April 2005
Volume 46, Issue 4
Free
Clinical and Epidemiologic Research  |   April 2005
Detecting Progression of Nuclear Sclerosis by Using Human Grading Versus Semiautomated Computer Grading
Author Affiliations
  • Barbara E. K. Klein
    From the Departments of Ophthalmology and Visual Sciences,
  • Larry Hubbard
    Fundus Photography Reading Center, University of Wisconsin, Madison, Wisconsin.
  • Nicola J. Ferrier
    Mechanical Engineering, and
  • Ronald Klein
    From the Departments of Ophthalmology and Visual Sciences,
  • Daniel J. Klein
    Mechanical Engineering, and
  • Kristine E. Lee
    From the Departments of Ophthalmology and Visual Sciences,
  • Andrew Ewen
    From the Departments of Ophthalmology and Visual Sciences,
  • Karl Jensen
    From the Departments of Ophthalmology and Visual Sciences,
  • Michael D. Evans
    Biostatistics and Medical Informatics, and the
Investigative Ophthalmology & Visual Science April 2005, Vol.46, 1155-1162. doi:https://doi.org/10.1167/iovs.04-0239
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Barbara E. K. Klein, Larry Hubbard, Nicola J. Ferrier, Ronald Klein, Daniel J. Klein, Kristine E. Lee, Andrew Ewen, Karl Jensen, Michael D. Evans; Detecting Progression of Nuclear Sclerosis by Using Human Grading Versus Semiautomated Computer Grading. Invest. Ophthalmol. Vis. Sci. 2005;46(4):1155-1162. https://doi.org/10.1167/iovs.04-0239.

      Download citation file:


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

      ×
  • Supplements
Abstract

purpose. To assess indices of nuclear sclerosis derived from digitized images made from color (slide) photographs.

methods. Film-based slit lamp images taken at baseline and at 5- and 10-year follow-up examinations of the Beaver Dam Eye Study cohort were digitized, and optical traces were taken along an axis through the center of the cornea and lens. Four indices of the severity of sclerosis were calculated based on the optical densities. The associations of the original Beaver Dam grades and these indices to age, vision, and change in severity of sclerosis over two subsequent visits were compared.

results. At baseline photographs, the Spearman correlation between age and severity was 0.65 for the original film-based grading (n = 4518 right eyes) and varied between 0.46 and 0.71 for the measures from digitized images. Correlations of the indices to visual acuity were 0.38 for the film-based grading and ranged from 0.32 to 0.38 for the other indices. The authors assume that nuclear sclerosis does not regress and the percentage of regression is a reflection of error in grading. The percentage of regression and progression of sclerosis over 5- and 10-year intervals was determined for each index. After 5 years, 48.2% progressed and 4.9% regressed, using the Beaver Dam grades; progression occurred in 4.9% to 9.9%, and regression occurred in 4.5% to 7.0% for the other indices. After 10 years, 61.9% progressed and 3.2% regressed using the Beaver Dam grades; progression occurred in 8.0% to 19.7%, and regression occurred in 2.6% to 9.7% for the other indices.

conclusions. Semiautomated grading of the digitized images can be used to process thousands of images with little oversight by a trained grader. Indices of sclerosis that closely parallel human grading in their relationships to age and visual acuity can be easily computed. However, the indices appear to identify significantly less progression of nuclear sclerosis than does human grading. Further development to define a useful metric for identifying severity and progression of nuclear sclerosis is needed.

