January 2007
Volume 48, Issue 1
Free
Glaucoma  |   January 2007
Rule Extraction for Glaucoma Detection with Summary Data from StratusOCT
Author Affiliations
  • Mei-Ling Huang
    From the Department of Industrial Engineering and Management, National Chin-Yi Institute of Technology, Taipei, Taichung, Taiwan; and the
  • Hsin-Yi Chen
    Glaucoma Service, Department of Ophthalmology, China Medical University Hospital Taichung City, Taiwan.
  • Jian-Cheng Lin
    From the Department of Industrial Engineering and Management, National Chin-Yi Institute of Technology, Taipei, Taichung, Taiwan; and the
Investigative Ophthalmology & Visual Science January 2007, Vol.48, 244-250. doi:10.1167/iovs.06-0320
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Mei-Ling Huang, Hsin-Yi Chen, Jian-Cheng Lin; Rule Extraction for Glaucoma Detection with Summary Data from StratusOCT. Invest. Ophthalmol. Vis. Sci. 2007;48(1):244-250. doi: 10.1167/iovs.06-0320.

      Download citation file:


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

      ×
  • Supplements
Abstract

purpose. To extract and induce rules of association for differentiating between normal and glaucomatous eyes based on the quantitative assessment of summary data reports from the StratusOCT (optical coherence tomography; Carl Zeiss Meditec, Inc., Dublin, CA) in a Taiwan Chinese population.

methods. One randomly selected eye of each of the 64 patients with glaucoma and each of the 71 normal subjects was included in the study. Measurements of glaucoma variables (retinal nerve fiber layer thickness and optic nerve head analysis results) were obtained with the StratusOCT. A self-organizing map and decision tree were applied to extract features and determine rules of association for glaucoma detection.

results. The average visual field mean deviation was −0.55 ± 0.57 dB in the normal group and −4.30 ± 3.32 dB in the glaucoma group. Vertical cup-to-disc (C/D) ratio and inferior quadrant thickness were extracted from the decision tree, and three association rules were determined for glaucoma detection.

conclusions. The precise rules of association induced by a novel application of the decision tree may enhance glaucoma detection.

