Automated Analysis of Optic Nerve Images for Detection and Staging of Papilledema

Abstract

Purpose.: To develop an automated system that analyzes digital fundus images for staging and monitoring of optic disc edema (i.e., papilledema), due to raised intracranial pressure.

Methods.: A total of 294 retrospective, digital photographs of the right and left eyes of 39 subjects with papilledema acquired over the span of 2 years were used. Software tools were developed to analyze three features of papilledema from digital fundus photographs: (1) sharpness of the optic disc border, (2) discontinuity along major vessels overlying the optic nerve, and (3) texture properties of the peripapillary retinal nerve fiber layer (RNFL). A classifier used these features to assign a grade of papilledema according to a standard protocol used by an expert neuro-ophthalmologist (RK).

Results.: The algorithm showed substantial agreement (κ = 0.71, P < 0.001) with the neuro-ophthalmologist when grading papilledema per patient. Vessel features showed statistical significance (P < 0.05) in differentiating grades 0, 1, and 2 from grades 3 and 4, whereas disc obscuration differentiated grades 0 or 1 from the rest (P < 0.05).

Conclusions.: These results show that this algorithm can be used to automatically grade papilledema. The algorithm provides objective and quantitative assessment of the stage of papilledema with accuracy that is comparable to grading by a neuro-ophthalmologist. One application is in rapid assessment of digital optic nerve photographs acquired in clinical, intensive care, and emergency response settings by nonophthalmologists to evaluate for the presence and severity of papilledema, due to intracranial hypertension.

Papilledema ^1,2 is a swelling of the optic disc secondary to elevated intracranial pressure. ³ Pathophysiologically, the disc swelling in papilledema is the result of reduced axoplasmic flow within the bundle of nerves in the optic disc. Because regions of the brain filled with cerebrospinal fluid are continuous, with the same fluid enclosed within the optic nerve sheath, an increase in the cerebrospinal fluid pressure will be transmitted to the optic nerve head, causing stasis of axoplasmic flow in the optic nerve as it exits the eye. ⁴ The buildup of intracellular and extracellular fluid at the lamina cribrosa results in swelling of the nerve head. ⁵  

Patients with papilledema require urgent attention, not only to minimize vision loss, but also to avoid elevated intracranial pressure, which can be fatal. Causes of elevated intracranial pressure include idiopathic intracranial hypertension, traumatic brain injury, intracranial tumors, subarachnoid hemorrhage, infection, inflammation, and venous sinus thrombosis. ⁶ Prevalence of idiopathic intracranial hypertension in the general population is 1 case per 100,000 persons, increasing to 13 cases per 100,000 among individuals ages 20 to 44 who are 10% above ideal body weight. ⁷ Traumatic brain injury can occur as a result of sports injuries, motor vehicle accidents, and warfare. ⁸ Timely and accurate detection and assessment of papilledema is essential to effectively triage patients for adequate treatment. 

Tracking progression of papilledema is a difficult task for neurosurgeons, neurologists, ophthalmologists, and emergency treatment physicians. ⁹ To address this problem, a widely accepted staging scheme for optic disc swelling using ophthalmoscopic features, was proposed by Frisén. ¹⁰ In his scheme, Frisén divides optic disc swelling into six grades from 0 (normal) to 5 (severe). Each grade is characterized by a set of visual features observed on the optic disc and peripapillary retina. Scott et al. ¹¹ modified the Frisén staging scale by identifying a key observable feature for each grade and more precisely redefining some of Frisén's original features, allowing more objective classification of papilledema grade in the clinic. A summary of the features defined in the Modified Frisén Scale (MFS) is shown in Table 1. 

Table 1.

 

The Modified Frisén Scale ¹¹

Grade	Description
Grade 0: Normal
Grade 1: Minimal	C-shaped ring that is subtle and grayish with a temporal gap; obscures underlying retinal details
	Temporal disc margin normal, nasal disc margin obscured
Grade 2: Low degree	Circumferential ring
	Elevation, nasal border
	Temporal disc margin is obscured in addition to nasal margin
	No major vessel obscuration
Grade 3: Moderate	Obscurations of one or more segments of major blood vessels as they pass over the optic disc, but not obscured at their origin in the center
	Elevation, all borders
	Ring, irregular outer fringe with finger-like extensions
Grade 4: Marked	Total obscuration of a segment of a major artery or vein at its central origin
	Ring, complete
Grade 5: Severe	Obscuration of all vessels both on the disc and leaving the disc

 Reprinted with permission from Scott CJ, Kardon RH, Lee AG, Frisén L. Wall M. Diagnosis and grading of papilledema in patients with raised intracranial pressure using optical coherence tomography vs clinical expert assessment using a clinical staging scale. Arch Ophthalmol. 2010;128(6):705–711. Copyright © American Medical Association.