Nuclear sclerosis is progressive throughout life. Although changes in the nucleus of the lens can be seen through a dilated pupil by trained observers, such changes are not usually labeled as cataract until there is obscuration of the discrete lamellae and sulcus of the lens nucleus. Progressive nuclear sclerosis may be associated with changes in refractive error and in visual acuity, as well as other measures of visual function. Nuclear cataract was the most common type to precede lens extraction in the Beaver Dam Eye Study cohort. 1  
Several schemes have been developed to evaluate nuclear sclerosis for use in clinical and epidemiologic studies. Some are based on clinical examination. 2 3 4 5 6 Because nuclear sclerosis is a slowly progressive process throughout life in Western populations, it is difficult to assign a change from examination to examination when the interval between examinations is relatively short, especially when there may be different examiners judging severity. Therefore, for most long-term longitudinal studies (e.g., for incidence studies or for clinical trials), photographic documentation has become the preferred technique for judging severity of nuclear sclerosis. 
Several protocols have been developed for photographic or other imaging documentation. 7 8 9 10 Usually, the photographs are taken with a slit lamp camera which produces an image that replicates a clinical slit lamp view at the instant the photograph is taken. The advantage of these systems is that they produce images that are similar to the view of the lens as seen by a clinician at the slit lamp. Different systems have been developed to grade such slit lamp images. 10 11 12 13 14 15 16 These systems are similar, in that they entail comparing study photographs to sets of slit lamp photographic standards. 
There are different choices of cameras available for slit lamp photography of the lens. Those cameras that are designed according to the Scheimpflug principle—that is, the plane of the film parallels the optical section through the lens—can provide an evenly focused representation of the section of the lens. 7 9 14 17 Slit lamp cameras that use a conventionally designed film arrangement—that is, the film parallels the flat camera back and plane of the camera lens and the illuminated section of the lens is at 30° or 45° from the viewer or the camera—can provide an image with best focus corresponding to only part of the image of the lens (ideally the sulcus of the nucleus). Other parts of the image are not as well focused. 8 10 11 The advantage of cameras of this design is that they are relatively inexpensive, readily available, and, in general, they are reliable (or easily repairable) in a field setting. 
Digital rather than film-based imaging systems have recently been used with Scheimpflug design cameras. 18 19 20 Magno et al. 20 reported that with the use of such a system, variability in estimates of progression were decreased compared with a clinical grading system, so that they were able to detect change in sclerosis over only a 1-year interval. This system uses differences in density, as measured in reflected light from an optical trace through the center of the lens as a measure of sclerosis. The outcome is a truly continuous measure, unlike the human grading systems, which measure against standard photographs. (Some investigators attempt to approximate a continuous measure with film-based human grading by having a grader estimate in tenths between standards. 3 8 11 ) The greatest disadvantage of the Scheimpflug digital system is its cost. Also, like other imaging–grading systems, it has difficulty in estimating the severity of sclerosis when other lens opacities are also present. 21  
No matter which camera system is used, certain specific lens regions are more informative about the severity and change in severity of nuclear sclerosis (or lens density) than others. 14 15 22 Qian et al. 14 22 indicate that eliminating information from the lens cortex is important in defining a “common lens nuclear area” and measuring the optical density in that area. Duncan et al. 15 defines a “region of interest” that similarly excludes the lens cortex. Within the nucleus there are identifiable regions of density, both anterior and posterior to a central region of decreased density (sulcus). This defines the important parts of the lens for estimating the severity of nuclear sclerosis. 23 Qian et al. 24 have also suggested that densitometry—taking weighted information from specific regions of the nucleus—may permit the detection of changes in nuclear density over a short period. 
There are large epidemiologic studies and clinical trials that have been completed and others that are ongoing that employ the non-Scheimpflug, slit lamp, film-based camera technology with images graded by human graders. 10 11 The variability in the grading of such photographs leads to relatively broad confidence intervals in estimating the prevalence of nuclear cataract, in estimating the strengths of association of a variety of factors to severity and in measuring change over time. Gradings derived from semiautomated densitometry measurements of the photographs from these projects might diminish grading variability and therefore permit better estimates of change. 
We have developed a semiautomated grading procedure for nuclear sclerosis from black-and-white digitized images that were derived from color-based film images. We compare this technique with the results of human grading of the original film-based color images in data derived from a longitudinal population-based study. 
Methods
Visual acuities were measured according to a modification of the Early Treatment Diabetic Retinopathy Study protocol for each eye at each visit. 25 Slit lamp photographs of the lenses of participants in the Beaver Dam Eye Study were taken at baseline and 5 and 10 years after. Signed consent was obtained from each participant at each examination. Institutional review board approval was obtained yearly, and the protocol adhered to the tenets of the Declaration of Helsinki. The age range of the 4926 persons seen at baseline was 43 to 86 years. The photographs were taken after pharmacologic dilation of the pupil with a specially modified slit lamp camera (Topcon, Paramus, NJ; using Ektachrome 200ASA film; Eastman Kodak, Rochester, NY). This camera is not designed according to the Scheimpflug principle. The camera was modified to fix the angle of the slit beam at 45° from the visual axis, the beam width was fixed at 0.3 mm, and the height at 9 mm, and the flash intensity was set at 5. The slit beam illumination was always to the left of the photographer. Unmovable fixation targets were added to the camera. 10 The camera was evaluated for these parameters every 6 months during the 2.5 years of each examination phase, or more frequently if any question about its function arose. The photographer was instructed to focus the camera at the sulcus of the lens nucleus. All photographs were developed by a commercial laboratory and sent to the study offices for grading the severity of sclerosis. The grading scheme required comparison of the participant’s photograph with photographic reference standards of increasing severity of sclerosis. Graders assigned a grade for the severity of sclerosis by comparing primarily the density of the sulcus and secondarily the distinctness of the other nuclear landmarks with those in standard photographs. The scheme resulted in five levels of increasing severity. The grader evaluated overall photograph quality and plane of focus of the image. The grader was masked to subject characteristics. Right and left eyes were graded independently. A quality control program for photography and grading was ongoing through all examination phases. Intergrader reliability was 64.7% for exact agreement and 99.8% within one category. 10  