Glaucoma is an ocular disease that causes progressive damage in the optic nerve fibers and leads to visual field loss. The detection of early glaucomatous damage is very important for early treatment. In addition, the development of screening programs for glaucoma relies on methods for the detection of glaucoma at an early stage. It has been demonstrated that structural damage to the optic nerve head (ONH) and peripapillary retinal nerve fiber layer (RNFL) can occur well before any detectable functional visual loss. 1 2 3 Several imaging instruments have been designed to measure peripapillary RNFL thickness and optic disc topography objectively and quantitatively. 4 5 6 7 8 9 The ability to detect glaucoma using these instruments has been widely described and discussed. 4 5 6 7 8 9 10 11 12 13 14 15 Optical coherence tomography (OCT) is the optical equivalent of ultrasonography with high in vivo resolution. 16 17 18 The latest generation of StratusOCT (Carl Zeiss Meditec, Inc., Dublin, CA) even provides objective, quantitative, and reproducible measurements of the retina, RNFL thickness, and ONH. 19 20  
The quest for an efficient and robust classifier is an important issue in medical decision-making. An overview of recent developments in machine learning for medical decisions and knowledge-based decision support systems is given in Kononenko 21 and Wetter. 22 Recently, machine learning classifiers of OCT measurements have provided a simple and accurate index for diagnosing the presence or absence of glaucoma, as well as its severity. 23 In our previous report, we developed automated classifiers to improve the discriminating power between glaucomatous and normal eyes with input parameters from StratusOCT. 24 In this study, data-mining methods were applied to detect relationships between different attributes in large data sets. Machine learning is a classification method and a part of the field of data mining. Decision-tree building is one of the machine learning methods and has been studied extensively as a solution for classification problems. However, in the analysis of glaucoma datasets, this technique has been used only in a limited investigation. 25 Our study with the combination of automatic labeling with a self-organizing map (LabelSOM) and decision-tree methods was designed to determine rules of association for detection of glaucoma. The study was split into two stages: (1) cluster analysis and feature selection through LabelSOM; (2) determination of rules of association based on the application of decision-tree methods. 
Methods
Subjects
One eye from each of 64 patients with glaucoma and 71 normal subjects was included in this observational cross-sectional study. This research adhered to the tenets of the Declaration of Helsinki. Informed consent was obtained from all participants, and the study was approved by the Institutional Review Board of the China Medical University Hospital, which is the major medical center in mid Taiwan. Subjects with a best corrected visual acuity of less than 20/40, a spherical equivalent outside ± 5.0 D, and a cylinder correction of more than 3.0 D were excluded. To increase imaging quality and accuracy, patients with marked peripapillary atrophy were also excluded, to avoid instrumentation problems in the algorithms used to find the layers. All subjects underwent a complete ophthalmic examination, including slit lamp biomicroscopy, measurement of intraocular pressure (IOP), stereoscopic fundus examination, and standard full-threshold automated perimetry (30-2 mode, Humphrey Field Analyzer, model 750, HFA; Carl Zeiss Meditec, Inc.). 
Inclusion criteria for the patients with glaucoma included an initial untreated IOP higher than 22 mm Hg, an open angle, and a reproducible glaucomatous visual field defect in the absence of any other abnormalities to explain the defect. IOP was measured three times, and the average value was obtained. The patients with glaucoma were recruited from a group of patients with high-tension type open-angle glaucoma who had received at least 6 months of regular follow-up at the glaucoma service at the China Medical University Hospital between March 2005 and November 2005. 
Inclusion criteria for normal subjects included no history of eye disease, no family history of glaucoma, IOP lower than 21 mm Hg when measured by Goldmann applanation tonometry, and normal optic disc appearance based on clinical stereoscopic examination (no diffuse or focal rim thinning, optic disc hemorrhage, or RNFL defects) by the same experienced doctor (H-YC, glaucoma specialist). A normal result on the Glaucoma Hemifield Test and corrected pattern SD (HFA, program 30-2) within normal limits were required. Subjects with normal eyes were volunteers from the staff at the China Medical University Hospital. 
Visual Field Testing
Achromatic automated perimetry was performed with an HFA, with the central full-threshold visual field testing program 30-2. Visual field reliability criteria included fixation losses and false-positive and -negative rates of less than 20%. The evaluation of glaucomatous visual field defects was made based on the following liberal criteria: two or more contiguous points with a pattern deviation sensitivity loss of P < 0.01, or three or more contiguous points with sensitivity loss of P < 0.05 in the superior or inferior arcuate areas, or a 10-dB difference across the nasal horizontal midline at two or more adjacent locations and an abnormal result on the glaucoma hemifield test. 26  
StratusOCT Imaging
The StratusOCT (ver. A 4.0.1; Carl Zeiss Meditec Inc.) consists of an infrared-sensitive video camera to provide a view of the scanning probe beam on the fundus, a low-coherence interferometer as a light source and detection unit, a video monitor, a computer, and an image-analysis system. The StratusOCT is calibrated to an axial resolution of ≤10 μm and a transverse resolution of 20 μm. Our quantitative OCT protocol including computing the mean of three regular 3.4-mm circular scans if 512 A-scans, centered on the optic disc to determine RNFL thickness. All scans were completed in a single session by a trained operator after pupil dilation with tropicamide 1%, to achieve a minimum pupillary diameter of 6 mm. The fast ONH radial scan protocol consisted of six linear scans crossing the optic scan. This protocol acquires six 4-mm radial scans in 1.92 seconds. The machine automatically determined the edge of the ONH as the end of the retinal pigment epithelium–choriocapillaris layer. This determination could be manually corrected in cases in which the machine did not identify the edge correctly. A straight line connected the edges of the retinal pigment epithelium-choriocapillaris, and a parallel line was constructed 150 μm anteriorly. The structure below this line was defined as the disc cup, and the structure above the line was defined as the neuroretinal rim. 
Quality assessment of StratusOCT scans was determined by an experienced examiner. Good-quality scans of RNFL thickness had to have focused ocular fundus images, the signal strength had to be greater than 6, and a centered circular ring around the optic disc had to be present. Besides, a good-quality ONH printout, the machine had to determine automatically and correctly the edge of the ONH as the end of the retinal pigment epithelium–choriocapillaris layer with a signal strength greater than 6. 
We selected the average RNFL thickness, quadrant thickness (temporal, superior, nasal, inferior), 12-clock-hour (30° sector) RNFL thicknesses, and ONH analysis results (vertical integrated rim area, horizontal integrated rim area, disc area, cup area, rim area, cup/disc area ratio, horizontal cup/disc ratio, vertical cup/disc ratio) as our 25 input parameters. 
The perimetry and OCT examinations were all performed within a maximum period of 2 weeks. If the tests were performed on the same day, the perimetry examination was performed first. 
Data Processing Procedure I: Automatic Labeling with the SOM
The SOM, an unsupervised learning scheme, is particularly well suited for the combined task of mapping a high-dimension data distribution to a low-dimension topology so as to allow one to determine the number of clusters visually. 27 28 29 There are numerous applications of SOM for unsupervised clustering and visualization. 30 31 An overview of the multifaceted applications of SOM is given in Oja et al. 32  