Although there are many advantages to using fundus images with the Frisén scale to document the state of the optic disc and the retina in general, it requires expert interpretation, ¹¹ which is prone to inter- and intragrader variability, as is the case in other eye diseases. ^{12 –14} Computer-aided image analyses techniques provide a means of overcoming the variability inherent in human assessment of papilledema. To rapidly assess for the presence the papilledema and its progression, computer image analysis can be applied to help solve the variability problem. A computer-aided tool would be useful in several clinical settings including emergency rooms where expert assessment of the optic disc's appearance may not be available in a timely fashion. With the advent of fundus imaging devices that are becoming more widely available in nonophthalmology settings, digital fundus photographs are now more accessible, establishing a need for a computer-aided tool to detect and stage papilledema in a variety of settings. 

The purpose of this study was to extend the work of Scott et al. ¹¹ by demonstrating a system that automatically characterizes papilledema by using mathematically derived, visually based features in fundus images and that uses these mathematical features to classify papilledema grade from digital images according to the MFS. 

Methods

Data

For this study, 294 retrospective, longitudinal, digital photographs of the right and left optic disc from 39 subjects were used that were taken at the Department of Ophthalmology and Visual Sciences (University of Iowa, Iowa City, IA). The study was approved by the University of Iowa's Institutional Review Board and adhered to the tenets of the Declaration of Helsinki. Patients were identified in a random manner among those patients who were seen in the neuro-ophthalmology clinic who had had papilledema diagnosed by lumbar puncture. The diagnosis of papilledema was given as the underlying cause for raised intracranial pressure and the cause of the optic disc edema. Patients who had stereo digital disc photos, visual field testing, and spectral-domain OCT (Cirrus; Carl Zeiss Meditec, Dublin, CA) were included. The images were acquired during 96 visits, with each subject having between one and four visits over the span of 2 years (2008–2009; RC 50-DX retinal cameras [Topcon, Tokyo, Japan] with a Megavision 6 megapixel back [2392 × 2048 actual image size], which is a full-size [35 mm] CCD sensor [OIS, Sacramento, CA]). Stereo pair images were acquired with a 30° or a 55° view. Only one image of the stereo pair was used for image analysis in this study. The images were graded by a neuro-ophthalmologist (RK) according to the MFS. ¹¹ The ground truth expert grading of papilledema consisted of grade 0 (n = 10), grade 1 (n = 48), grade 2 (n = 27), grade 3 (n = 5), and grade 4 (n = 6) images. There were no grade 5 examples in our data set. For this reason, our study did not attempt to classify grade 5. An example of each grade present in our study is shown in Figure 1. 

Figure 1.

 

Samples of the image dataset used to grade papilledema. (a) MFS grade 0, (b) MFS grade 1, (c) MFS grade 2, (d), MFS grade 3, and (e) MFS grade 4.

The center and margin of the optic disc were marked by a certified ophthalmic medical technician for the sole purpose of cropping and resizing the digital photographs taken at both the 30° and the 55° fields of view. The images were cropped, 3 disc diameters per side (900 × 900 pixels, TIFF, lossless compression), centered on the optic disc, and converted to right-eye orientation. The disc margin was approximated by an experienced ophthalmic technician who manually drew the margin on the digital photograph. The approximate disc diameter was defined as the average number of pixels from the margin of the optic disc to its geometric center. This process was performed only for the purpose of standardizing the image size and region of interest analyzed by the software algorithms, so that all digital images being analyzed had constant magnification. The optic disc margin and resulting disc diameter as defined by the technician, were used only as a means of prestudy sorting of the images by magnification and did not otherwise factor into the study, since the precise location of the disc margin is often ambiguous in eyes with severe papilledema. 