The original images on film were first obtained during the 1988 to 1990 examination. Digital imaging cameras were not widely available and were (and still are) costly. Our current intent was to take advantage of the many attractive features of digital techniques that are now available that might facilitate automated grading of our images. To do this, we had to digitize our film-based images. The scanner was chosen based on availability, cost, and reputed fidelity of the scanning process. We cannot directly compare our digitized film-based images with images taken originally with a digital slit lamp camera, because we did not have one available. Because our comparisons are all made on images captured, processed, and digitized the same way, our internal comparisons will be consistent. 
The slit lamp photographic images were scanned into gray-scale images with 256 integer levels (Cool Scan LS-2000, ver. 1.31 firmware, 2.5 software; Nikon, Melville, NY). The slides were removed from their stored plastic sleeves and placed face up (label to the right); scanning parameters were selected for black and white scanning; and the images were previewed to ensure that the slit lamp image was centered and that both the front and back edges of the image were contained within the desired “box.” The image was scanned and saved to a subdirectory and filed according to digit code. The quality control procedures entailed checking scanning parameters against the written protocols at the beginning of each session. The first scan of every session was a repeat of the last scan of the previous session. Rescanned photographs were visually inspected for any difference between it and the original scan. If no differences were found, the image was saved to a unique file. If any differences were noted on inspection, all parameters for the original and rescanned images were compared and the software and hardware were checked to determine whether there were errors in the scanning parameters. Any corrections were made, and the image was rescanned until it was judged to be the same as the original. These adjustments were needed infrequently. 
All image files were processed in batch mode through a program that automatically detected or found a central visual axis and generated a trace along that axis (which eliminated artifacts including keyhole-shaped light artifacts appearing in the interior chamber). From the trace, several landmarks were identified. Algorithms for identification of axis and landmarks were tested on sample photographs, the quality of which was judged to be excellent and yielded consistent grading. A senior grader (AE) reviewed, and was allowed to modify, axis and landmark placement for all images, including those sent to a “manual” folder by the computer program when automatic identification was not possible. Such modification of landmarks relevant to calculating an index of sclerosis was needed in approximately 1.7% of images (Ferrier N, et al. IOVS 2002;43:ARVO E-Abstract 435). These techniques were applied to slit lamp images of right and left eyes from participants in the Beaver Dam Study at the baseline and at the 5- and 10-year follow-up evaluations. The total number of images examined was 20,093. 
Finding the Axis
We assumed that the visual axis bisects the nucleus horizontally. The cornea was readily identified in all images, regardless of the degree of sclerosis. In most cases the corneal bow was roughly symmetric with respect to the axis, and, in virtually all cases, the cortex was symmetric with respect to this axis as well. However, the surface of the cortex was often difficult to detect. Thus, we used the location and shape of the corneal bow in initially identifying the location of the axis (and used the cortex when a “good” edge was obtained). 
Our approach involves the following steps:
  1.  
    Detect the cornea (the first edge and the ridge of the first peak when traversing the image from left to right in our system).
  2.  
    Use the shape of the cornea to estimate an initial axis location.
  3.  Iterate (until axis location converges).  
    •  
      a. Detect landmark locations along the current axis (described later).
    •  
      b. Based on the estimated location of the sulcus and lentils, determine the line location of the best fit (a mini-max point—described mathematically by a hyperbolic iparaboloid) to the region of the sulcus.
    •  
      c. Relocate the axis to that location.
  4.  
    Detect landmark locations along the final axis and compute luminances at the landmark locations (the mean of a rectangular region of interest centered on the location of each feature).
Detecting Features
Many of the landmarks are local maxima or minima in the luminances along the trace of the axis. The data, however, were quite noisy, and hence care was taken because the signal-to-noise ratio was low in the regions of the image that contained the nucleus. Several possible solutions for extracting features were explored, including methods to smooth the data and filtering techniques (e.g., wavelet/Fourier methods). Smoothing the raw data using conventional methods (e.g., convolution with a Gaussian mask) was found to be ineffective. Filtering techniques with wavelets and/or Fourier techniques were found to be superior. Evaluating these approaches, we adopted a variable-order Fourier approach—that is, the trace was reconstructed at different orders (or “resolutions”) based on estimated landmark locations. 
The raw data (luminances along the visual axis) were preprocessed: A local median filter was used perpendicular to the axis. Because Gaussian smoothing affects the localization of critical points, the luminances along the visual axis are only smoothed vertically (i.e., we took the median of seven values perpendicular to the axis). The discrete Fourier series for this signal is used to represent the trace. This approach had two advantages. First, the series could be computed with a variable number of coefficients; thus, we could reconstruct the trace with varying degrees of “fidelity to the data” by varying the number of coefficients. Using all coefficients replicates the data precisely. The zeroth coefficient represents the average value of the data (and thus is an extremely crude representation of the trace). The second advantage of using a Fourier series representation is that the signal becomes differentiable. Differentiation of discrete data in images has been known to be problematic (see e.g., Ref. 26 ). With this representation, computation of derivatives (and interpolating function values at locations between data points) is straightforward. 
Figure 1is an example of the digitized image. The trace appears at the top of the figure. It is taken along the axis (1). Lens landmarks needed to assess severity of nuclear sclerosis are labeled (2, 3, 4). 
Developing the Indices.
The severity of nuclear sclerosis was taken to be represented by the relationship of the gray level between selected anatomic features of the trace. While several different combinations of these measurements were possible (and were tried), we present data for four models, all of which were rescaled to 1 to 10:
  1.  
    \(Log\ \left(\ \frac{lentils\ {-}\ sulcus\ {+}\ offset}{sulcus}\right),\)
     
