May 2015
Volume 56, Issue 5
Free
Eye Movements, Strabismus, Amblyopia and Neuro-ophthalmology  |   May 2015
Automatic Computer-Aided Diagnosis of Retinal Nerve Fiber Layer Defects Using Fundus Photographs in Optic Neuropathy
Author Affiliations & Notes
  • Ji Eun Oh
    Biomedical Engineering Branch Division of Convergence Technology, National Cancer Center, Goyang, Korea
  • Hee Kyung Yang
    Department of Ophthalmology, Seoul National University College of Medicine, Seoul National University Bundang Hospital, Seongnam, Korea
  • Kwang Gi Kim
    Biomedical Engineering Branch Division of Convergence Technology, National Cancer Center, Goyang, Korea
  • Jeong-Min Hwang
    Department of Ophthalmology, Seoul National University College of Medicine, Seoul National University Bundang Hospital, Seongnam, Korea
  • Correspondence: Kwang Gi Kim, Biomedical Engineering Branch, Division of Convergence Technology, National Cancer Center, Goyang, 410-769, Korea; kimkg@ncc.re.kr
  • Jeong-Min Hwang, Department of Ophthalmology, Seoul National University College of Medicine, Seoul National University Bundang Hospital, 166 Gumiro, Bundang-gu, Seongnam, Gyeonggi-do 463-707, Korea; hjm@snu.ac.kr
  • Footnotes
     JEO and HKY are joint first authors.
  • Footnotes
     JEO and HKY contributed equally to the work presented here and should therefore be regarded as equivalent authors.
Investigative Ophthalmology & Visual Science May 2015, Vol.56, 2872-2879. doi:10.1167/iovs.14-15096
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to Subscribers Only
      Sign In or Create an Account ×
    • Get Citation

      Ji Eun Oh, Hee Kyung Yang, Kwang Gi Kim, Jeong-Min Hwang; Automatic Computer-Aided Diagnosis of Retinal Nerve Fiber Layer Defects Using Fundus Photographs in Optic Neuropathy. Invest. Ophthalmol. Vis. Sci. 2015;56(5):2872-2879. doi: 10.1167/iovs.14-15096.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose.: To evaluate the validity of an automatic computer-aided diagnosis (CAD) system for detection of retinal nerve fiber layer (RNFL) defects on fundus photographs of glaucomatous and nonglaucomatous optic neuropathy.

Methods.: We have proposed an automatic detection method for RNFL defects on fundus photographs in various cases of glaucomatous and nonglaucomatous optic neuropathy. In order to detect the vertical dark bands as candidate RNFL defects, the nonuniform illumination of the fundus image was corrected, the blood vessels were removed, and the images were converted to polar coordinates with the center of the optic disc. False positives (FPs) were reduced by using knowledge-based rules. The sensitivity and FP rates for all images were calculated.

Results.: We tested 98 fundus photographs with 140 RNFL defects and 100 fundus photographs of healthy normal subjects. The proposed method achieved a sensitivity of 90% and a 0.67 FP rate per image and worked well with RNFL defects with variable depths and widths, with uniformly high detection rates regardless of the angular widths of the RNFL defects. The average detection accuracy was approximately 0.94. The overall diagnostic accuracy of the proposed algorithm for detecting RNFL defects among 98 patients and 100 healthy individuals was 86% sensitivity and 75% specificity.

Conclusions.: The proposed CAD system successfully detected RNFL defects in optic neuropathies. Thus, the proposed algorithm is useful for the detection of RNFL defects.