LabelSOM, proposed by Andreas, 33 automatically labels the features of clusters generated by SOM. Automatic labeling is designed to filter automatically the large amounts of study variables in a cluster to form features. In this study, we present the LabelSOM neural network, used to determine the number of clusters and to select the features from StratusOCT as the input parameters for decision tree. Details on SOM parameter settings were as follows: (1) hextop topology function and linkdist distance function were used; (2) the neighborhood distance was set to 30; (3) dimensions of the map were 7 × 7; (4) the ordering phase learning rate was 0.7; (5) ordering phase steps were 40; and (6) the turning phase learning rate was 0.02. 
Data Processing Procedure II: Decision Tree
A decision tree is a chart that illustrates decision rules, which is a nonlinear discrimination method using a set of independent variables to split a sample into progressively smaller subgroups. A classification and regression tree (CART) is constructed by splitting subsets of the data set using all predictor variables to create two child nodes repeatedly. 
Two main objectives include a minimum tree size and maximum classification accuracy. A decision tree is pruned by an error-based method and by replacing nodes or a whole subtree to retain the classification’s accuracy. Association rules generated by the decision tree can be used to classify new cases with maximum accuracy. The software we used was AnswerTree (SPSS Inc., SPSS, Chicago, IL) Three basic parameter settings in AnswerTree are as follows: (1) maximum tree depth, 4; (2) minimum number of cases parent node, 3; and (3) minimum number of cases child node, 3. 
Results
Demographic Data
Good-quality RNFL and ONH images were achieved in 64 glaucomatous and 71 normal eyes. In addition, the optic disc did not reveal diffuse or focal rim thinning, optic disc hemorrhage, or RNFL defects in all the 71 normal eyes. The demographic details are presented in Table 1 . The mean age was 45.01 ± 12.41 years in the normal group and 42.53 ± 13.34 years in the glaucoma group. There was no significant difference in age between the two groups (P > 0.05). The mean deviation in the normal group was −0.55 ± 0.57 dB and in the glaucoma group was −4.30 ± 3.32 dB. There was a significant difference in mean deviation between the two groups (P < 0.0001). 
StratusOCT
Table 2lists the statistical results in both groups of 25 glaucoma variables measured from StratusOCT. There were 25 comparisons of variables; for α < 0.05, the Bonferroni adjustment required P < 0.002 (0.05/25 = 0.002) for the difference to be considered significant. The t-test revealed that all parameters were significantly different between both groups except temporal quadrant thickness, 8- to 10-clock-hour segment thickness, and disc area (P > 0.002). 
Table 3summarizes the sensitivities, specificities, and areas under the receiver operating characteristic (ROC) curves for 25 individual parameters. The inferior quadrant thickness was the best individual parameter for differentiating between normal eyes and glaucomatous eyes (ROC area, 0.879). The average RNFL thickness was the second best parameter (ROC area, 0.854). There was no significant difference in ROC areas between inferior quadrant thickness and average RNFL thickness (P = 0.247). In addition, vertical C/D ratio was the best individual parameter among ONH parameters (ROC area, 0.778). There was a significant difference in ROC areas between inferior quadrant thickness and vertical C/D ratio (P = 0.02). 
Feature Selection through LabelSOM
Table 4displays the average distance for five clusters from LabelSOM. The maximum average distances occur in cluster 3 and 4, which in total include 103 (55 + 48; 76%) eyes. Most of the subjects in cluster 3 were normal, whereas most of the patients in cluster 4 had glaucoma. Clusters 2 and 5 with only five and two patients, respectively, were ignored for further processing. The last three groups containing the fewest data numbers were ignored, and the largest two groups containing 103 data points were considered into the next processing phase. The first three parameters in the most two groups, which were nasal quadrant thickness, inferior quadrant thickness, average RNFL thickness, horizontal C/D ratio and vertical C/D ratio, were finally selected into CART to generate association rules for glaucoma detection. 
Determination of Association Rules
A decision tree was used to generate association rules on glaucoma detection. Five features were selected by LabelSOM, and the whole data set was split into training and testing. The training set contained 45 normal patients and 43 patients with glaucoma, and the testing set contained 26 normal patients and 21 patients with glaucoma. After pruning, there were only three nodes in the tree. The important StratusOCT variables for glaucoma detection were vertical C/D ratio and inferior quadrant thickness. Figure 1shows the association rules for glaucoma detection. The sensitivity, specificity, and overall accuracy were 73%, 92%, and 83%, respectively. Three association rules are summarized as follows:
  1.  
    If vertical C/D ratio is more than 0.60, the subject is classified as glaucomatous.
  2.  
    If vertical C/D ratio is less than 0.60 and inferior quadrant thickness is less than 106.5 μm, the subject is classified as glaucomatous
  3.  
    If vertical C/D ratio is less than 0.60 and inferior quadrant thickness is more than 106.5 μm, the subject is classified as normal.
Linear Discriminant Function
To compare the overall classification abilities of different classifiers, we classified the same data set by linear discriminant function (LDF). The sensitivity, specificity, and overall accuracy were 75%, 83%, and 81%, respectively, similar to CART and LDF. However, only the linear discriminant function can be generated from LDF, whereas we extracted more specific information on association rules related to glaucoma from CART. 
Discussion
Recently, several studies have been conducted to investigate and compare the discrimination performance of the RNFL thickness and optic nerve head (ONH) of the StratusOCT 24 25 34 35 36 37 38 In Mederios et al., 38 the highest discrimination power of the individual StratusOCT parameters was from the inferior quadrant thickness (ROC area, 0.91). Similar to our study, which had the highest ROC AUC (0.891) among all parameters, Manassakorn et al. 25 found that RNFL thickness at the 7-o’clock sector and the inferior quadrant and the vertical C/D ratio had the highest ROC AUC (0.93, 0.92, and 0.90, respectively). They used CART analysis to find the optimal parsimonious combination of parameters for glaucoma detection. The first parameter to be used for classification by CART was inferior quadrant RNFL thickness followed by vertical C/D ratio under the fast optic disc algorithm. The combination of these two parameters achieved the best classification (misclassification rate, 6.2%). 