    where the values for the lentils were the height of the peaks labeled 2 and 4 in Figure 1 .
  2.  
    Area under the curve for the anterior lentil endpoints was taken from the trace segment corresponding to the anterior lentil in the digital image 27 where its boundaries were defined by the half-height gray level between the anterior lentil itself and the anterior cortex (fore) and sulcus (aft).
  3.  
    Where the height was the median of the maximum height: \(\left(\ \frac{gray\ level\ at\ the\ height\ of\ anterior\ lentil\ curve\ segment\ {-}\ gray\ level\ at\ the\ height\ of\ sulcus\ {+}\ offset}{gray\ level\ at\ the\ height\ of\ posterior\ lentil\ curve\ segment\ {-}\ gray\ level\ at\ the\ height\ of\ sulcus\ {+}\ offset}\right)\)
  4.  
    Gray levels of the anterior lentil (2 in Fig. 1 ), sulcus (3 in Fig. 1 ), and posterior lentil (4 in Fig. 1 ).
For models 1 and 4, the individual measures of the lentils were combined to a single measure using k medians. 28 The k medians calculation interpolates how far observed values are from “central” values on a predetermined scale, then weights the average of these values to arrive at a single value. The rationales for the four approaches to estimating severity are as follows:
  1.  
    Reflects the degree of optical definition of the lens landmarks. The more sclerotic the lens nucleus, the closer the lentils and sulcus are in value (more homogeneous).
  2.  
    It accounts for the amount of light that is reflected back from the anterior lentil.
  3.  
    It describes a reflectance gradient from the relevant landmarks. As sclerosis progresses, there is a relative enhancement of the light reflected from the anterior compared to the posterior lentil.
  4.  
    No adjustments are made to the gray-level values for the nuclear region; they are combined for an overall estimate of reflectance.
The offset in approaches 1 and 3 refer to the gray-scale values measured in the anterior chamber, to adjust for film emulsion and degree of development of the image. For these four measures, we used an arbitrary continuous scale of 1 to 10. Based on the minimum and maximum values in an initial test sample, the scaling factor was determined and applied to all final gradings. 
Statistical Analyses
Several approaches were taken to evaluate whether the proposed indices are measuring nuclear sclerosis. We expected them to behave similarly to established measures of nuclear sclerosis, such as the Beaver Dam grade (BDG), and so we assessed the relationships between these measures of sclerosis and the original BDG through Spearman correlation coefficients. Distributions of each index by the BDG are displayed using box plots 29 to indicate the mean, median, intraquartile range (IQR = 25th to 75th quartiles) with whiskers to 1.5 · IQR. Similarly, the proposed indices should behave like the BDG in their relationships with age and visual acuity and were evaluated by using Spearman correlations. 
The ability of a measure of sclerosis to detect progression was estimated by the difference in sclerosis between baseline and 5 years and between baseline and 10 years. Definite change in the continuous measures was based on a difference of two standard deviations of the change in a given measure between baseline and the 5-year follow-up for the population. This corresponds to a one-step change in the BDG (5-level scale) and a two-step change for the other indices (10-level scale). 
Results
The measures of central tendency for the BDG and other indices of sclerosis are given in Table 1 . All indices except the BDG are continuous. Therefore, we give the median, the mean and standard deviation, and the coefficient of variation for the new indices, but only the median for the BDG. For index 3, there was an increase over the three time points and the coefficient of variation decreased across the time points for this index. There were smaller numbers of lenses with BDG, because film-based gradings were more often scored as “can’t grade” than semiautomated gradings. The Spearman correlation coefficients of the new indices to the BDG are given in Table 1 . Index 1 correlated most highly with the BDG at each visit. The distribution of values for each index by the BDG at baseline was plotted in Figure 2for index 1. For the sake of space, such plots for the other indices are not given. All showed separation of the median values (notches did not overlap). Generally, index 1 had the largest IQR and index 4 had the smallest when BDG was 2 or 3. The opposite was true when BDG was 4. There were few lenses with BDG level 5 (as can be seen by box width). For every step in the BDG, the median for index 1 changed by approximately 2. The other indices had less consistent change across the range of BDG, which was visually reflected in apparently different slopes. Indices 2 and 3 both had smaller differences in medians between BDG 1 and 2 (flatter slope) than between BDG 3 and 4 (steeper slope). Index 4 had a much larger difference between BDG 3 and 4. Index 4 had a very small difference in the median from BDG of 2 to 3 (a nearly flat slope). This may explain the poorer correlation of this index with the BDG. This characteristic may also be related to the poorer ability of index 4 to identify progression (described later). 
Because nuclear sclerosis is an age-related phenomenon, we evaluated the relationships of the indices to age. The mean ± SD age in this population was 60.6 ± 11.3 years. 25 The median increased with age for all measures (Fig. 3) . The Spearman correlation coefficient between age and the indices is given in Table 2 . Correlation coefficients decreased from baseline to the 10-year follow-up visits for each index. At each visit, the correlations were similar, except for index 3, which had a much lower correlation coefficient. 
Nuclear sclerosis in its early stages has not been found to be a cause of visual impairment. Therefore, visual acuity cannot be used as a validating measure to judge the severity of sclerosis. However, we did expect a direct association between the severity of nuclear sclerosis and vision. The mean ± SD visual acuity in this population was 51.2 ± 15.1 letters of the logMAR (logarithm of the minimum angle of resolution) scale (Snellen equivalent of approximately 20/30) in right eyes. 25 Correlation coefficients of visual acuity and the indices are given in Table 2 . The Spearman correlations between visual acuity and the indices of sclerosis ranged from 0.32 to 0.38. There was no consistent pattern of correlation among the three visits, and no index appeared to have a different relationship to vision. 
We next sought to evaluate the ability of the indices to detect change over three visits, each 5 years apart. We report this as percentage progressed and percentage regressed, as defined in the Methods section. These data, along with the coefficient of variation for change (Table 3) , indicate that for indices 1 and 3, a slightly greater percentage progressed than regressed and for indices 2 and 4, a slightly greater percentage regressed than progressed after 5 years (Table 3) . Then, at 10 years, most indices showed more progression than regression. The BDG showed substantially more progression than regression at the end of each interval, even though the median change for the first 5 years was zero. 
It has been apparent to graders in our study that poorly focused photographs (not focused at the sulcus) yield more uncertainty about the severity of sclerosis. We found that the width of the corneal bright band, a reflection of the plane of focus of the photograph, was related to the estimated severity of sclerosis. A wide corneal width occurs with more posterior focus and was associated with increased estimates of severity. This was true for all indices (data not shown). Attempts to adjust the semiautomated gradings systematically to account for this were not uniformly successful. 
We have previously described a relationship of lens color to age. Lens color was graded dichotomously from the same slit lamp photographs that were used to grade the severity of sclerosis. 30 To evaluate whether this influenced estimates of progression of sclerosis, we computed the percentage of lenses with severity of nuclear sclerosis that regressed or progressed in those lenses in which the color classification was the same at baseline and at follow-up visits. The percentage that regressed and progressed was nearly identical with the values in Table 3for all indices, including the BDG. 
The BDGs are discrete, whereas the other indices are continuous. To evaluate how a more continuous human grade (decimalized system) compares with the indices, a subset of the Beaver Dam Eye Study slit lamp photographs was graded by a single grader from the Age-Related Eye Disease Study (AREDS) grading team using the AREDS grading protocol. 11 The grader was masked to the Beaver Dam gradings. These gradings resulted in a pattern similar to that using the BDGs—namely, the gradings correlated with age and vision. Although we had such gradings at baseline and at the 5-year follow-up only, more progression than regression was found when this human grading system was used, as well. 
