illumination has been inadequate.Since illumination is usually adequate in the center of the image, poor image quality regions are s with poor image quality may cause errors in abnormality detection.That is why they should be detected and removed before detection of In our technique, we create binary noise segmentation mask which includes the noisy pplied on retinal image to ensure not to process the noisy area in upcoming steps i.e. feature extraction and abnormality detection.In this segmentation technique, we convert RGB (Red, Green, Blue) retinal image into HSI color space because firstly it is closer to the way a human experiences colors and secondly noise can be easily removed in HSI color SI color space using equations 3, 5 and 6 .
where here R, G and B represent RED, GREEN and BLUE components of RGB retinal image.
The algorithm for noise removal mask is as below: Step 1: Divide the input retinal image Step 2: Use histogram equalization to enhance the contrast Step 3: Use a 3x3 median filter to reduce the noise in background of image.
Step 4: Convert the equalized and filtered RGB retinal image into Step 5: Calculate N (noise factor) which is a ratio of Hue and Intensity Step 6: Select a threshold value empirically.
Step 7: for each pixel, Calculate N(I)<Threshold?if true, add pixel in normal retinal image area pixels if false, add pixel in noisy area pixels end_for Figure 2 shows the images.These segmentation masks contain single pixel and edge pixel noise which will be removed in fine segmentation.

Fine Segmentation
Background and noise segmentation masks that are formed by coarse segmentation contain single pixel noise and edge pixels.Fine segmentation is done to remove these noises from segmentation masks.In fine segmentation, morphological operations i.e. morphological dilation, morphological erosion and morphological opening are applied to remove noise from binary masks [19].We have used 5 x 5 square structuring element for morphological operations [19].Background segmentation mask contains black single pixel noise ( order to remove the black pixel noise, square structur removes all black single pixel noise and edge pixels and it gives a fin segmentation mask.Noise segmentation mask contains white single pixel noise and edge pixels (Figure 2).In order to remove the w opening followed by erosion.Opening removes all edge pixels and erosion removes all white single pixel noise and it gives a fi segmentation masks and fine segmentation masks.Background and noise segmentation masks that are formed by coarse segmentation single pixel noise and edge pixels.Fine segmentation is done to remove these noises from segmentation masks.In fine segmentation, morphological operations i.e. morphological dilation, morphological erosion and morphological opening are applied to remove noise from binary masks [19].We have used 5 x 5 square structuring element for morphological operations [19].Background segmentation mask contains black single pixel noise ( order to remove the black pixel noise, square structuring element is used for dilation.Dilation removes all black single pixel noise and edge pixels and it gives a fin Noise segmentation mask contains white single pixel noise and edge pixels ).In order to remove the white pixel noise, square structuring element is used for opening followed by erosion.Opening removes all edge pixels and erosion removes all white single pixel noise and it gives a fine noise segmentation mask.( overlapping blocks with size w x w.
HSI color space.
noise segmentation masks for retinal segmentation masks contain single pixel and edge pixel noise which will be Coarse segmentation.a) Original retinal image; b) Background segmentation mask; c) Background and noise segmentation masks that are formed by coarse segmentation single pixel noise and edge pixels.Fine segmentation is done to remove these noises from segmentation masks.In fine segmentation, morphological operations i.e. morphological dilation, morphological erosion and morphological opening are applied to remove single pixel noise from binary masks [19].We have used 5 x 5 square structuring element for morphological operations [19].Background segmentation mask contains black single pixel noise (Figure 2).In ing element is used for dilation.Dilation removes all black single pixel noise and edge pixels and it gives a fine background Noise segmentation mask contains white single pixel noise and edge pixels hite pixel noise, square structuring element is used for opening followed by erosion.Opening removes all edge pixels and erosion removes all white

Final Segmentation Mask
Final segmentation mask is prepaid by combining fine background and fine noise segmentation mask.For more fine segmentation mask, small regions are removed by filtering the combined mask by a medium size median filter [19].Final noise free segmentation mask is then applied on retinal im final segmentation masks for retinal image segmentation.