The concept of machine learning has also been widely used in ophthalmology, especially for diagnosis of glaucoma. Some studies evaluated the application of machine classifiers in visual field interpretation of glaucoma. 39 40 41 Goldbaum et al. 39 reported that using the method of mixture of Gaussian (MoG), interpreted standard automated perimetry (SAP) better than the global indices of STATPAC. Their experience with machine learning classifiers indicates that there is additional useful information in visual field tests for glaucoma. Machine classifiers are able to discover and use perimetric information not obvious to experts in glaucoma. In another study, Sample et al. 40 reported that machine learning classifiers can learn complex patterns and trends in data and adapt to create a decision surface without the constraints imposed by statistical classifiers. This adaptation allowed the machine learning classifiers to identify abnormality in visual field converts much earlier than the traditional methods. In another study, also by Sample et al., 41 they found that without training-based diagnosis (unsupervised learning), the variational Bayesian mixture of factor analysis (vbMFA) identifies four important patterns of field loss in eyes with glaucomatous optic neuropathy in a manner consistent with years of clinical experience. Meanwhile, several automated classifiers were developed through different techniques, such as artificial neural networks (ANN), linear discriminant analysis (LDA), support vector machine (SVM), on glaucoma detection using summary reports from confocal scanning laser ophthalmoscopy (CSLO), 42 scanning laser polarimetry 43 (SLP) and StratusOCT. 24 25 Zangwill et al., 42 they reported that use of machine learning classifiers, trained with adequate cross-validation methods, can assist in identifying which combination of HRT parameters can best detect glaucoma. The application of these results in clinical practice could result in a more accurate diagnosis of glaucoma than is possible with any single optic disc parameter such as cup-disc ratio or rim area. 43 Bowd et al., 43 reported that results from RVM (relevance vector machine) and SVM (support vector machine) trained on SLP RNFL thickness measurements are similar and provide accurate classification of glaucomatous and healthy eyes. In our previous study, 24 we developed several automated classifiers and compared their performance using ANN, LDA, and Mahalanobis distance. Because the processing procedure for building those classifiers are complex and nontransparent, most of the results are unreadable and inexplicable. Although the automated classifiers showed promise for differentiating glaucomatous from normal eyes in the Taiwan Chinese population using summary data from StratusOCT, there was motivation to find more concise diagnostic rules, which was the main objective of this study. Reliable diagnostic regulations or precise disease association rules can be treated as handy diagnostic guidelines that help clinicians daily with glaucoma detection. Currently, there is limited research available regarding association rules for glaucoma detection. Our study is the first one to use the extraction of association rules to evaluate the glaucoma diagnosis in a Chinese population based on the summary data reports from Status OCT. 
In this study, the inferior quadrant RNFL thickness and the vertical C/D ratio were found to be the most major parameters on glaucoma detection from our decision tree. Three precise association rules with 86% accuracy on discriminating glaucomatous from normal eyes were established in our study. Compared our result with related work, Yan et al. 44 analyzed subclassifications of glaucoma via SOM-based clustering on the optic nerve head. Manassakorn et al. 25 reported the combination of inferior quadrant thickness and vertical C/D ratio for glaucoma detection, and the misclassification ratio was 6.2%. Our current result is consistent with the result of Manassakorn et al. In both studies, the most important two parameters are inferior quadrant thickness and vertical C/D ratio. However, there is still some difference in the association rules—the most possible reason being the different age groups. The average age in our study was younger than that in Manassakorn et al. Comparison between different groups was somewhat difficult. We had three association rules in our decision tree, while Manassakorn et al. had four. The first key parameter we found in our decision tree was vertical C/D ratio, whereas the first key parameter in their decision tree was inferior quadrant thickness. Although there are some differences between our study and theirs, we confirmed the previous work of Manassakorn et al., even with a different study population. As they pointed out in their study, there were still some other pair-wise potential combinations of different parameters what could work as well. The CART analysis is potentially helpful for defining clinically useful cutoff points that have immediate application for the clinicians to evaluate whether the optic disc is glaucomatous or nonglaucomatous. Further study is needed in the near future to find more consistent rules for glaucoma diagnosis. 
Our result is also consistent with the general concept for glaucomatous optic disc evaluation that vertical C/D ratio is important. As far as we know, ophthalmoscopic estimation of vertical cup-to-disc ratio (VCDR) of the ONH is very important in the management and follow-up of glaucoma; but it has only a moderate interobserver agreement and relies on observer experience. 45 Recently, there were some studies comparing OCT analysis with stereophotography. One of them, a report by Arnalich-Montiel et al. 46 found that ONH analysis with OCT shows good agreement with slit lamp indirect ophthalmoscopy for horizontal C/D ratio and vertical C/D ratio evaluation in greater C/D ratio and disc areas. However, for smaller C/D ratio and disc areas, the OCT values tended to be higher. Medeiros et al. 38 reported that there was no difference in the mean vertical C/D ratio between stereophotography and OCT. However, for lower values of vertical C/D ratio, the OCT measurements were higher; whereas for greater vertical C/D ratio, the OCT measurements were lower. A recent study by Arthur et al. 47 compared the level of agreement between subjective (stereoscopic ONH photographs) and objective methods (HRT II, StratusOCT) in estimating horizontal and vertical cup-to-disc ratios (HCDR and VCDR, respectively) to determine whether objective techniques may be used as surrogates for subjective cup-to-disc estimation. Their results showed that the agreement in subjectively assessed HCDR and VCDR was substantial (ICC = 0.84 and 0.85, respectively), and for all three methods, overall agreement was good (ICC = 0.75 and 0.77 for the HCDR and VCDR, respectively). StratusOCT provided the largest overall mean ± SD. HCDR (0.68 ± 0.14) and VCDR (0.62 ± 0.13). Although the overall agreement between various methods was good, the mean estimates were significantly different. Therefore, evaluating C/D ratio with the StratusOCT is still imperfect at this moment, but it still provides us more reliable and reproducible measures of optic disc topography. Additional studies are needed to evaluate the sources of variability, their level of significance, and longitudinal agreement between various methods of the CDR estimation. 47  
Although our result is interesting and promising, there are some limitations in our study. First, the substrate for studies is usually a clinic-based population of patients with glaucoma. These patients have been identified on the basis of particular patterns of structural and functional abnormality that meet preconceived notions that bias the outcome of the comparisons. 48 For example, inclusion criteria for normal subjects included a normal optic nerve appearance judged from examination of stereoscopic optic disc photographs. This criterion was necessary to avoid including subjects with glaucomatous optic neuropathy but normal visual fields in the control group, but as a matter of fact, not all normal subjects have normal looking optic nerves. Therefore, this could overestimate the diagnostic accuracy of OCT instruments. However, this problem is a common limitation in this type of case–control study, just as in the other studies mentioned. 38 49 Besides, there were some problems in the imaging selection process. To increase accuracy and obtain good-quality scans, we excluded the patients with marked peripapillary atrophy or some optic disc shape that could not be analyzed by StratusOCT software version A 4.0.1, making those individuals poor candidates for OCT examination. We know that there are some degrees of peripapillary atrophy usually present in the population of an age to be affected by glaucoma. However, it is inevitable to have image selection bias in a technical imaging study. One more limitation is the small sample used to generate the association rules. Larger sample sizes are recommended to provide more precise and robust estimations for glaucoma diagnosis. Therefore, caution should be used when applying the results in this first study of combining LabelSOM and decision-tree methods to daily glaucoma practice. 
In conclusion, our results can be used as the basis for further improving the diagnostic accuracy of glaucoma in the Taiwan Chinese population in the near future. The precise association rules induced from a novel application of the decision tree may enhance glaucoma detection. 
 