Discussion
Grading schemes are most useful when variability is kept at a minimum so that prevalence, incidence, and risk factor estimates can be identified and quantified. However, valid and reproducible techniques to document and assign a severity level to nuclear sclerosis have been problematic. In the Beaver Dam Eye Study, we attempted to diminish obvious sources of film-based image variability by fixing the mechanical settings of the slit lamp camera, adding nonmoving fixation targets, changing flash bulbs periodically, maximizing pupil dilation, and standardizing the training program for photographers. Even so, there were differences in illumination, film emulsion and processing, fixation, focus of the photograph, and pupil size from subject to subject and for the same subject at different times. Still, such photographic documentation of nuclear sclerosis appears to be useful for epidemiologic longitudinal studies. 8 11 13 21 Scheimpflug photography may reduce some of the photographic artifacts because of the uniform plane of focus of the image of the lens nucleus and other lens landmarks. This may permit the detection of more subtle amounts of progression. However, well-trained graders learned to discriminate some of the effects of photographic artifact and variability of focus on lens images and are able to grade the less than perfect photographs of the lens from non-Scheimpflug images reliably, and thus they had the ability to detect age-related changes and progression. Human graders also had the color image, which may have influenced grading, despite our attempts to detect such an effect. 
The current project was undertaken to decrease variability of grading by using an objective semiautomated procedure while accepting the likelihood that such a method would be likely to have weaknesses of its own relating to image quality that decreased the fidelity of the imaging system to represent the underlying disease. Our intent was not to replicate the grading obtained in the film-based grading scheme but to provide an alternative that would also improve on it. The gray-level data were analyzed using several different approaches to determine the sclerosis of the nucleus. Our algorithms took ratios or differences in reflectance of various landmarks in the lens, adjusted for overall photographic illumination by taking into account the gray level of the anterior chamber, and tested the resultant indices for their relationships to age, vision, and change in estimated nuclear sclerosis over a 10-year interval. Although one or another index was more strongly associated with age or vision, no index was substantially better than the human grading (BDG), and no other measure identified as much progression. It is possible that other metrics could be derived from the traces, possibly using other or more of the landmark information than was used in the indices we describe herein, and thereby improve on our results. In addition, it is possible that using the full array of information that may be obtained from digitizing the film images such as including data from three color channels (blue, green, and red) might have been a more successful approach. It is likely that digitizing images that were originally captured on film resulted in more “noise” in the resultant images. Some reasons for this are resolution when reducing an image from film emulsion to one of 100 dpi. Differences in scanning sensitivity at different gray scale levels could also affect the fidelity of the new image. Images that are initially captured digitally may avoid some of these potential problems. We are unable to compare our digitized images with those captured by a digital system, and so we cannot evaluate the comparability. 31 32  
The automated grading was significantly more efficient in utilization of grader time than the standard method. The procedures described herein enabled the senior grader to review all 20,000+ images over approximately 2 months, interspersed with other projects. Balanced against this efficiency must be considered the effort to scan all the slides in question. 
We have found some degree of inconsistency in the two ways we approached age in these analyses (longitudinally by looking at severity at baseline and 5 and 10 years, as seen in Table 1and cross-sectionally as correlation coefficients of severity with age at each of the three visits). We are not certain as to why this is true. It is possible that the computer grading algorithms include those regions in which age-related artifacts are problematic and that graders who are evaluating color photographs are able to discern these and discard them when assigning a grade. 
We made certain arbitrary decisions in choosing the measurements to be included in our indices of sclerosis. These included assuming that the clarity of the sulcus at the center of the nucleus and more homogeneous light-scattering are typical of more severe nuclear sclerosis. However, lenses with apparently homogeneous nuclear areas can be seen at young ages when nuclear density is low. Also, homogeneous density with no obvious sulcus can occasionally be seen at older ages when there is more nuclear sclerosis. In such cases, using formulas that depend on finding a sulcus could produce inaccurate assessment. 
We computed several indices of sclerosis based on gray levels from the optical trace. We present data for only the most promising four approaches. Even these four were not entirely satisfactory. Two of the derived indices relied in part on the gray scale of the offset—the value for the anterior chamber after the obvious bright artifacts from a reflection of the slip lamp mirror and reflections from lashes were eliminated. The offset values were not constant at all levels of severity of nuclear sclerosis, as measured from the standard photographs and in the actual study images. Attempts to model this variability were not successful and thus were not included in our computation of index values (for indices 1 and 3). Another assumption we made was that nuclear sclerosis increases with age. Thus, although progression of nuclear sclerosis with age is an observable clinical phenomenon, there is no way to document and quantitate this biochemically in vivo, and observable progression at the slit lamp requires years of observation. We extended this thesis to expect that the age–sclerosis relationship is progressive and therefore that regression or improvement does not occur, so we regarded the latter change as error. Moreover, we also assume that an index that detects a greater amount of change is a better measure of the underlying biology. It is possible that further attempts to identify and correct variability in how the lens is imaged and how to identify and eliminate artifacts may result in a more valid result. However, another approach to evaluating digitized images such as ours is the further development of a neural network similar to that described by Duncan et al. 15 The underlying assumption for that approach is that the human grader is correct and that a semiautomated system can be “trained” to yield similar grades. Our assumption had been that an automated uniform approach might improve on human grading, because it would reduce human grader variability. In fact, our findings indicate that the experience and judgment of human graders is desirable. Thus, it is possible that further development of the neural net approach to grading digital images or digitized images from film-based photographs will deliver a method that replicates human grading, but may do so more rapidly than humans. Fan et al. 33 have also developed a model that assumes that human graders are correct in grading the severity of sclerosis. This has yielded promising results in a small validation sample (data not shown). It is also possible that there is more useful information in color photographs on which the original Beaver Dam gradings were made than in the digitized gray-scale analogues. Our conclusion is that, at this time, we plan to continue careful training of human graders for our film-based images of the lens nucleus. 
In summary, we have developed a system for digitizing images of the lens nucleus from slit lamp images and for identifying important landmarks in the image in a semiautomated fashion. The procedure is efficient in its use of grader time and is reproducible. Although we have not achieved our goal to develop a metric that is as satisfactory as human grading of slit lamp photographs, we plan to continue to refine our algorithms to minimize further the effects of artifacts and other sources of variability. 
 