Retinal nerve fiber layer (RNFL) defects such as localized thinning or loss of the RNFL, including papillomacular bundle defects,1,2 are a major sign that precedes detectable optic disc changes and visual field loss in the early stages of glaucoma or other nonglaucomatous optic neuropathies.3,4 Glaucoma leads to structural changes of the optic disc and RNFL, and advanced visual field loss. Because such changes are generally irreversible, early diagnosis and treatment are critical.5,6 Papillomacular bundle defect is also a major sign of various hereditary, toxic, and mitochondrial optic neuropathies, and early detection of such changes can facilitate the clinical diagnosis and decision-making.3,4 
Various methods are available for examining RNFL defects. Heidelberg retina topography (HRT), scanning laser polarimetry (SLP), and optical coherence tomography (OCT) can accurately produce quantitative measurements of the optic disc and RNFL.5 However, these technologies are expensive and require a trained technician.7 Fundus photography is the most commonly used and cost-effective imaging tool that can be performed as part of the routine medical check-up; thus, it has an enormous potential for community glaucoma screening in remote or nonhospital environments.8 However, RNFL defects are thin and shallow in the early stages of glaucoma, which makes them difficult to detect. Fundus photographs may have sharper boundaries for RNFL defects showing early defects on fundus photographs, while OCT results may not be definitely decreased in thickness compared with the age-matched normal population, with only a relative thinning compared with other sectors observed. Further, the detection rates are highly dependent on the experience of the examiner. Therefore, the development of a method that is more sensitive and reliable for detecting RNFL defects on fundus photographs for use as a diagnostic tool for glaucoma and various optic neuropathies is important. 
In this work, we have proposed a simple and efficient algorithm for the automatic detection of RNFL defects that can alert ophthalmologists to the location of possible RNFL defects. We tested its validity in patients with localized RNFL defects with glaucoma and other hereditary and toxic optic neuropathies. 
Methods
Subjects and Fundus Photographs
This study included the fundus photographs of 98 subjects with 140 localized RNFL defects and 100 fundus photographs of healthy normal subjects who did not have any history of ocular diseases and had normal RNFL and optic disc appearance on fundus photographs. The subjects consisted of 89 patients with early to moderate glaucoma, and their mean deviation was −4.06 ± 3.44 dB (range, −12.87 to approximately −0.03) by the Humphrey Field Analyzer, and nine patients with nonglaucomatous optic neuropathy with papillomacular bundle defects: four patients with Leber's hereditary optic neuropathy, three patients with autosomal dominant optic atrophy, and two patients with toxic optic neuropathy. 
The fundus images were obtained from Seoul National University Bundang Hospital. All images were photographed by using a fundus camera (KOWA VX-10; Kowa Company Ltd., Tokyo, Japan), and were taken as 24-bit color images. Images of various sizes were resized to a resolution of 1278 × 848 pixels in order to reduce the computational time and to minimize the effects of small pathological regions. Two ophthalmologists independently marked the region of the RNFL defects and when both reviewers determined the presence of RNFL defect, the case was considered to have an RNFL defect. The results of manual detection were confirmed by Stratus OCT (Carl Zeiss Meditec, Germany) or Spectralis OCT (Heidelberg Engineering, Germany) results showing a definite or relative localized thinning of RNFL analysis compared to other sectors. The research adhered to the tenets of the Declaration of Helsinki. 
Flowchart of Proposed Computer-Aided Diagnosis (CAD)
The proposed CAD method was performed in three steps. First, the algorithm automatically detected the location of the main features in the fundus image, such as the optic disc, macula, and blood vessels. Then, it corrected the nonuniform illumination of the green channel image by using the bias image. Second, blood vessels were removed. Blood vessel segmentation is essential in order to achieve RNFL defect detection. Then, the image was converted to polar coordinates according to the center of the optic disc. Finally, the candidate RNFL defects were observed as vertical dark bands, and the false positives were subsequently reduced by using knowledge-based rules of the average pixel value in the candidate region, the vertical length, and the angular location. To reduce perceptual errors, two ophthalmologists performed independent double reading of the fundus photographs. A flowchart showing the proposed CAD system for RNFL defects and the resulting images at each step is provided in Figure 1
Figure 1
 
Flowchart of the proposed CAD system for automatic RNFL defect detection and result images at each step.
Figure 1
 
Flowchart of the proposed CAD system for automatic RNFL defect detection and result images at each step.
On fundus images, the RNFL defect locations were determined as clock hours, and we measured the clock-hour location of the RNFL defects in all of the cases. The clock-hour location was defined as the location from a reference line in a clockwise direction for the right eyes and in a counterclockwise direction for the left eyes (Fig. 2).9 The reference line was a vertical line through the center of the circle. 
Figure 2
 
Clock-hour locations of RNFL defects on fundus photographs of the (a) right eye and (b) left eye.
Figure 2
 