Image Features

Assessment of papilledema focuses on visual features on the optic disc and the peripapillary region to grade optic disc edema. The features can be divided into five groups (1) vasculature features, (2) optic disc margin changes, (3) surrounding gray ring, (4) retinal nerve fiber layer (RNFL) changes, and (5) depth information. The stereo depth information was not used in this study, nor is it taken into account in the MFS; therefore, only the first five grades of papilledema (grade 0–4) were analyzed for features. Vasculature features include any obscuration of the parent central retinal artery or vein and their major branches emanating from the center of the optic disc and coursing over its borders. Optic disc margin features include the asymmetric obscuration between the temporal and nasal poles of the optic disc present in grade 1. The gray ring refers to the annular obscuration at the disc margin that starts nasally and progresses temporally in a circumferential C-shaped pattern. The RNFL visual features refer to changes in the texture of the superficial nerve fibers in the peripapillary region. In the following sections, it will be shown how each visual feature was characterized by the algorithm. 

Vasculature-Related Features.

The obscuration of the major blood vessels on the optic disc surface by the surrounding swollen nerve fibers and extracellular fluid is related to the degree of optic disc swelling. As the disease progresses, the blood vessels become less well defined due to the thickening of the nerve fiber layer and extracellular fluid. This visual feature was used to introduce a vascular obscuration model derived from a computer-based vessel segmentation algorithm. This is a previously unreported feature and is defined in this article as the vessel discontinuity index (VDI). Although the VDI is directly derived from the vessel segmentation algorithm, its performance is based on the segmentation of large vessels. Small vessels, which are the source of most errors in other vessel segmentation applications, are not important in our application. The Frisén scale is likewise based on the envelopment of the large vessels, which is consistent with the VDI analysis. An additional characteristic is that the major blood vessels coursing over the optic disc become affected (grade 3) before their distal extensions at the disc margin are obscured. Thus, the location of the obscuration along the length of the artery or vein is an indicator of grade of papilledema, with more advanced grades representing vessel obscurations closer to the center of the optic disc. This characteristic is represented by disc proximity feature in Table 2. 

Table 2.

 

The 24 Features Used in the Algorithm, by Category

Blood Vessel Features   Vessel discontinuity index   Vessel discontinuity Index to disc proximity   Area of the largest region in the vessel map   Area of the smallest region in the vessel map   Mean area of regions in the vessel map   Standard deviation of the distribution of areas on the regions in the vessel map   Kurtosis of the distribution of areas on the regions in the vessel map

Peripapillary Texture   Energy of the peripapillary region   Entropy of the peripapillary region   Contrast of the peripapillary region   Cluster shade of the peripapillary region   Correlation of the peripapillary region

Disc Margin Obscuration   Obscuration around the disc margin   Disc margin obscuration in the nasal pole   Disc margin obscuration in the temporal pole   Superior segment of nasal to temporal ratio of disc margin obscuration   Central segment of nasal to temporal ratio of disc margin obscuration   Inferior segment of nasal to temporal ratio of disc margin obscuration   Nasal to superior segment of temporal ratio of disc margin obscuration   Nasal to central segment of temporal ratio of disc margin obscuration   Nasal to inferior segment of temporal ratio of disc margin obscuration   Maximum nasal segment to temporal segment ratio of disc margin obscuration   Minimum nasal to temporal segment ratio of disc margin obscuration   Nasal to temporal ration of disc margin obscuration

In papilledema, as the larger vessels become obscured, the visual continuity along the length of the vessel decreases. A vessel will then begin to appear as discontinuous segments. The approach used in this study for detecting and quantifying vessel obscuration was to find a means for quantifying the connectivity along the length of arteries and veins on the disc and peripapillary regions. The process starts with the segmentation of vessels, which uses image processing to extract the vessel skeletons from the digital photos of the disc and peripapillary retina. To segment the vessels, an algorithm was developed based on techniques presented by Frangi et al. ¹⁵ for vessel enhancement and Chanwimaluang ¹⁶ for automatic segmentation. Because of the reduced field of view surrounding only the optic disc, uneven illumination correction was found to be unnecessary. The results of the vessel segmentation used in this study obtained an ROC of 0.92 when applied to one of the commonly used standard databases, DRIVE. ¹⁷ The algorithm is tuned to detect larger vessels; therefore, the contribution of the small vessels to the VDI is small. Using the larger vessels to calculate VDI is consistent with the MFS, which also considers the large vessels to differentiate between the middle stages and late stages of papilledema. Examples of fundus images and their corresponding vessel segmentations are shown in the top and bottom row, respectively, of Figure 2. In the advanced case, we can observe how the vessel segmentation algorithm includes hemorrhages in its segmentation. This serves to increase the VDI and supports the correct classification of the disease stage. 