Table 1.
 
Subject Demographics
Table 1.
 
Subject Demographics
Normal (n = 71) Glaucoma (n = 64) P *
Gender
 Male, n (%) 30 (42%) 41 (63%)
 Female, n (%) 41 (58%) 23 (37%)
Age, y (mean ± SD) 45.01 ± 12.41 42.53 ± 13.34 0.2667
MD, dB (mean ± SD) −0.55 ± 0.57 −4.30 ± 3.32 <0.0001
 >0 10 0
 0 to −1 51 1
 −1 to −2 10 12
 −2 to −3 0 14
 −3 to −6 0 26
 −6 to −10 0 5
 −10 to −15 0 6
Table 2.
 
Stratus OCT Glaucoma Variables Included in the 25 Input Set
Table 2.
 
Stratus OCT Glaucoma Variables Included in the 25 Input Set
Parameter Normal Glaucoma P *
Average RNFL thickness (μm) 111.72 ± 12.66 87.73 ± 19.81 <0.0001
Temporal quadrant 81.86 ± 13.84 77.75 ± 24.30 0.2369
Superior quadrant 136.96 ± 16.27 107.92 ± 28.36 <0.0001
Nasal quadrant 85.75 ± 21.67 63.53 ± 20.13 <0.0001
Inferior quadrant 142.31 ± 19.29 101.73 ± 28.98 <0.0001
Clock-hour segment thickness (μm)
 12 134.00 ± 24.93 107.95 ± 35.12 <0.0001
 11 (superior temporally) 147.62 ± 20.69 117.34 ± 33.07 <0.0001
 10 96.13 ± 15.30 86.85 ± 29.71 0.0273
 9 (temporal) 66.24 ± 12.46 65.56 ± 23.84 0.8392
 8 84.32 ± 19.88 81.50 ± 30.11 0.5262
 7 (inferior temporally) 159.24 ± 25.99 115.58 ± 43.43 <0.0001
 6 150.46 ± 26.07 105.46 ± 34.51 <0.0001
 5 118.38 ± 23.78 85.05 ± 25.87 <0.0001
 4 82.41 ± 23.33 62.41 ± 21.89 <0.0001
 3 (nasal) 75.51 ± 23.04 56.09 ± 20.86 <0.0001
 2 100.08 ± 26.36 72.61 ± 24.53 <0.0001
 1 130.48 ± 23.06 99.46 ± 31.72 <0.0001
ONH analysis result
 Vertical integrated rim area (mm3) 0.54 ± 0.51 0.29 ± 0.24 <0.0001
 Horizontal integrated rim width (mm2) 1.81 ± 0.24 1.49 ± 0.34 <0.0001
 Disc area (mm2) 2.55 ± 0.43 2.54 ± 0.63 0.9267
 Cup area (mm2) 0.69 ± 0.37 1.17 ± 0.66 <0.0001
 Rim area (mm2) 1.86 ± 0.40 1.40 ± 0.54 <0.0001
 Cup/disc area ratio 0.26 ± 0.12 0.46 ± 0.23 <0.0001
 Horizontal cup/disc ratio 0.54 ± 0.14 0.69 ± 0.16 <0.0001
 Vertical cup/disk ratio 0.46 ± 0.12 0.62 ± 0.16 <0.0001
Table 3.
 
The Sensitivity and Specificity of Individual Parameters
Table 3.
 
The Sensitivity and Specificity of Individual Parameters
Parameter Sensitivity ROC Area (%)
Specificity of 80% Specificity of 90%
Average RNFL thickness (μm) 68.7 67.2 85.4
Temporal quadrant 43.7 32.8 56.2
Superior quadrant 68.7 51.6 81.2
Nasal quadrant 65.6 48.4 77.7
Inferior quadrant 76.6 67.2 87.9
Clock-hour segment thickness (μm)
 12 51.6 42.2 72.2
 11 50.0 48.4 76.5
 10 45.3 40.6 61.3
 9 35.9 26.6 55.0
 8 35.9 21.9 54.1
 7 64.1 53.1 79.1
 6 76.6 71.9 84.7
 5 73.4 57.8 83.0
 4 57.8 26.6 73.9
 3 51.6 45.3 74.3
 2 65.6 45.3 78.0
 1 65.6 54.7 79.3
ONH analysis result
 Vertical integrated rim area (mm3) 59.4 42.2 75.9
 Horizontal integrated rim width (mm2) 54.7 42.2 77.4
 Disc area (mm2) 26.6 18.8 53.8
 Cup area (mm2) 48.4 59.4 73.8
 Rim area (mm2) 56.2 45.3 75.7
 Cup/disc area ratio 59.4 64.1 76.6
 Horizontal cup/disc ratio 62.5 54.7 74.9
 Vertical cup/disk ratio 65.6 59.4 77.8
Table 4.
 