Clock-hour locations of RNFL defects on fundus photographs of the (a) right eye and (b) left eye.
Preprocessing
Preprocessing of the proposed algorithm can be divided into two major steps: noise reduction and illumination correction. We used the green channel of the color fundus image because this maximizes the contrast between the RNFL defects and blood vessels.10 
Noise Reduction.
To remove unwanted text images such as patient information, we applied a binary mask by using a morphological open operation. The binary mask extracted the region of interest (ROI) in the entire image. Then, a median filter of size 3 × 3 was applied to the masked image to suppress noise and to maintain an edge of the image. 
Illumination Correction.
Nonuniform illumination by nonideal acquisition condition and the spherical geometry of the eye can cause severe distortions in the color fundus image. This nonuniform illumination artifact is often described as “shading” or “bias.” Correcting the nonuniform illumination can contribute to accurate blood vessel segmentation by enhancing the contrast of the blood vessel at the periphery of the fundus image.1113 Zheng et al.14 proposed a retrospective illumination-correction method based on the sparsity of the image gradient distribution. This method can automatically correct the illumination of an arbitrary fundus image by using the bias. The bias of the image denotes the spatial variations of intensity caused by illumination changes for images taken by a digital camera. The bias is a smooth field in any format, which can be represented by a bivariate polynomial, B-spline, etc. In this study, we used the method described by Zheng et al.14 A given green channel image (G) assumes the product of the uniformly illuminated fundus image (I) and the bias field (B), as follows:  where (i, j) is the pixel position in the image.  
An example of nonuniform illumination correction is shown in Figure 3. The given green channel image is shown in Figure 3a and the bias image as the shading artifact is shown in Figure 3b. The corrected image was generated by the given green channel image divided by the bias image, as shown in Figure 3c. The intensity profiles of one row in the given green channel image and in the corrected image is shown in Figure 3d. This shows that the background intensity in the corrected image had greater uniformity compared to the given green image. 
Figure 3
 
An example of nonuniform illumination correction: (a) green channel image, (b) estimated bias field for correction, (c) corrected green image, and (d) intensity profiles of one row in the given green channel image and in the corrected green image.
Figure 3
 
An example of nonuniform illumination correction: (a) green channel image, (b) estimated bias field for correction, (c) corrected green image, and (d) intensity profiles of one row in the given green channel image and in the corrected green image.
Blood Vessel Removal
Blood vessel removal (BVR) is an essential step to detect RNFL defects more accurately. Many methods have been reported for blood vessel extraction from fundus images.9,1519 In this study, the blood vessels were extracted by using a morphological bottom-hat transform in order to detect the blood vessels as dark regions. The bottom-hat transform is defined as the difference between the closing by a disk-shaped structural element (SE) and the input image. Since the maximum blood vessel width is approximately 10 pixels, the SE diameter was set to 10 pixels. Although the bottom-hat method can be used simply to detect the blood vessels, this method is not sufficient for exact extraction of the entire blood vessel structure. Therefore, we also used the Kirsch method, which is an edge detector that determines the maximum edge strength with an 8-directional filter.15 The Kirsch method is useful for detecting the edge of large blood vessel structure, but it is not suitable in detecting small vessels and pathological regions. 
The results of blood vessel extraction after applying the bottom-hat transform, Kirsch method, and combined method are shown in Figure 4. To compensate for the disadvantages of the two methods, we combined the two methods for iterative use. 
Figure 4
 
Results of blood vessel extraction from (a) bottom-hat transformation, (b) Kirsch method, and the (c) combined method.
Figure 4
 
Results of blood vessel extraction from (a) bottom-hat transformation, (b) Kirsch method, and the (c) combined method.
To reduce the false-positive (FP) rates, the pathological regions in the fundus images were detected with blood vessels, as shown in Figure 4. For each pixel of the segmented vessel region determined by using the combined method, the intensity value was replaced by a mean intensity that was computed over all pixels in a 61 × 61 neighborhood of the pixel location. The size of the neighborhood region was chosen to be a sufficiently large area to minimize the influence caused by other pathological lesions and blood vessel regions. 
Polar Transformation
Defects of the RNFL appear radially from the optic disc in various forms, such as fan-shaped, wedge-shaped, slit-like shape, and spindle-like shape. In this study, to detect the RNFL defects effectively, the BVR image was transformed to the polar coordinates with a reference point as the location of maximum cup depth in the optic disc. Because the optic disc has high intensity in the green image, the reference point in the optic disc was detected by using a local maximum method. Further, we detected the macular center by using a local minimum method to reduce the FP rate. In the polar coordinates, a retinal location (pixel) is represented by a radial distance and angle from the reference point. Because the RNFL defects do not spread out in the radial direction completely around the optic disc, the RNFL defect lines in the polar coordinates are slightly tilted toward the macular position. The difference of the radial widths of the defect in the polar coordinate was also negligible. Thus, we assumed that the RNFL defects in the polar coordinates were a relatively straight line in the vertical direction. The detection results of the reference point in the optic disc and the macular center are shown in Figure 5a. The polar coordinate transformation of the BVR image results is shown in Figure 5b, where the RNFL defect is marked by black arrows. 
Figure 5
 
(a) Detection results for the optic disc and macular centers and (b) polar coordinate transformation of blood vessel removal image and RNFL defects (black arrow). The circle region (yellow dotted line) indicates the region for polar transformation and is transformed in the clockwise direction (white arrow) for the right eye from the start position.
Figure 5
 