Figure 2.

 

Original images (top). Corresponding associated vessel structure after segmentation (bottom). Left: grade 0; middle: grade 2; right: grade 4.

Using the vessel mask, the VDI was calculated by counting the number of disjointed regions in the mask within optic disc and peripapillary area, 1.5 disc diameters from the center of the disc. A vessel region is defined as pixels that are contiguous. As vessels become more obscured, more gaps appear in the vessel mask; suggesting that the more advanced the swelling, the higher the VDI. Large and small vessels are not differentiated and add equally to the VDI. The density or number of pixels representing a segmented vessel is not factored into the VDI, but could be a worthwhile refinement to explore. 

In addition to the VDI, which made use only of disconnected regions, total pixel features, maximum, mean, minimum, standard deviation, and skewness of the distribution of the number of pixels in each region of the vessel mask were also calculated and used as features for papilledema staging. 

Optic Disc Margin.

As the swelling due to the papilledema becomes more severe, the margin of the optic disc becomes less defined. According to the MFS (Table 1), the nasal pole is the first to become less defined and is associated with the mildest grade of papilledema (grade 1). At the next higher grade, grade 2, obscuration of the margin progresses circumferentially to involve the temporal pole. Higher grades (i.e., grades 3 and 4), show equally obscured margins all around the disc, although the extent of the obscuration of the disc margin increases with increasing edema. A quantitative, automated method was implemented to measure the “blurriness” for sectors of the optic disc margin. 

To avoid the interference from the vessels that cross the optic disc margins, our system implements the inpainting algorithm developed by Criminisi et al. ¹⁸ This inpainting algorithm eliminates the vessels around the optic disc while preserving the integrity of the surrounding structures. Points that fall in the inpainted vessel regions carry information of the pixels in the neighborhood and can be thought of as a means of interpolation. Since the variance is calculated for a large number of points on the boundary (n = 180) and vessels represent less than 5% of the points, any interpolation errors introduced by the inpainting did not greatly affect the blurriness measure. An example comparing an optic disc with its inpainted version is shown in Figure 3. 

Figure 3.

 

Results of inpainting. Optic disc image (left). Automated vessel segmentation (center). Optic disc with vessels close to the optic disc removed (right).

The optic disc margin was found automatically by selecting the pixels every 2° with the largest radial intensity gradient that defines the disc margin at that point (i.e., change in pixel contrast along a radial line moving from brighter disc area to darker retina). As the swelling progresses and disc margin obscuration increases, the locations of the points with the largest radial intensity gradient become less correlated with the natural, elliptical shape of the optic disc. In other words, the less defined the optic disc margin becomes due to the edema, the greater the variation in the distance from the center of the optic disc to the points located at each circumferential point along the obscured disc margin. This distance variation, measured as the statistical variance between a best-fit circle and the actual points, is measured along each radial meridian from the center of the disc to the location where the change in pixel intensity is greatest (location of maximum derivative of the pixel intensity along a radial line originating from the center of the disc). Optic disc boundaries determined by this method for grade 0 and 1 optic discs from the same patient, superimposed on their original vascular images, are shown in Figure 4. The automatically selected points were divided into four arc segments of 90° each: superior, inferior, temporal, and nasal. The superior and inferior poles are arc sectors that are not used in the Frisén scale system due to the large variability in the appearance of the disc margin in these sectors of the normal population. Similarly, superior and inferior poles are not used in our current algorithm, in keeping with the modified Frisén grading of papilledema. The nasal and temporal poles are further subdivided into three segments of equal 30° arc angles to refine the model of obscuration progression. A graph illustrating this division is shown in Figure 5. 

Figure 4.

 

Red points: the margin of the disc as defined by the distance from the center of the disc at with the maximum change in radial pixel intensity gradient. The point location found by the algorithm is shown with respect to the optic disc margin in a grade 0 (left) and a grade 1 (right) image of the right eye (temporal border is to the left of the disc and nasal border is to the right side). This example illustrates how there is more variation in the location of the points with respect to the disc margin in the grade 1 image at the nasal pole compared to the same disc when no papilledema was present.