Figure 1.
 
Example of a scanned image used to grade nuclear sclerosis with axis and landmarks.
Figure 1.
 
Example of a scanned image used to grade nuclear sclerosis with axis and landmarks.
Table 1.
 
Measures of Central Tendency for Indices of Nuclear Sclerosis at Baseline and 5- and 10-Year Follow-up Examinations (Right Eyes)
Table 1.
 
Measures of Central Tendency for Indices of Nuclear Sclerosis at Baseline and 5- and 10-Year Follow-up Examinations (Right Eyes)
Visit Index n Median Mean Standard Deviation Coefficient of Variation Correlation* with BDG, †
Baseline BDG 4518 2.00
1 4566 4.11 4.46 2.34 52.31 0.68
2 4566 3.82 4.01 2.06 51.54 0.61
3 4566 4.61 4.70 1.93 41.06 0.60
4 4566 3.57 3.88 2.02 52.17 0.55
5-year follow-up BDG 3166 3.00
1 3207 4.02 4.37 2.23 50.99 0.63
2 3207 3.59 3.87 1.97 50.81 0.57
3 3207 4.87 4.87 1.95 39.98 0.62
4 3207 3.50 3.71 1.90 51.26 0.49
10-year follow-up BDG 2261 3.00
1 2275 3.75 4.02 2.20 54.80 0.62
2 2275 3.10 3.52 1.96 55.72 0.55
3 2275 5.16 5.15 1.85 35.91 0.62
4 2275 3.22 3.33 1.88 56.40 0.46
Figure 2.
 
Box plot showing intraquartile range (IQR = 25th to 75th percentiles) with whiskers to ± 1.5 · IQR for index 1 grouped by the human grade (BDG) at baseline in right eyes only. The box is notched from median ± 1.58 (IQR/ \batchmode \documentclass[fleqn,10pt,legalpaper]{article} \usepackage{amssymb} \usepackage{amsfonts} \usepackage{amsmath} \pagestyle{empty} \begin{document} \(\sqrt{n}\) \end{document} ), and box width is scaled proportional to the number of eyes in the group. +, the mean. Lines connect the medians.
Figure 2.
 
Box plot showing intraquartile range (IQR = 25th to 75th percentiles) with whiskers to ± 1.5 · IQR for index 1 grouped by the human grade (BDG) at baseline in right eyes only. The box is notched from median ± 1.58 (IQR/ \batchmode \documentclass[fleqn,10pt,legalpaper]{article} \usepackage{amssymb} \usepackage{amsfonts} \usepackage{amsmath} \pagestyle{empty} \begin{document} \(\sqrt{n}\) \end{document} ), and box width is scaled proportional to the number of eyes in the group. +, the mean. Lines connect the medians.
Figure 3.
 
Medians of BDGs (B) and each proposed index (symbol 1 through 4) by age. To avoid overlapping lines, a constant of 2, 4, and 6 was added to indices 2, 3, and 4, respectively, for display purposes only.
Figure 3.
 
Medians of BDGs (B) and each proposed index (symbol 1 through 4) by age. To avoid overlapping lines, a constant of 2, 4, and 6 was added to indices 2, 3, and 4, respectively, for display purposes only.
Table 2.
 
Spearman Correlations of Indices of Sclerosis with Age and Visual Acuity (Right Eyes)
Table 2.
 
Spearman Correlations of Indices of Sclerosis with Age and Visual Acuity (Right Eyes)
Visit Index n Correlation with Age Correlation with Visual Acuity
Baseline BDG 4518 0.65 0.38
1 4566 0.71 0.38
2 4566 0.68 0.37
3 4566 0.46 0.32
4 4566 0.64 0.34
5-year follow-up BDG 3166 0.60 0.32
1 3207 0.69 0.35
2 3207 0.68 0.34
3 3207 0.45 0.30
4 3207 0.63 0.31
10-year follow-up BDG 2261 0.55 0.36
1 2275 0.60 0.33
2 2275 0.58 0.32
3 2275 0.44 0.34
4 2275 0.52 0.28
Table 3.
 
Regression and Progression of Nuclear Sclerosis for 5- and 10-Year Intervals (Right Eyes)
Table 3.
 