Experimental Results
In medical image processing, the validity and evaluation of automated diagnostic systems are very important so we have used four standard retinal image databases i.e. diaretdb0 [12], diaretdb1 [20], DRIVE [21] and STARE [22] to extensively check the propos method.Diaretdb0 and DiaretDB1 d resolution of 1500 X 1152 pixels and of different qualities in terms of noise and illumination.DRIVE and STARE databases consist of 40 and 20 images respectively.We hav objective and subjective criteria for evaluation of proposed method.segmented retinal images and it depicts the subjective validity of proposed method.
Statistical results of our segmentation technique table 1 and table 2. Table 1 separately shows the accuracy of background segmentation mask, noise segmentation mask and final segmentation mask.segmentation, fine segmentation after applying morphological operations and final segmentation after applying median filter.These tables show the number and percentage of accurate segmented retinal images and poorly segmented evaluate the proposed system, manually labeled preprocessing masks are created by the two experts and these manually labeled masks act as ground truth for accuracy calculations.Table 3 shows the accuracy results of prop segmentation masks.Retinal images of different illumination and noise values are shown in shows the background segmentation masks, noise segmentation masks, final segmentation masks and final segmented retinal image for each color retinal image.These results support the validity of our technique and show that our technique gives good results noisy areas.Figure 7 shows the manually created preprocessing masks and the masks created by proposed algorithm.In order to find out the accuracy, pix order to check the validity of proposed system, characteristics curves using manually labeled data as ground truth and comparing the segmented masks with it.Figure 8 shows the ROC curves for DiaretDB0 and DiaretDB1 databases.Retinal images of different illumination and noise values are shown in shows the background segmentation masks, noise segmentation masks, final segmentation masks and final segmented retinal image for each color retinal image.These results support the validity of our technique and show that our technique gives good results for both low and high shows the manually created preprocessing masks and the masks created by proposed algorithm.In order to find out the accuracy, pixel by pixel comparison is done.In order to check the validity of proposed system, we further calculate the receiver operating characteristics curves using manually labeled data as ground truth and comparing the segmented masks with it.Figure 8 shows the ROC curves for DiaretDB0 and DiaretDB1  Table 3. Accuracy results for proposed preprocessing method Retinal images of different illumination and noise values are shown in Figure 6. Figure 6 shows the background segmentation masks, noise segmentation masks, final segmentation masks and final segmented retinal image for each color retinal image.These results support the for both low and high shows the manually created preprocessing masks and the masks created el by pixel comparison is done.In we further calculate the receiver operating characteristics curves using manually labeled data as ground truth and comparing the segmented masks with it.Figure 8 shows the ROC curves for DiaretDB0 and DiaretDB1

Conclusion
The problem with retinal images is that the quality of acquired images is usually not good.They contained uneven illuminated, blurred and noisy regions.It is very vital to enhance the quality of colored retinal images for reliable detection of abnormalit presented a method for the preprocessing of retinal image prior to screening and diagnosis of diabetic retinopathy.The proposed method consisted of background estimation using mean and variance based method and noise modeling and re divided our algorithm into two phases i.e. coarse and fine segmentation.is prepaid by combining the background mask and noise segmentation mask.For further improving the quality of segmentation mask to remove small regions.Final segmented masks are then applied on poor quality retinal images for their segmentation.The results are confirmed by visual inspection of segmented images taken from the standard diabetic retinopathy databases.ROC curves to validate the proposed method.

References
[1] E J Susman, W J Tsiaras, K A Soper Receiver operating characteristic curves for proposed method using diaretDB0 and diaretDB1 databases The problem with retinal images is that the quality of acquired images is usually not good.They contained uneven illuminated, blurred and noisy regions.It is very vital to enhance the quality of colored retinal images for reliable detection of abnormalities.In this paper, presented a method for the preprocessing of retinal image prior to screening and diagnosis of diabetic retinopathy.The proposed method consisted of background estimation using mean and variance based method and noise modeling and removal using HSI color model.We further divided our algorithm into two phases i.e. coarse and fine segmentation.Final segmented mask is prepaid by combining the background mask and noise segmentation mask.For further improving the quality of segmentation mask, median filter are applied on final segmentation mask to remove small regions.Final segmented masks are then applied on poor quality retinal images for their segmentation.The results are confirmed by visual inspection of segmented the standard diabetic retinopathy databases.We further used accuracy and ROC curves to validate the proposed method.The problem with retinal images is that the quality of acquired images is usually not good.They contained uneven illuminated, blurred and noisy regions.It is very vital to enhance ies.In this paper, we presented a method for the preprocessing of retinal image prior to screening and diagnosis of diabetic retinopathy.The proposed method consisted of background estimation using mean and moval using HSI color model.We further Final segmented mask is prepaid by combining the background mask and noise segmentation mask.For further mask, median filter are applied on final segmentation mask to remove small regions.Final segmented masks are then applied on poor quality retinal images for their segmentation.
Figure 1.Flow Diagram of Retinal Image Preprocessing

Figure 2 .
Figure 2. Coarse segmentation.a) Original retinal image; b) Background segmentation mask; c) 2012 : 537 -544 here R, G and B represent RED, GREEN and BLUE components of RGB retinal image.The algorithm for noise removal mask is as below: : Divide the input retinal image I(i,j) into non-overlapping blocks with size : Use histogram equalization to enhance the contrast : Use a 3x3 median filter to reduce the noise in background of image.: Convert the equalized and filtered RGB retinal image into HSI color space.(noise factor) which is a ratio of Hue and Intensity : Select a threshold value empirically.<Threshold?, add pixel in normal retinal image area pixels , add pixel in noisy area pixels shows the coarse background and noise segmentation masks for retinal segmentation masks contain single pixel and edge pixel noise which will be segmentation.Coarse segmentation.a) Original retinal image; b) Background segmentation mask; c) Noise segmentation mask

Figure
G and B represent RED, GREEN and BLUE components of RGB retinal image.