Figure 5.

 

Graph illustrating the subdivisions of the poles of the optic disc. The optic disc is divided into four poles (superior, inferior, temporal, and nasal). The temporal and nasal poles are subdivided into three segments each (superior, central, and inferior).

The SD of the radial distance from the center of the optic disc to each point on its margin where the change in pixel intensity is greatest was calculated on the nasal and temporal poles at each circumferential location every 2° (Fig. 5). This variation defined the disc margin obscuration score of an image. The ratio between their total and subsegment variation was used as a feature in the algorithm to differentiate early grades of papilledema. Bar plots of papilledema grade versus obscuration per pole for all our dataset is shown in Figure 6. In the plots, it can be observed how the nasal pole disc margin varies more than the temporal pole as the disease progresses due to obscuration at the margin. The maximum and minimum deviation of each of the three segments within the temporal and nasal poles was also used to characterize each grade. 

Figure 6.

 

Amount of obscuration in the central segments of the nasal and temporal poles. According to the MFS, for an image to be considered grade 1, the nasal pole should appear obscured while the temporal pole should not. For grade 2, both poles should present lack of definition across the disc margin. This graph agrees with the Frisén definitions in that the difference in obscuration between the central nasal segment presents a greater obscuration value starting in grade 1 (points A), while the central temporal segment remains relatively low until grade 2 (point B), significant difference in the standard deviation at P < 0.05.

Peripapillary Texture.

Changes in the nerve fiber layer appearance due to papilledema have been documented by Hoyt et al. ^{1,9 –11} and others. As the severity of papilledema advances, the nerve fiber layer loses its definition, becoming dark and featureless. A gray annular ring starts forming around the optic disc, first C-shaped (grade 1), then completely circumferential (grade 2), developing fringe extensions (grade 3) and with greater severity of papilledema, it can be perceived as a complete gray annulus (grade 4). In addition, as the disease progresses, the RNFL begins to exhibit an increasing disorderly pattern close to the optic disc margin. 

The algorithm developed for RNFL appearance captures these changes by calculating features based on the texture of the avascular image generated in the previous step, when segmented vessels were removed. The features include energy, entropy, contrast, cluster shade, and correlation (Table 2). The textural features represent a mathematical description of the gray level relationships between pixels. ¹⁹ An example showing a low entropy texture (grade 1) against high entropy texture (grade 3) image is shown in Figure 7, noting how in the image of more severe stage papilledema the texture shows a more disorderly (higher entropy) pattern when compared to the image of the lower grade papilledema. An average value for the texture features was calculated in the avascular annular region encompassing an area from 0.5 to 1.5 diameters from the center of the optic disc (Fig. 7). 

Figure 7.

 

Examples of low (left) and high (right) Haralick's entropy peripapillary textures. Extracted from a grade 1 (entropy = 0.003; left) and a grade 3 (entropy = 0.0420; right) image. Entropy was calculated in the annular region shown superimposed on the original images.

Classification

An ANOVA was performed on each feature against ground truth defined by expert Frisén grading of papilledema. Features not showing statistical significance (P < 0.05) for separating at least one grade of papilledema from a lesser grade were omitted from our final feature matrix. The final feature matrix consisted of 24 features, 7 from blood vessel features, 12 from disc margin obscuration, and 5 from peripapillary texture. The features used in this algorithm are shown in Table 2. 

For classification, a methodology known as decision tree forest was used. ^20,21 This methodology creates an ensemble of classifiers, where each is trained with a different portion of a training set. ²² Classification is achieved through an analytical combination of their individual results. This methodology avoids overfitting, and reduces instability and variable selection bias, which are well-known problems of tree-based methods. ²¹  

Testing

For testing the classifier, a leave-one-patient-out test was performed. ²³ All visits and all images from both of the selected patient's eyes were left out to ensure that no bias was introduced into the training. In each run, a patient was selected from our population (n = 39). Then, the algorithm was trained on the remaining patients, creating a classification model. This model was then tested on the excluded patient's images, generating the predicted grades for each of the patient visits. These steps were repeated for each patient in our dataset, generating a vector of predicted grades for each patient. A final classification per patient visit was chosen by selecting the maximum predicted papilledema grade for each patient visit between both eyes. In this way, we classified the patient on the eye with the most severe grade of papilledema, since this provides a more meaningful classification for urgency of referral. This approach is based on the observation that papilledema is typically bilateral and was found in this data set to typically have a difference of one or less Frisén scale categories between a patient's eyes, as shown in Table 3. One patient of the 39 had a bilateral difference in MFS of three grades in one of their visits. 