Average Distance for Clusters
Table 4.
 
Average Distance for Clusters
Network Normal Glaucoma Total Number Weight Average Distance
Cluster 1 12 13 25 0.028396
Cluster 2 4 1 5 0.018924
Cluster 3 47 8 55 0.030035
Cluster 4 7 41 48 0.035306
Cluster 5 1 1 2 0.014720
Figure 1.
 
Association rules generated by the decision tree. The original input was five features selected by LabelSOM. There are three nodes after pruning.
Figure 1.
 
Association rules generated by the decision tree. The original input was five features selected by LabelSOM. There are three nodes after pruning.
QuigleyHA, AddicksEM, GreenRW. Optic nerve damage in human glaucoma. III. Quantitative correlation of nerve fiber loss and visual field defect in glaucoma, ischemic neuropathy, papilledema, and toxic neuropathy. Arch Ophthalmol. 1982;100:135–146. [CrossRef] [PubMed]
SommerA, KatzJ, QuigleyHA, et al. Clinically detectable nerve fiber layer atrophy precedes the onset of glaucomatous field loss. Arch Ophthalmol. 1991;109:77–83. [CrossRef] [PubMed]
GreanyMJ, HoffmanDC, Garway-HeathDF, et al. Comparisons of optic nerve imaging methods to distinguish normal eyes from those with glaucoma. Invest Ophthalmol Vis Sci. 2002;43:140–145. [PubMed]
Sanchez-GaleanaC, BowdC, BlumenthalEZ, et al. Using optical imaging summary data to detect glaucoma. Ophthalmology. 2001;108:1812–1818. [CrossRef] [PubMed]
UchidaH, BrigattiL, CaprioliJ. Detection of structural damage from glaucoma with confocal laser image analysis. Invest Ophthalmol Vis Sci. 1996;37:2393–2401. [PubMed]
WollsteinG, Garway-HeathDF, HitchingsRA. Identification of early glaucoma cases with the scanning laser ophthalmoscope. Ophthalmology. 1998;105:1557–1563. [CrossRef] [PubMed]
ChoplinNT, LundyDC, DreherAW. Differentiating patients with glaucoma from glaucoma suspects and normal subjects by nerve fiber layer assessment with scanning laser polarimetry. Ophthalmology. 1998;105:2068–2076. [CrossRef] [PubMed]
HohST, GreenfieldDS, MistlbergerA, et al. Optical coherence tomography and scanning laser polarimetry in normal, ocular hypertensive, and glaucomatous eyes. Am J Ophthalmol. 2000;129:129–135. [CrossRef] [PubMed]
CioffiGA, RobinAL, EastmanRD, et al. Confocal laser scanning ophthalmoscope: reproducibility of optic nerve head topographic measurements with the confocal laser scanning ophthalmoscope. Ophthalmology. 1993;100:57–62. [CrossRef] [PubMed]
ZangwillL, ShakibaS, CaprioliJ, et al. Agreement between clinicians and a confocal scanning laser ophthalmoscope in estimating cup/disk ratios. Am J Ophthalmol. 1995;119:415–421. [CrossRef] [PubMed]
WeinrebRN, ShakibaS, SamplePA, et al. Association between quantitative nerve fiber layer measurement and visual field loss in glaucoma. Am J Ophthalmol. 1995;120:732–738. [CrossRef] [PubMed]
SchumanJS, HeeMR, PuliafitoCA, et al. Quantification of nerve fiber layer thickness in normal and glaucomatous eyes using optical coherence tomography. Arch Ophthalmol. 1995;113:586–596. [CrossRef] [PubMed]
HohST, IshikawaH, GreenfieldDS, et al. Peripapillary nerve fiber layer thickness measurement reproducibility using scanning laser polarimetry. J Glaucoma. 1998;7:712–715.
HornFK, JonasJB, MartusP, et al. Polarimetric measurement of retinal nerve fiber layer thickness in glaucoma diagnosis. J Glaucoma. 1999;8:353–362. [PubMed]
ChenHY, HungPT, HungTJ. Optical coherence tomography study on retinal nerve fiber layer thickness in open angle glaucoma (abstract). Am J Ophthalmol. 2005;139(suppl)S53.
HeeMR, IzattJA, SwansonEA, et al. Optical coherence tomography of the human retina. Arch Ophthalmol. 1995;113:325–332. [CrossRef] [PubMed]
SchumanJS, HeeMR, AryaAV, et al. Optical coherence tomography: a new tool for glaucoma diagnosis. Curr Opin Ophthalmol. 1995;6:89–95. [CrossRef] [PubMed]
HuangD, SwansonEA, LinCP, et al. Optical coherence tomography. Science. 1991;254:1178–1181. [CrossRef] [PubMed]
Gurses-OzdenR, TengC, VessaniR, et al. Macular and retinal nerve fiber layer thickness measurement reproducibility using optical coherence tomography (OCT-3). J Glaucoma. 2004;13:238–244. [CrossRef] [PubMed]
PaunescuLA, SchumanJS, PriceLL, et al. Reproducibility of nerve fiber thickness, macular thickness, and optic nerve head measurements using StratusOCT. Invest Ophthalmol Vis Sci. 2004;45:1716–1724. [CrossRef] [PubMed]
KononenkoI. Machine learning for medical diagnosis: history, state of the art and perspective. Artif Intell Med. 2001;23:89–109. [CrossRef] [PubMed]
WetterT. Lessons learnt from bringing knowledge-based decision support into routine use (editorial). Artif Intell Med. 2002;24:195–203. [CrossRef] [PubMed]
Burgansky-EliashZ, WollsteinG, ChuT, et al. Optical coherence tomography machine learning classifiers for glaucoma detection: a preliminary study. Invest Ophthalmol Vis Sci. 2005;46:4147–4152. [CrossRef] [PubMed]