(a) Detection results for the optic disc and macular centers and (b) polar coordinate transformation of blood vessel removal image and RNFL defects (black arrow). The circle region (yellow dotted line) indicates the region for polar transformation and is transformed in the clockwise direction (white arrow) for the right eye from the start position.
Detection of RNFL Defects
Hough transformation was applied to detect the candidates with straight lines. The procedure for detecting RNFL defects consisted of three steps. First, the polar transformed image was filtered with a 2D-Gaussian filter to smooth the boundary of the blood vessel region. Second, we determined the edge of the smoothed image by using the Canny edge detection algorithm. Third, the Hough transform for line detection was performed on the edge image. Finally, the candidate RNFL defects were detected, including the misdetected candidates. If the detected region of the arbitrary candidate overlapped with the region of the gold standard, it was considered a true positive (TP) detection. 
FP Reduction
In this study, an FP represented a misdetected RNFL defect candidate. The misdetected candidates of the RNFL defects were reduced by using knowledge-based rules. For each candidate, we classified the average pixel values in the candidate region, the vertical length, and the angular location. With these three features, the true RNFL defects were determined. First, we extracted the average pixel values in the candidate region that were smaller compared with the average value in the surrounding background region, except for the blood vessel region. Then, the candidates were classified by the vertical length to suppress noise. We selected the vertical lengths which were more than two times of the maximum blood vessel width in order to reduce the false detection of the blood vessel and other noise. We selected the angular location within a main vascular region (±79° from a reference line) corresponding to the temporal sector, as this is where clinically significant loss of glaucomatous and nonglaucomatous RNFL thinning takes place. The reference line was drawn from the center of the optic disc to the macular center on the fundus image. We also excluded the macula region and the blood vessel region from the inner region of the main vascular region. The false positives were removed after FP reduction, leaving the true detected candidates. 
Main Outcome Measures
The intensities and widths of the RNFL defects were measured and the detection rates according to each feature were determined for glaucoma and nonglaucomatous optic neuropathy with papillomacular bundle defects. The comparison was based on agreement in the position of the RNFL defect band and not the width. 
To evaluate the performance of the proposed algorithm quantitatively, we used free-response receiver operating characteristics (FROC) analysis.20,21 The curve of FROC is a tool for characterizing the performance of the proposed algorithm at all decision thresholds simultaneously.22 The thresholds were decided by varying the number of the initial candidate in the detection step. We obtained the FROC curve by plotting the sensitivity as a function of the number of false positives per image (FPs/image). The sensitivity, which was defined as the number of TPs divided by the sum of TPs and false negatives (FN), indicated the ability of the algorithm to detect RNFL defects correctly. We defined the sensitivity and the number of FP per image in this evaluation as follows:   where RFP is the number of remaining false positives after application of the FP reduction method and Nimg is the total number of images.  
Results
Validity of the Proposed CAD in Automated RNFL Detection
We tested 98 patients with 140 RNFL defects, including nine patients with nine images with papillomacular bundle defects. The distribution of the number of RNFL defects per case was as follows: 60 patients, single RNFL defect; 34 patients, two RNFL defects; and four patients, three RNFL defects. Among the 131 defects of 89 patients with glaucoma, the algorithm identified 117 defects and failed to identify 14. Among the nine defects of nine patients with papillomacular bundle defects, the algorithm identified all nine defects. The results obtained by using the proposed algorithm are shown in Figure 6. The white arrows indicate the region of the RNFL defect marked by the two ophthalmologists and the yellow lines indicate the TPs. The first column shows the original images, the second column shows the correct detection results, and the third column shows the red-free fundus photographs of the original images. 
Figure 6
 
Representative detection results using the proposed algorithm on three patients consisting of a typical superior temporal RNFL defect in glaucoma (top row), slit-like RNFL defect in early stage preperimetric glaucoma (middle row) and papillomacular bundle RNFL defect in a patient with Leber's hereditary optic neuropathy (bottom row). (a) First column: Original image. (b) Second column: Detection results. (c) Third column: Red-free fundus image of original image. White arrows indicate the region of RNFL defect marked by two ophthalmologists.
Figure 6
 