Table 3.

 

Intrapatient Visit Variation in Papilledema between the Right and Left Eyes on a Given Clinic Visit Day in the Dataset Used in This Study

Modified Frisén Scale Grade	Modified Frisén Scale Grade
	0	1	2	3	4
0	10	17	0	1	0
1		31	16	0	0
2			11	0	0
3				4	1
4					5

 The entries in bold indicate cases where the left and the right eye showed the same papilledema grade.

The importance of each feature was evaluated by permuting its value across the entire population, and then measuring the classifier mean squared error against the expert grade (ground truth). This permutation was repeated for each feature, allowing the features to be ranked in order of importance for the model. Features that are more influential in the model will cause the classifier to produce a greater error from ground truth when altered; therefore, they are ranked as more important than features that did not alter the error as much. 

To evaluate the performance of the algorithm, the algorithm's predicted grades were compared against the grades given by the neuro-ophthalmologist (RK). To measure agreement, the quadratic Cohen's weighted κ ²⁴ coefficient between the expert grade of papilledema and that of the algorithm was calculated. Cohen's κ can be any value in −1 to 1, with 1 showing perfect agreement and −1 showing complete disagreement. 

Results

The mathematical models that were developed in this study to characterize the Frisén grading features in digital fundus images were successful in differentiating each papilledema grade according to the MFS. Results are shown in Figure 8 for the VDI, in Figure 9 for the Tukey's honest significance difference (HSD), ^{25 –27} of the total disc margin obscuration (disc obscuration), and in Figure 10 for the Tukey's HSD of the mean Haralick's entropy of the peripapillary RNFL. Tukey's HSD is a conservative multicomparison statistical test similar to the t-test, to determine statistically significant separation between the means of different populations. This test corrects for type I errors keeping our P value fixed. 

Figure 8.

 

Boxplot of the VDI versus the Frisén scale for all eyes. The center line in each box represents the median. Top and bottom of the box: the 25th and 75th percentiles, respectively. Dashed lines: the 5^th (top) and 95th (bottom) percentiles. +, outliers. A t-test showed that the VDI of images grade 0, 1, and 2 were significantly different from the VDI of images grade 3 and 4 (P < 0.05). These results follow clinical observations, as the major vessels start getting obscured in grade 3. The VDI showed correlation with the MFS grade (Spearman correlation coefficient = 0.50).

Figure 9.

 

Tukey's HSD between grades and their obscuration index, with P < 0.05. The total disc margin obscuration (nasal + temporal) is defined as the standard deviation of the distance of points in the optic disc margin to the center of the disc. The difference was significant between grade 0 and the other grades, and between grade 1 and the other grades. The metric did not show a significant difference between grades 2, 3, and 4. The separation follows clinical interpretation as in grade 0 neither the nasal nor the temporal pole contribute to the total obscuration. In grade 1, the nasal pole starts contributing to the total disc obscuration, followed by the temporal pole in grade 2 papilledema.

Figure 10.

 

Tukey's HSD between grade of papilledema and the mean Haralick's entropy of the RNFL from images in the peripapillary region with P < 0.05. A higher mean Haralick's entropy represents more disorganization of structure and striations of the RNFL in the peripapillary retina over the entire circumferential region of interest analyzed. The difference was significant between grade 0 and the others and between grade 1 and the other grades. The metric did not show a significant difference between grades 2, 3, and 4. The first separation of grade 1 shows the initial formation of the ring while the later separation of grades 2, 3, and 4 from the rest represents the completion of the circumferential ring, which agrees with clinical interpretation.

In Figure 8, it can be observed how the VDI increases directly with Frisén grade (Spearman correlation coefficient = 0.50), confirming our hypothesis that the more gaps that occur in the vessel segmentation image, the higher the VDI. The outliers in grades 0, 1, and 2 (+ markers in Fig. 7), represent cases where the algorithm over- and undersegmented the vasculature, because of blur in the image caused by either defocus or cataracts. 