HuangML, ChenHY. Development and comparison of automated classifiers for glaucoma diagnosis using Stratus optical coherence tomography. Invest Ophthalmol Vis Sci. 2005;46:4121–4129. [CrossRef] [PubMed]
ManassakornA, Nouri-MahdaviK, CaprioliJ. Comparison of retinal nerve fiber layer thickness and optic disk algorithms with optical coherence tomography to detect glaucoma. Am J Ophthalmol. 2006;141:105–115. [CrossRef] [PubMed]
CaprioliJ, ParkHJ, UgurluS, et al. Slope of the peripapillary nerve fiber layer surface in glaucoma. Invest Ophthalmol Vis Sci. 1998;39:2321–2328. [PubMed]
LeeRCT, SlagleJR, BlumH. A Triangulation method for the sequential mapping of points from N-space to two-space. IEEE Trans Comp. 1977;26:288–292.
PykettCE. Improving the efficiency of Sammon’s nonlinear mapping by using clustering archetypes. Electron Lett. 1978;14:799–800. [CrossRef]
KohonenT. Self-organized formation of topologically correct feature maps. Biol Cybern. 1982;43:59–69. [CrossRef]
SomervuoP, KohonenT. Self-organizing maps ad learning vector quantization for feature sequences. Neur Process Lett. 1999;10:151–159. [CrossRef]
TokutakaH, YoshiharaK, FujimuraK, et al. Application of self-organizing maps (som) to auger electron spectroscopy (aes). Surf Interface Anal. 1999;27:783–788. [CrossRef]
OjaE KaskiS eds. Kohonen Maps. 2003;Elsevier Science B.V. Amsterdam.
AndreasR. Automatic Labeling of Self-Organizing Maps: Making a Treasure- Map Reveal its Secrets. Proceedings of the 3rd Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD ‘99) Bejing, China: 1999, April, 26–28. ;
ChenHY, HuangML. Discrimination between normal and glaucomatous eyes using Stratus optical coherence tomography in Taiwan Chinese subjects. Graefe Arch Clin Exp Ophthalmol. 2005;243:894–902. [CrossRef]
WollstenG, IshikawaH, WangJ, et al. Comparison of three optical coherence tomography scanning areas for detection of glaucomatous damage. Am J Ophthalmol. 2005;139:39–43. [CrossRef] [PubMed]
MedeirosFA, ZangwillLM, BowdC, et al. Influence of disease severity and optic disc size on the diagnostic performance of imaging instruments in glaucoma. Invest Ophthalmol Vis Sci. 2006;47:1008–1015. [CrossRef] [PubMed]
LeungCK, ChanWM, YungWH. Comparison of macular and peripapillary measurements for the detection of glaucoma: an optical coherence tomography study. Ophthalmology. 2005;112:391–400. [CrossRef] [PubMed]
MedeirosFA, ZangwillLM, BowdC, et al. 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. [CrossRef] [PubMed]
GoldbaumMH, SamplePA, ChanK, et al. Comparing machine learning classifiers for diagnosing glaucoma from standard automated perimetry. Invest Ophthalmol Vis Sci. 2002;43:162–169. [PubMed]
SamplePA, GoldbamMH, ChanK, et al. Using machine learning classifiers to identify glaucomatous change earlier in standard visual fields. Invest Ophthalmol Vis Sci. 2002;43:2660–2665. [PubMed]
SamplePA, ChanK, BodenC, et al. Using unsupervised learning with variational bayesian mixture of factor analysis to identify patterns of glaucomatous visual field defects. Invest Ophthalmol Vis Sci. 2004;45:2596–2605. [CrossRef] [PubMed]
ZangwillLM, ChanK, BowdC, et al. Heidelberg retina tomograph measurements of the optic disc and parapapillary retina for detecting glaucoma analyzed by machine learning classifiers. Invest Ophthalmol Vis Sci. 2004;45:3144–3151. [CrossRef] [PubMed]
BowdC, MedeirosFA, ZhangZ, et al. Relevance vector machine and support vector machine classifier analysis of scanning laser polarimetry retinal nerve fiber layer measurements. Invest Ophthalmol Vis Sci. 2005;46:1322–1329. [CrossRef] [PubMed]
YanS, AbidiSS, ArtesPH. Analyzing sub-classification of glaucoma via SOM based clustering of optic nerve images. Stud Health Technol Inform. 2005;116:483–488. [PubMed]
TielschJM, KatzJ, QuigleyHA, et al. Intraobserver and interobserver agreement in measurements of optic disc characteristics. Ophthalmology. 1988;95:350–356. [CrossRef] [PubMed]
Arnalich-MontielF, Munoz-NegreteFJ, RebolledaG, et al. Cup-to-disc ratio: agreement between slit-lamp indirect ophthalmoscopic estimation and stratus optical coherence tomography measurement. Eye. .Published online May 5, 2006
ArthurSN, AldridgeAJ, De Leon-OrtegaJ, et al. Agreement in assessing cup-to-disc ratio measurement among stereoscopic optic nerve head photographs, HRT II, and StratusOCT. J Glaucoma. 2006;15:183–189. [CrossRef] [PubMed]
Garway-HeathDF, HitchingsRA. Sources of bias in studies of optic disc and retinal nerve fiber morphology. Br J Ophthalmol. 1988;82:986.
BournePR, MedeirosFA, BowdC, et al. Comparability of retinal nerve fiber layer thickness measurements of optical coherence tomography instruments. Invest Ophthalmol Vis Sci. 2005;46:1280–1285. [CrossRef] [PubMed]
Figure 1.
 