Representative detection results using the proposed algorithm on three patients consisting of a typical superior temporal RNFL defect in glaucoma (top row), slit-like RNFL defect in early stage preperimetric glaucoma (middle row) and papillomacular bundle RNFL defect in a patient with Leber's hereditary optic neuropathy (bottom row). (a) First column: Original image. (b) Second column: Detection results. (c) Third column: Red-free fundus image of original image. White arrows indicate the region of RNFL defect marked by two ophthalmologists.
We derived the FROC curve of the proposed algorithm for the 98 patients by varying the number of the initial candidates in the detection step (Fig. 7). The sensitivity by using nonuniform illumination was 74% and the sensitivity after illumination correction was increased 1.2-fold from 74% to 90%. Finally, a sensitivity of 90% was obtained at 0.67 FPs per image. The 100 fundus photographs of healthy normal subjects showed a false positive (FP) rate of 0.25 per image. Consequently, the overall diagnostic accuracy of the proposed algorithm for detecting RNFL defects among 98 patients and 100 healthy individuals was 86% sensitivity and 75% specificity. 
Figure 7
 
The curve of FROC indicating the performance of the proposed algorithm for 98 cases containing 140 RNFL defects. The proposed algorithm achieved 90% sensitivity at 0.67 false positives per image.
Figure 7
 
The curve of FROC indicating the performance of the proposed algorithm for 98 cases containing 140 RNFL defects. The proposed algorithm achieved 90% sensitivity at 0.67 false positives per image.
Intensity of RNFL Defects
In fundus photography, the intensity of the RNFL represents the thickness of the RNFL, and the thinning or loss of the RNFL appears as low intensity. Therefore, we compared the mean intensity of the blood vessel, RNFL defects, and normal RNFL region (Fig. 8). In order to maintain the same grayscale level for each image, the grayscales of the images were normalized by using an arbitrary image. We automatically selected a region of interest (ROI) with a 50 × 50 square region in the 8 o'clock direction according to the clock-hour locations on Figure 2, which was independent of eye orientation, at a distance of 100 pixels from the optic disc in the illumination corrected image. We set the ROI to minimize the effect of the darkened macula and RNFL defect region. The mean intensity of the normal RNFL region was calculated from the ROI excluding blood vessel region in the fundus image. The blood vessel region was selected in the same ROI and the mean intensity of blood vessel was calculated. The defect region of the RNFL was selected as an arbitrary region with the same size. The regions of the RNFL defects were marked by two ophthalmologists. The box plot shows the differences in the mean intensity. The mean intensity (±SD) of the RNFL defects (38.20 ± 7.60) was significantly higher compared to the blood vessel region (36.06 ± 7.19), but was lower compared with the normal RNFL region (40.20 ± 8.29; F value = 22.64, P < 0.005, by ANOVA).23 Therefore, if the difference between the mean intensity of each candidate defect region and its surrounding background region is greater than the half value of the difference between the mean intensity of RNFL defects and normal RNFL, the candidate defect region was selected as an RNFL defect. 
Figure 8
 
Mean intensity of the blood vessels, RNFL defects, and normal RNFL regions. A central line of the box plot indicates the median value of the data. Lower and upper boundary lines of the central box are at the 25% and 75% quartile of the data. Box represents 95% confidence interval.
Figure 8
 
Mean intensity of the blood vessels, RNFL defects, and normal RNFL regions. A central line of the box plot indicates the median value of the data. Lower and upper boundary lines of the central box are at the 25% and 75% quartile of the data. Box represents 95% confidence interval.
Location of RNFL Defects
The distribution of the clock-hour locations for the RNFL defects is shown in Figure 9. The RNFL defects were most frequently located at 7 o'clock, followed by 11 o'clock and 10 o'clock. 
Figure 9
 
Distribution of clock-hour locations for the RNFL defects.
Figure 9
 
Distribution of clock-hour locations for the RNFL defects.
Detection Rate According to the Angular Widths of the RNFL Defects
The detection accuracy as the angular widths of the RNFL defects increased is shown in the Table. The angular width of the RNFL defects was defined as the angle between the proximal and distal border lines, with the height of each bar representing the true detection rate of the RNFL defects. Although the numbers of RNFL defects varied, the detection rates of the proposed algorithm were almost uniformly high, regardless of the angular widths of the RNFL defects. The average detection accuracy was approximately 0.94. 
Table
 
Performance of the Proposed Algorithm as a Function of the Angular Widths of the RNFL Defects
Table
 