Regression and Progression of Nuclear Sclerosis for 5- and 10-Year Intervals (Right Eyes)
Interval Index n Median Mean Standard Deviation Coefficient Variation % Regressed % Progressed
5 Years BDG 3139 0.00 4.94 48.20
1 3160 0.22 0.22 1.36 618.73 7.03 8.80
2 3160 0.23 0.14 1.25 874.18 6.01 4.94
3 3160 0.42 0.41 1.39 336.17 4.46 9.87
4 3160 0.05 0.09 1.33 1419.64 6.23 5.73
10 Years BDG 2252 1.00 3.15 61.90
1 2255 0.24 0.30 1.59 534.61 9.71 14.32
2 2255 0.16 0.14 1.47 1029.78 8.34 8.96
3 2255 0.94 0.98 1.47 149.58 2.62 19.69
4 2255 0.01 0.02 1.51 7549.20 8.82 8.03
The authors thank Lisa Grady and Janet Klosterman for technical support. 
KleinBE, KleinR, MossSE. Incident cataract surgery: The Beaver Dam Eye Study. Ophthalmology. 1997;104:573–580. [CrossRef] [PubMed]
ChylackLT, Jr, LeskeMC, McCarthyD, KhuP, KashiwagiT, SperdutoR. Lens Opacities Classification System II (LOCS II). Arch Ophthalmol. 1989;107:991–997. [CrossRef] [PubMed]
ChylackLT, Jr, WolfeJK, SingerDM, et al. The Lens Opacities Classification System III. The Longitudinal Study of Cataract Study Group. Arch Ophthalmol. 1993;111:831–836. [CrossRef] [PubMed]
ThyleforsB, ChylackLT, Jr, KonyamaK, et al. A simplified cataract grading system. The WHO Cataract Grading Group. Ophthalmic Epidemiol. 2002;9:83–95. [CrossRef] [PubMed]
TaylorHR, WestSK. The clinical grading of lens opacities. Aust NZ J Ophthalmol. 1989;17:81–86. [CrossRef]
SparrowJM, BronAJ, BrownNA, AyliffeW, HillAR. The Oxford Clinical Cataract Classification and Grading System. Int Ophthalmol. 1986;9:207–225. [CrossRef] [PubMed]
BrownN. An advanced slit-image camera. Br J Ophthalmol. 1972;56:624–631. [CrossRef] [PubMed]
WestSK, RosenthalF, NewlandHS, TaylorHR. Use of photographic techniques to grade nuclear cataracts. Invest Ophthalmol Vis Sci. 1988;29:73–77. [PubMed]
HockwinO, DragomirescuV, LaserH, et al. Scheimpflug photography of the anterior eye segment: principle, instrumentation and application to clinical and experimental ophthalmology. J Ophthalmic Photogr. 1986;9:104–111.
KleinBE, KleinR, LintonKL, MagliYL, NeiderMW. Assessment of cataracts from photographs in the Beaver Dam Eye Study. Ophthalmology. 1990;97:1428–1433. [CrossRef] [PubMed]
Age-Related Eye Disease Study Research Group. The Age-Related Eye Disease Study (AREDS) system for classifying cataracts from photographs: AREDS Report No. 4. Am J Ophthalmol. 2001;131:167–175. [CrossRef] [PubMed]
SasakiK, FujisawaK, SakamotoY. Quantitative evaluation of nuclear cataract using image analysis. Ophthalmic Res. 1992;24:26–31. [CrossRef] [PubMed]
WestSK, MuñozB, WangF, TaylorH. Measuring progression of lens opacities for longitudinal studies. Curr Eye Res. 1993;12:123–132. [CrossRef] [PubMed]
QianW, SoderbergPG, ChenE, MagniusK, PhilipsonB. A common lens nuclear area in Scheimpflug photographs. Eye. 1993;7:799–804. [CrossRef] [PubMed]
DuncanDD, ShuklaOB, WestSK, ScheinOD. New objective classification system for nuclear opacification. J Opt Soc Am A. 1997;14:1197–1204. [CrossRef]
CotlierE. Cataract Image Analysis System. Eye. 1999;13:457–463. [CrossRef] [PubMed]
ChylackLT, Jr, McCarthyD, KhuP. Use of Topan SL-45 Scheimpflug slit photography to measure longitudinal growth of nuclear cataracts in vivo. Lens Res. 1988;5:83–93.
GarrettSK, RobmanLD, McCartyCA, ThomasAP, McNeilJJ, TaylorHR. Reproducibility of automatic standard digital analysis of lens opacities. Aust NZ J Ophthalmol. 1998;26(suppl 1)S29–S31. [CrossRef]
HallNF, LempertP, ShierRP, ZakirR, PhillipsD. Grading nuclear cataract: reproducibility and validity of a new method. Br J Ophthalmol. 1999;83:1159–1163. [CrossRef] [PubMed]
MagnoBV, FreidlinV, DatilesMB, III. Reproducibility of the NEI Scheimpflug Cataract Imaging System. Invest Ophthalmol Vis Sci. 1994;35:3078–3084. [PubMed]
FooKP, MacleanH. Measured changes in cataract over six months: sensitivity of the Nidek EAS-1000. Ophthalmic Res. 1996;28(suppl 2)32–36.
QianW, SoderbergP, LindstromB, ChenE, MagniusK, PhilipsonB. Spatial distribution of back scattering in the nuclear area of the non-cataractous human lens. Eye. 1994;8:524–529. [CrossRef] [PubMed]
RobmanLD, McCartyCA, GarrettSK, et al. Comparison of clinical and digital assessment of nuclear optical density. Ophthalmic Res. 1999;31:119–126. [CrossRef] [PubMed]
QianW, SoderbergPG, ChenE, MagniusK, PhilipsonB. 3 year simvastatin treatment and lens nuclear back scattering. Br J Ophthalmol. 2000;84:512–516. [CrossRef] [PubMed]
KleinR, KleinBE, LintonKL, De MetsDL. The Beaver Dam Eye Study: visual acuity. Ophthalmology. 1991;98:1310–1315. [CrossRef] [PubMed]
CastlemanK. Digital Image Processing. 1996;Prentice-Hall Englewood Cliffs, NJ.
RobmanLD. Measurement of change in lens opacities over 2 years with the Nidek EAS-1000 image analysis system (dissertation). 2000;University of Melbourne Melbourne, Australia.
BradleyPS, MangasarianOL, StreetWN. Clustering via concave minimization.MozerM JordanMI PetsceT eds. Advance in Neural Information Processing Systems 9. 1997;368–374.MIT Press Cambridge, MA.
McGillR, TukeyJW, LarsenWA. Variations of box plots. Am Stat. 1978;32:12–16.
KleinBE, KleinR, LintonKL. Prevalence of age-related lens opacities in a population: The Beaver Dam Eye Study. Ophthalmology. 1992;99:546–552. [CrossRef] [PubMed]
FultonW. A Few Scanning Tips. 1998; 2nd ed; Available at http://www.scantips.com. Accessed November 15, 2004.
Eastman Kodak Company. CCD Primer, Conversion of Light (Photons) to Electronic Charge. August 5, 1999;Available at http://www.kodak.com/go/ccd. Accessed November 15, 2004
FanS, DyerCR, HubbardL, KleinB. An automatic system for classification of nuclear sclerosis from slit lamp photographs.Lecture Notes in Computer Science, Volume 2878: Medical Image Computing and Computer-Assisted Intervention. 2003;592–601.Springer-Verlag Heidelberg.
Figure 1.
 