Association rules generated by the decision tree. The original input was five features selected by LabelSOM. There are three nodes after pruning.
Figure 1.
 
Association rules generated by the decision tree. The original input was five features selected by LabelSOM. There are three nodes after pruning.
Table 1.
 
Subject Demographics
Table 1.
 
Subject Demographics
Normal (n = 71) Glaucoma (n = 64) P *
Gender
 Male, n (%) 30 (42%) 41 (63%)
 Female, n (%) 41 (58%) 23 (37%)
Age, y (mean ± SD) 45.01 ± 12.41 42.53 ± 13.34 0.2667
MD, dB (mean ± SD) −0.55 ± 0.57 −4.30 ± 3.32 <0.0001
 >0 10 0
 0 to −1 51 1
 −1 to −2 10 12
 −2 to −3 0 14
 −3 to −6 0 26
 −6 to −10 0 5
 −10 to −15 0 6
Table 2.
 
Stratus OCT Glaucoma Variables Included in the 25 Input Set
Table 2.
 
Stratus OCT Glaucoma Variables Included in the 25 Input Set
Parameter Normal Glaucoma P *
Average RNFL thickness (μm) 111.72 ± 12.66 87.73 ± 19.81 <0.0001
Temporal quadrant 81.86 ± 13.84 77.75 ± 24.30 0.2369
Superior quadrant 136.96 ± 16.27 107.92 ± 28.36 <0.0001
Nasal quadrant 85.75 ± 21.67 63.53 ± 20.13 <0.0001
Inferior quadrant 142.31 ± 19.29 101.73 ± 28.98 <0.0001
Clock-hour segment thickness (μm)
 12 134.00 ± 24.93 107.95 ± 35.12 <0.0001
 11 (superior temporally) 147.62 ± 20.69 117.34 ± 33.07 <0.0001
 10 96.13 ± 15.30 86.85 ± 29.71 0.0273
 9 (temporal) 66.24 ± 12.46 65.56 ± 23.84 0.8392
 8 84.32 ± 19.88 81.50 ± 30.11 0.5262
 7 (inferior temporally) 159.24 ± 25.99 115.58 ± 43.43 <0.0001
 6 150.46 ± 26.07 105.46 ± 34.51 <0.0001
 5 118.38 ± 23.78 85.05 ± 25.87 <0.0001
 4 82.41 ± 23.33 62.41 ± 21.89 <0.0001
 3 (nasal) 75.51 ± 23.04 56.09 ± 20.86 <0.0001
 2 100.08 ± 26.36 72.61 ± 24.53 <0.0001
 1 130.48 ± 23.06 99.46 ± 31.72 <0.0001
ONH analysis result
 Vertical integrated rim area (mm3) 0.54 ± 0.51 0.29 ± 0.24 <0.0001
 Horizontal integrated rim width (mm2) 1.81 ± 0.24 1.49 ± 0.34 <0.0001
 Disc area (mm2) 2.55 ± 0.43 2.54 ± 0.63 0.9267
 Cup area (mm2) 0.69 ± 0.37 1.17 ± 0.66 <0.0001
 Rim area (mm2) 1.86 ± 0.40 1.40 ± 0.54 <0.0001
 Cup/disc area ratio 0.26 ± 0.12 0.46 ± 0.23 <0.0001
 Horizontal cup/disc ratio 0.54 ± 0.14 0.69 ± 0.16 <0.0001
 Vertical cup/disk ratio 0.46 ± 0.12 0.62 ± 0.16 <0.0001
Table 3.
 
The Sensitivity and Specificity of Individual Parameters
Table 3.
 
The Sensitivity and Specificity of Individual Parameters
Parameter Sensitivity ROC Area (%)
Specificity of 80% Specificity of 90%
Average RNFL thickness (μm) 68.7 67.2 85.4
Temporal quadrant 43.7 32.8 56.2
Superior quadrant 68.7 51.6 81.2
Nasal quadrant 65.6 48.4 77.7
Inferior quadrant 76.6 67.2 87.9
Clock-hour segment thickness (μm)
 12 51.6 42.2 72.2
 11 50.0 48.4 76.5
 10 45.3 40.6 61.3
 9 35.9 26.6 55.0
 8 35.9 21.9 54.1
 7 64.1 53.1 79.1
 6 76.6 71.9 84.7
 5 73.4 57.8 83.0
 4 57.8 26.6 73.9
 3 51.6 45.3 74.3
 2 65.6 45.3 78.0
 1 65.6 54.7 79.3
ONH analysis result
 Vertical integrated rim area (mm3) 59.4 42.2 75.9
 Horizontal integrated rim width (mm2) 54.7 42.2 77.4
 Disc area (mm2) 26.6 18.8 53.8
 Cup area (mm2) 48.4 59.4 73.8
 Rim area (mm2) 56.2 45.3 75.7
 Cup/disc area ratio 59.4 64.1 76.6
 Horizontal cup/disc ratio 62.5 54.7 74.9
 Vertical cup/disk ratio 65.6 59.4 77.8
Table 4.
 
Average Distance for Clusters
Table 4.
 
Average Distance for Clusters
Network Normal Glaucoma Total Number Weight Average Distance
Cluster 1 12 13 25 0.028396
Cluster 2 4 1 5 0.018924
Cluster 3 47 8 55 0.030035
Cluster 4 7 41 48 0.035306
Cluster 5 1 1 2 0.014720
×
×

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.

×