Performance of the Proposed Algorithm as a Function of the Angular Widths of the RNFL Defects
Discussion
In this study, we have proposed a fully automatic method for detecting various forms and widths of RNFL defects in color fundus images. Fundus photography is the most common screening tool to detect RNFL defects in various optic neuropathies. However, the detection of RNFL defects by using fundus photographs depends on the experience of the examiner, and early defects may be missed because of the low contrast of the RNFL. Therefore, we developed a simple and efficient algorithm to assist the ophthalmologist for the detection of RNFL defects. 
The strength of the proposed algorithm is that it can detect very narrow defects in early stage glaucoma to nonglaucomatous optic neuropathy involving the papillomacular bundle accurately, as shown in the representative cases in Figure 6. To our knowledge, no previous studies described specific methods for detecting RNFL defects with various forms and widths in fundus images. Our results showed that the proposed algorithm was successful, with a sensitivity of 90% for glaucoma and 100% for papillomacular bundle defects in non-glaucomatous optic neuropathies. 
Many algorithms have been proposed for detecting RNFL defects. Prageeth et al.24 used texture analysis by utilizing only intensity information about the RNFL around the optic disc in the red-free fundus image. Odstrčilík et al.25 proposed the use of texture analysis by utilizing Gaussian Markov random fields (GMRF) for classification of healthy and glaucomatous RNFL tissue in fundus images.25 These results were compared with the OCT images as a gold standard.2426 Muramatsu et al.18 applied three sizes of Gabor filters to detect RNFL defects in the fundus image. The detection rates were 89% ∼ 91% at 1.0 FPs per image in these studies.18,25 However, because determining the filter width for detecting RNFL defects with various forms and widths is difficult with these methods, we applied the Hough transformation to detect RNFL defect candidates with straight continuous lines. In addition, these previous studies confined their study subjects to those with glaucoma with localized RNFL defects, which would be apparently visible and found on OCT; however, these studies did not include early stage preperimetric glaucoma, other nonglaucomatous optic neuropathies, and papillomacular bundle defects. 
The defects of the RNFL were most commonly located in the inferior temporal and superior temporal regions. These locations are the most frequently affected in the early stage of glaucoma.27 Additionally, the detection rate of the proposed algorithm was almost consistent, regardless of the angular widths of RNFL defects. However, among the total defects, the proposed algorithm performed worse in cases with shallow defects (i.e., in early-stage glaucoma or in images with poor resolution). 
This study had several limitations. First, the number of images used in this study was not large; thus, a larger database should be used in the future. Second, we did not consider structural changes of the optic disc, such as cupping, notching, or pallor of the rim. The change in the optic disc is an important indicator of the severity of glaucoma, and detection of optic disc parameters can provide a significant differential clue regarding glaucomatous and nonglaucomatous optic neuropathies.27,28 Therefore, future work should be focused on the detection of optic disc changes, and combining these findings with RNFL defects. Further, regarding the high rate of false positives per image, modification of the FP reduction method may improve the reliability of the current algorithm for early detection of various optic neuropathies. Finally, diffuse RNFL defects in advanced glaucoma or optic atrophy cannot be detected with our program, since the mean intensity of the RNFL is low in all clock hours and a localized lesion is not distinguishable. However, in these cases, the pathologic features of disc cupping or atrophy are more prominent than RNFL thinning, and can easily be detected. 
In conclusion, the proposed algorithm showed a reliable diagnostic accuracy for automatically detecting RNFL defects in fundus photographs of various optic neuropathies. This method has the potentional to assist the ophthalmologist in reading fundus photographs and replace double reading in the office, and can also be applied in mass screening using fundus photographs. 
Acknowledgments
Supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF); the Ministry of Science, ICT, and Future Planning (2013R1A1A2010606); and Grant no. 14-2014-041 from the SNUBH Research Fund. 
Disclosure: J.E. Oh, None; H.K. Yang, None; K.G. Kim, None; J.-M. Hwang, None 
References
Barboni P, Carbonelli M Savini G, et al. Natural history of Leber's hereditary optic neuropathy: longitudinal analysis of the retinal nerve fiber layer by optical coherence tomography. Ophthalmology. 2010; 117: 623–627.
Wang MY Sadun AA. Drug-related mitochondrial optic neuropathies. J Neuroophthalmol. 2013; 33: 172–178.
Haleem MS Han L, van Hemert J Li B. Automatic extraction of retinal features from colour retinal images for glaucoma diagnosis: a review. Comput Med Imaging Graph. 2013; 37: 581–596.