Example of a scanned image used to grade nuclear sclerosis with axis and landmarks.
Figure 1.
 
Example of a scanned image used to grade nuclear sclerosis with axis and landmarks.
Figure 2.
 
Box plot showing intraquartile range (IQR = 25th to 75th percentiles) with whiskers to ± 1.5 · IQR for index 1 grouped by the human grade (BDG) at baseline in right eyes only. The box is notched from median ± 1.58 (IQR/ \batchmode \documentclass[fleqn,10pt,legalpaper]{article} \usepackage{amssymb} \usepackage{amsfonts} \usepackage{amsmath} \pagestyle{empty} \begin{document} \(\sqrt{n}\) \end{document} ), and box width is scaled proportional to the number of eyes in the group. +, the mean. Lines connect the medians.
Figure 2.
 
Box plot showing intraquartile range (IQR = 25th to 75th percentiles) with whiskers to ± 1.5 · IQR for index 1 grouped by the human grade (BDG) at baseline in right eyes only. The box is notched from median ± 1.58 (IQR/ \batchmode \documentclass[fleqn,10pt,legalpaper]{article} \usepackage{amssymb} \usepackage{amsfonts} \usepackage{amsmath} \pagestyle{empty} \begin{document} \(\sqrt{n}\) \end{document} ), and box width is scaled proportional to the number of eyes in the group. +, the mean. Lines connect the medians.
Figure 3.
 
Medians of BDGs (B) and each proposed index (symbol 1 through 4) by age. To avoid overlapping lines, a constant of 2, 4, and 6 was added to indices 2, 3, and 4, respectively, for display purposes only.
Figure 3.
 
Medians of BDGs (B) and each proposed index (symbol 1 through 4) by age. To avoid overlapping lines, a constant of 2, 4, and 6 was added to indices 2, 3, and 4, respectively, for display purposes only.
Table 1.
 
Measures of Central Tendency for Indices of Nuclear Sclerosis at Baseline and 5- and 10-Year Follow-up Examinations (Right Eyes)
Table 1.
 
Measures of Central Tendency for Indices of Nuclear Sclerosis at Baseline and 5- and 10-Year Follow-up Examinations (Right Eyes)
Visit Index n Median Mean Standard Deviation Coefficient of Variation Correlation* with BDG, †
Baseline BDG 4518 2.00
1 4566 4.11 4.46 2.34 52.31 0.68
2 4566 3.82 4.01 2.06 51.54 0.61
3 4566 4.61 4.70 1.93 41.06 0.60
4 4566 3.57 3.88 2.02 52.17 0.55
5-year follow-up BDG 3166 3.00
1 3207 4.02 4.37 2.23 50.99 0.63
2 3207 3.59 3.87 1.97 50.81 0.57
3 3207 4.87 4.87 1.95 39.98 0.62
4 3207 3.50 3.71 1.90 51.26 0.49
10-year follow-up BDG 2261 3.00
1 2275 3.75 4.02 2.20 54.80 0.62
2 2275 3.10 3.52 1.96 55.72 0.55
3 2275 5.16 5.15 1.85 35.91 0.62
4 2275 3.22 3.33 1.88 56.40 0.46
Table 2.
 
Spearman Correlations of Indices of Sclerosis with Age and Visual Acuity (Right Eyes)
Table 2.
 
Spearman Correlations of Indices of Sclerosis with Age and Visual Acuity (Right Eyes)
Visit Index n Correlation with Age Correlation with Visual Acuity
Baseline BDG 4518 0.65 0.38
1 4566 0.71 0.38
2 4566 0.68 0.37
3 4566 0.46 0.32
4 4566 0.64 0.34
5-year follow-up BDG 3166 0.60 0.32
1 3207 0.69 0.35
2 3207 0.68 0.34
3 3207 0.45 0.30
4 3207 0.63 0.31
10-year follow-up BDG 2261 0.55 0.36
1 2275 0.60 0.33
2 2275 0.58 0.32
3 2275 0.44 0.34
4 2275 0.52 0.28
Table 3.
 
Regression and Progression of Nuclear Sclerosis for 5- and 10-Year Intervals (Right Eyes)
Table 3.
 
Regression and Progression of Nuclear Sclerosis for 5- and 10-Year Intervals (Right Eyes)
Interval Index n Median Mean Standard Deviation Coefficient Variation % Regressed % Progressed
5 Years BDG 3139 0.00 4.94 48.20
1 3160 0.22 0.22 1.36 618.73 7.03 8.80
2 3160 0.23 0.14 1.25 874.18 6.01 4.94
3 3160 0.42 0.41 1.39 336.17 4.46 9.87
4 3160 0.05 0.09 1.33 1419.64 6.23 5.73
10 Years BDG 2252 1.00 3.15 61.90
1 2255 0.24 0.30 1.59 534.61 9.71 14.32
2 2255 0.16 0.14 1.47 1029.78 8.34 8.96
3 2255 0.94 0.98 1.47 149.58 2.62 19.69
4 2255 0.01 0.02 1.51 7549.20 8.82 8.03
×
×

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.

×