Figure 7 :-
Figure 7: Accuracy results; row 1: Original retinal images taken from diaretdb database; row 2: Manually created masks by ophthalmologists used as ground truth; row 3: Preprocessing masks created by proposed algorithm

Figure 7 :
Figure 7: Accuracy results; row 1: Original retinal images taken from diaretdb database; row 2: Manually created masks by ophthalmologists used as ground truth; row 3: Preprocessing masks created by proposed 3234.[2] Effective Health Care -Complications of diabetes: Screening for retinopathy and Management of foot ulcers.Royal Society of Medicine Press [3] A Osareh, M Mirmehdi, B Thomas and R Exudates in Digital Colour Images [4] S C Lee, E T Lee, R M Kingsley, Y early retinal lesions of diabetic retinopathy between a co Arch.Clin.Exp.Ophtalmol.[5] C Sinthanayothin, J F Boyce, H L Cook, T. H.Williamson fovea and retinal blood vessels from digital color fundus images 83(8): 231-238.[6] M Foracchia, E Grisan, A Ruggeri. model of vessel structure.IEEE transa [7] N K M N Subhasis Chaudhuri, Shankar Chatterjee retinal images using twodimensional matched filters 8(3): 263-269.[8] T Spencer, R P Phillips, P F Sharp, J V Forrester microaneurysms in fluorescein angiograms [9] A J Frame, P E Undill, M J Cree, J A Olson, K C McHardy, P F Sharp, J F Forrester computer based classification methods applied to the detection of microaneurysms in op fluorescein angiograms.Comput.Biol.Med e-ISSN: 2087-278X Retinal Image Preprocessing: Background and Noise Segmentation The results are confirmed by visual inspection of segmented We further used accuracy and 1982; 247( 23),3231-Complications of diabetes: Screening for retinopathy and Management of foot Automated Identification of Diabetic Retinal ): 1220-1223.Comparison of diagnosis of mputer system and human experts.Graefes Automated localization of the optic disc, ournal of Opthalmol.1999; Detection of optic disc in retinal images by means of a geometrical 1189-1195.Detection of blood vessels in transactions on medical imaging.1989; Automated detection and quantification of 1991; 230(1): 36-41.P E Undill, M J Cree, J A Olson, K C McHardy, P F Sharp, J F Forrester.A comparison of computer based classification methods applied to the detection of microaneurysms in ophtalmic e-ISSN: 2087-278X

Table 3 .
Accuracy results for proposed preprocessing method

Table 1 .
Background and Noise segmentation results

Table 2 .
Coarse and fine segmentation results

Table 3 .
Accuracy results for proposed preprocessing method Tsiaras, K A Soper.Diagnosis of diabetic eye disease.JAMA.1982 Complications of diabetes: Screening for retinopathy and Management of foot Royal Society of Medicine Press.1999; 5(4): ISSN 0965-0288.hdi, B Thomas and R Markham.Automated Identification of Diabetic Retinal Exudates in Digital Colour Images.British Journal of Ophthalmology.2003; 87(10 S C Lee, E T Lee, R M Kingsley, Y Wang, D Russell, R Klein, A Wanr.Comparison of diagnosis of early retinal lesions of diabetic retinopathy between a computer system and human experts ch.Clin.Exp.Ophtalmol.2001; 119(4): 509-515.Sinthanayothin, J F Boyce, H L Cook, T. H.Williamson.Automated localization of the optic disc, fovea and retinal blood vessels from digital color fundus images.British Journal of Grisan, A Ruggeri.Detection of optic disc in retinal images by means of a IEEE transactions on medical imaging.2004; 23(10): 1189 M N Subhasis Chaudhuri, Shankar Chatterjee, Michael goldbaum.Detection of blood vessels in twodimensional matched filters.IEEE transactions on medical imaging T Spencer, R P Phillips, P F Sharp, J V Forrester.Automated detection and quantification of ysms in fluorescein angiograms.Graefes Arch.Clin.Exp.Ophtalmol.1991; P E Undill, M J Cree, J A Olson, K C McHardy, P F Sharp, J F Forrester computer based classification methods applied to the detection of microaneurysms in op Comput.Biol.Med.1998; 28(3): 225-238.
Retinal Image Preprocessing: Background and Noise Segmentation (Ibaa Jamal) 543 characteristic curves for proposed method using diaretDB0 and