Hwang J-M Kim T-W, Park KH Kim DM, Kim H. Correlation between topographic profiles of localized retinal nerve fiber layer defects as determined by optical coherence tomography and red-free fundus photography. J Glaucoma. 2006; 15: 223–228.
Hoh ST Greenfield DS, Mistlberger A Liebmann JM, Ishikawa H Ritch R. Optical coherence tomography and scanning laser polarimetry in normal, ocular hypertensive, and glaucomatous eyes. Am J Ophthalmol. 2000; 129: 129–135.
Cheng J Liu J, Xu Y et al. Superpixel classification based optic disc and optic cup segmentation for glaucoma screening. IEEE T Med Imaging. 2013; 32: 1019–1032.
Lu C-K, Tang TB Laude A, Deary IJ Dhillon B, Murray AF. Quantification of parapapillary atrophy and optic disc. Invest Ophthalmol Vis Sci. 2011; 52: 4671–4677.
Nduaguba C Lee RK. Glaucoma screening: current trends, economic issues, technology, and challenges. Curr Opin Ophthalmol. 2006; 17: 142–152.
Kim T-W Park U-C, Park KH Kim DM. Ability of Stratus OCT to identify localized retinal nerve fiber layer defects in patients with normal standard automated perimetry results. Invest Ophthalmol Vis Sci. 2007; 48: 1635–1641.
Khalid NEA Noor NM, Mahmud Z Yahya S, Ariff NM. Bridging Quantitative and Qualitative of glaucoma detection. World Acad Sci Eng Technol. 2012; 6: 584–588.
Youssif AA Ghalwash AZ, Ghoneim AS. Comparative study of contrast enhancement and illumination equalization methods for retinal vasculature segmentation. Paper presented at: Cairo International Biomedical Engineering Conference; December 21–24 2006 ; Cairo, Egypt. Available at: https://scholar.google.co.kr/scholar?hl=ko&q=Comparative+study+of+contrast+enhancement+and+illumination+equalization+methods+for+retinal+vasculature+segmentation.&btnG=&lr=. Accessed April 20, 2015.
Kubecka L, Jan J Kolar R. Retrospective illumination correction of retinal images. Int J Biomed Imaging. 2010; 2010: 780262.
Kolar R Odstrčilík J, Jan J Harabis V. Illumination correction and contrast equalization in colour fundus images. Paper presented at: 19th European Signal Processing Conference (EUSIPCO 2011); August 29–September 2, 2011; Barcelona, Spain: 298–302. Available at: https://scholar.google.co.kr/scholar?q=Illumination+correction+and+contrast+equalization+in+colour+fundus+images&btnG=&hl=ko&as_sdt=0%2C5. Accessed April 20, 2015.
Zheng Y Vanderbeek B, Xiao R et al. Retrospective illumination correction of retinal fundus images from gradient distribution sparsity. In: 9th IEEE International Symposium on Biomedical Imaging (ISBI). Barcelona, Spain: IEEE; 2012; 972–975.
Karasulu B. Automatic extraction of retinal blood vessels: a software implementation. Eur Sci J. 2012; 8: 47–57.
Fraz MM Basit A, Barman S. Application of morphological bit planes in retinal blood vessel extraction. J Digit Imaging. 2013;26: 274–286.
Fraz MM Remagnino P, Hoppe A et al. Blood vessel segmentation methodologies in retinal images—a survey. Comput Methods Programs Biomed. 2012; 108: 407–433.
Muramatsu C, Hayashi Y Sawada A, et al. Detection of retinal nerve fiber layer defects on retinal fundus images for early diagnosis of glaucoma. J Biomed Opt. 2010; 15: 016021.
Zheng J Lu P-R, Xiang D et al. Retinal image graph-cut segmentation algorithm using multiscale hessian-enhancement-based nonlocal mean filter. Comput Math Methods Med. 2013; 2013: 927285.
Metz CE. Some practical issues of experimental design and data analysis in radiological ROC studies. Invest Radiol. 1989; 24: 234–245.
Chakraborty DP Winter L. Free-response methodology: alternate analysis and a new observer-performance experiment. Radiology. 1990; 174: 873–881.
Bandos AI Rockette HE, Song T Gur D. Area under the free-response ROC curve (FROC) and a related summary index. Biometrics. 2009; 65: 247–256.
Mertler CA Vannatta RA. Advanced and Multivariate Statistical Methods. Los Angeles: Pyrczak; 2002.
Prageeth P David J, Sukesh Kumar A. Early detection of retinal nerve fiber layer defects using fundus image processing. In: Recent Advances in Intelligent Computational Systems (RAICS). Trivandrum, India: IEEE; 2011: 930–936.
Odstrčilík J Kolar R, Harabis V Gazarek J, Jan J. Retinal nerve fiber layer analysis via Markov random fields texture modelling. Paper presented at: 18th European Signal Processing Conference (EUSPICO-2010); August 23–27 2010 ; Aalborg, Denmark: 1650–1654. Available at https://scholar.google.co.kr/scholar?q=Retinal+nerve+fiber+layer+analysis+via+Markov+random+fields+texture+modeling&btnG=&hl=ko&as_sdt=0%2C5. Accessed April 20, 2015.
Jan J, Odstrčilík J Gazarek J, Kolar R. Retinal image analysis aimed at blood vessel tree segmentation and early detection of neural-layer deterioration. Comput Med Imaging Graph. 2012; 36: 431–441.
Medeiros FA Zangwill LM, Bowd C Vessani RM, Susanna RJr Weinreb RN. Evaluation of retinal nerve fiber layer, optic nerve head, and macular thickness measurements for glaucoma detection using optical coherence tomography. Am J Ophthalmol. 2005; 139: 44–55.
Wong D Liu J, Lim J Li H, Wong T. Automated detection of kinks from blood vessels for optic cup segmentation in retinal images. Proc SPIE. 2009; 72601J:.
Figure 1
 