In Figure 9, one can see the obscuration of the optic disc margin caused an increase in deviation of the border determination, helping to significantly separate grades 0 and 1 from the other grades. 

In Figure 10 shows that the texture of the peripapillary region allowed a significant statistical separation of grade 1 from grades 2, 3, and 4. 

The algorithm incorporating all significant features showed substantial agreement (κ = 0.71, P < 0.001) with ground truth when evaluating per visit and substantial agreement (κ = 0.61, P < 0.001) when evaluating per image. The confusion matrix of the predictions after the leave-one-patient-out test per patient is shown in Table 4. 

Table 4.

 

Confusion Matrix for Patient-Visit Classification of Grade of Papilledema, Comparing the Expert Grading (Ground Truth) with the Feature Algorithm Prediction of Grade of Papilledema for Each Patient

Predicted Modified Frisén Scale	Ground Truth Modfied Frisén Scale
	0	1	2	3	4
0	0	1	0	0	0
1	9	42	7	1	0
2	1	5	18	2	2
3	0	0	1	0	0
4	0	0	1	2	4

 Numbers in the matrix below the bold numbers (where the predicted and ground truth agree) are cases where the predicted grade was greater than the expert's grade. Numbers above the bold numbers are cases where the predicted grade was less than the expert's grade.

The Tukey's HSD test ^{24 –26} was applied to each feature across all grades to analyze which features separated each grade of papilledema. The HSD is a conservative test that determines whether two population means are significantly different from each other, which takes into consideration the random error in each comparison. A matrix counting the number of features that significantly separated the mean of one grade of papilledema from another (P < 0.05) is shown in Table 5. 

Table 5.

 

Number of Features That Differentiate One Grade of Papilledema from Another, According to Tukey's HSD Test

Modified Frisén Scale Grade	Modified Frisén Scale Grade
	0	1	2	3	4
0
1	7
2	20	12
3	21	15	1
4	24	23	10	4

The top 10 features, in order of importance, are shown in Table 6. 

Table 6.

 

Top 10 Features in Order of Importance

Feature
1. Vessel discontinuity index
2. Obscuration around the disc margin
3. Nasal to temporal ratio of disc margin obscuration
4. Vessel Discontinuity index to disc proximity
5. Area of the largest region in the vessel map
6. Superior segment of nasal to temporal ratio of disc margin obscuration
7. Central segment of nasal to temporal ratio of disc margin obscuration
8. Disc margin obscuration in the nasal pole
9. Disc margin obscuration in the temporal pole
10. Inferior segment of nasal-to-temporal ratio of disc margin obscuration

Discussion

After the established Frisén grading protocol for papilledema, we were able to identify and measure features which differentiate (with a P < 0.05) each of the grades for which we had samples. These features were grouped into three main categories, vessel features, optic disc margin features, and peripapillary texture features. In this article, we defined a metric called the vessel discontinuity index which increases as the grade of papilledema increases (Fig. 8). The Spearman correlation coefficient showed a 0.50 correlation between papilledema grade and VDI. This feature showed its power in separating grades 0, 1, and 2 from grades 3, and 4, with a P < 0.05. 

The optic disc margin features were important in separating (P < 0.05) grade 0 (normal) from the other categories of papilledema, and papilledema grade 1 from the more advanced stages (i.e., grades 2 through 4; Fig. 9). This finding agrees with the MFS on which basis the progression of papilledema on asymmetric obscuration in grade 1 and an increasing general lack of definition of the disc margin as the disease progresses. In other words, for grade 0, the obscuration metric remains low, as neither the nasal nor the temporal poles contribute to the total obscuration. Grade 1, where the nasal pole becomes obscured while the temporal remains sharp, is captured by the quantitative obscuration metric. Another increase in asymmetric obscuration occurs once the disease progresses to grade 2, where both poles lack definition and contribute to the total obscuration. 

The texture feature was equally effective at separating normal from grade 1 and both from the advanced grades (Fig. 10) with a P < 0.05. This supporting information increased the power of the classifier. 