Flowchart of the proposed CAD system for automatic RNFL defect detection and result images at each step.
Figure 1
 
Flowchart of the proposed CAD system for automatic RNFL defect detection and result images at each step.
Figure 2
 
Clock-hour locations of RNFL defects on fundus photographs of the (a) right eye and (b) left eye.
Figure 2
 
Clock-hour locations of RNFL defects on fundus photographs of the (a) right eye and (b) left eye.
Figure 3
 
An example of nonuniform illumination correction: (a) green channel image, (b) estimated bias field for correction, (c) corrected green image, and (d) intensity profiles of one row in the given green channel image and in the corrected green image.
Figure 3
 
An example of nonuniform illumination correction: (a) green channel image, (b) estimated bias field for correction, (c) corrected green image, and (d) intensity profiles of one row in the given green channel image and in the corrected green image.
Figure 4
 
Results of blood vessel extraction from (a) bottom-hat transformation, (b) Kirsch method, and the (c) combined method.
Figure 4
 
Results of blood vessel extraction from (a) bottom-hat transformation, (b) Kirsch method, and the (c) combined method.
Figure 5
 
(a) Detection results for the optic disc and macular centers and (b) polar coordinate transformation of blood vessel removal image and RNFL defects (black arrow). The circle region (yellow dotted line) indicates the region for polar transformation and is transformed in the clockwise direction (white arrow) for the right eye from the start position.
Figure 5
 
(a) Detection results for the optic disc and macular centers and (b) polar coordinate transformation of blood vessel removal image and RNFL defects (black arrow). The circle region (yellow dotted line) indicates the region for polar transformation and is transformed in the clockwise direction (white arrow) for the right eye from the start position.
Figure 6
 
Representative detection results using the proposed algorithm on three patients consisting of a typical superior temporal RNFL defect in glaucoma (top row), slit-like RNFL defect in early stage preperimetric glaucoma (middle row) and papillomacular bundle RNFL defect in a patient with Leber's hereditary optic neuropathy (bottom row). (a) First column: Original image. (b) Second column: Detection results. (c) Third column: Red-free fundus image of original image. White arrows indicate the region of RNFL defect marked by two ophthalmologists.
Figure 6
 
Representative detection results using the proposed algorithm on three patients consisting of a typical superior temporal RNFL defect in glaucoma (top row), slit-like RNFL defect in early stage preperimetric glaucoma (middle row) and papillomacular bundle RNFL defect in a patient with Leber's hereditary optic neuropathy (bottom row). (a) First column: Original image. (b) Second column: Detection results. (c) Third column: Red-free fundus image of original image. White arrows indicate the region of RNFL defect marked by two ophthalmologists.
Figure 7
 
The curve of FROC indicating the performance of the proposed algorithm for 98 cases containing 140 RNFL defects. The proposed algorithm achieved 90% sensitivity at 0.67 false positives per image.
Figure 7
 
The curve of FROC indicating the performance of the proposed algorithm for 98 cases containing 140 RNFL defects. The proposed algorithm achieved 90% sensitivity at 0.67 false positives per image.
Figure 8
 
Mean intensity of the blood vessels, RNFL defects, and normal RNFL regions. A central line of the box plot indicates the median value of the data. Lower and upper boundary lines of the central box are at the 25% and 75% quartile of the data. Box represents 95% confidence interval.
Figure 8
 
Mean intensity of the blood vessels, RNFL defects, and normal RNFL regions. A central line of the box plot indicates the median value of the data. Lower and upper boundary lines of the central box are at the 25% and 75% quartile of the data. Box represents 95% confidence interval.
Figure 9
 
Distribution of clock-hour locations for the RNFL defects.
Figure 9
 
Distribution of clock-hour locations for the RNFL defects.
Table
 
Performance of the Proposed Algorithm as a Function of the Angular Widths of the RNFL Defects
Table
 
Performance of the Proposed Algorithm as a Function of the Angular Widths of the RNFL Defects
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×