As one can see from Table 4, the algorithm successfully classifies subjects with papilledema using nonstereo fundus images. The agreement with the expert obtained in this study (κ = 0.71) is considered substantial agreement, as shown by Landis and Koch. ²⁸ The algorithm agreed perfectly with the expert 67% of the time and within one grade 94% of the time. The agreement in these results is comparable to the intergrader variability of human experts as shown in the study by Scott et al. ¹¹ where three experts agreed perfectly 54% of the time and within one grade 97% of the time in a dataset of 28 images. The agreement between medical experts is limited by natural differences in their visual perception. With more data, our algorithm is likely to further improve its correct classification rate, making it more consistent across all papilledema grades. 

The imbalance in distribution of grades in our data set made automatic classification more challenging. Another limiting factor was that a constraint was set not to use the test patient's other examinations for training. Most images (n = 48) presented papilledema at grade 1, with only 11 visits presenting grades 3 or 4. It is intended to refine the algorithm further in future studies with larger numbers of examples of the more severe grades of papilledema. In addition, because agreement among expert neuro-ophthalmologists on the classification of papilledema is not always achieved, ¹¹ alternative approaches for establishing a better ground truth papilledema grade for each image in the future may be to institute a winner-take-all adjudication among three expert graders. 

Other alternative methods are being developed within our group (led by RK), such as the use of optical coherence tomography (OCT) to derive disc volume above the plane of the outer retina as a method of specifying severity of papilledema, which could be used as a separate ground truth on which to train the algorithms. Such approaches are complimentary and are likely to further improve the performance of the algorithms for classifying papilledema from fundus photos and will be the subject of a future study. 

It was observed that contiguous grades have fewer features separating their means (Table 5). This is expected because, although the MFS is discrete, changes in the optic disc occur continuously. An interesting example is the difference between Frisén grades 2 and 3, which only shows 1 feature (VDI) separating their means. Although, according to the MFS, the obscuration of major vessels (excluding their origin in the center of the disc) is a feature of grade 3, the addition of other relevant features would provide a better separation between grades. Another feature that the MFS uses to separate grades 3 from 2 is the shape of the peripapillary annulus overlying the RNFL (showing fringes in grade 3). Further work on adding features from the shape of the ring should further improve the algorithm's performance. 

Analysis of the top 10 features shown in Table 6 reveals that the VDI and definition of the disc margin obscuration (standard deviation in the location of the margin of the optic disc) represent the most power in the classification. The VDI was expected for differentiating grades 2, 3, and 4, as it varies with those MFS grades. The dominance of the optic disc margin obscuration factors can be attributed to the greater distribution of lower grades of papilledema in our dataset, where this feature is important in differentiating lower grades of papilledema. One interesting result was the fifth most significant feature listed in Table 5, the “area of the largest region in the vessel map.” This finding may reflect enlargement and engorgement of venous divisions of the central retinal vein, as it is known that increases in optic disc edema may be associated with varying degrees of venous stasis and may prove to be an important key feature in further refinement of the algorithm. 

Although the results reported in this article are promising, all visual features have not been taken into consideration. The elevation of the optic disc characterized by stereo information could allow us to obtain better accuracy and would contribute to a continuous scale in relation to the severity of papilledema. 

These results show that correct classification of papilledema grade is possible using region-based features of optic disc images. This algorithm could be used for rapid assessment of digital photos acquired when papilledema is being considered in clinical, intensive care, and emergency response settings, given that more user-friendly, nonmydriatic retinal digital cameras are becoming more widely available in non–eye care settings. Early detection of papilledema is very important in emergency response, as the cause can be life threatening in some cases. The system, once implemented at an emergency room, will allow the emergency treatment team to obtain rapid yet accurate papilledema grading from patients even if an expert in evaluating papilledema is not available, allowing them to take an appropriate, timely course of action. 

Footnotes

 Supported by the Department of Veterans Affairs Rehabilitation Research and Development Division through the Iowa City VA Center for the Prevention and Treatment of Visual Loss; Research to Prevent Blindness New York, NY (unrestricted departmental grant). VisionQuest Biomedical, LLC sponsored the research project.

Footnotes

 Disclosure: S. Echegaray, VisionQuest Biomedical, LLC (F, E); G. Zamora, VisionQuest Biomedical, LLC (F, E); H. Yu, VisionQuest Biomedical, LLC (F, E); W. Luo, VisionQuest Biomedical, LLC (F); P. Soliz,VisionQuest Biomedical, LLC (F, E); R. Kardon, VisionQuest Biomedical, LLC (F)