Automated segmentation of the optic disc from stereo color photographs using physiologically plausible features

A guide to stereoscopic 3D displays in medicine

Stereoscopic displays can potentially improve many aspects of medicine. However, weighing the advantages and disadvantages of such displays remains difficult, and more insight is needed to evaluate whether stereoscopic displays are worth adopting. In this article, we begin with a review of monocular and binocular depth cues. We then apply this knowledge to examine how stereoscopic displays can potentially benefit diagnostic imaging, medical training, and surgery. It is apparent that the binocular depth information afforded by stereo displays 1) aid the detection of diagnostically relevant shapes, orientations, and positions of anatomical features, especially when monocular cues are absent or unreliable; 2) help novice surgeons orient themselves in the surgical landscape and perform complicated tasks; and 3) improve the three-dimensional anatomical understanding of students with low visual-spatial skills. The drawbacks of stereo displays are also discussed, including extra eyewear, potential three-dimensional misperceptions, and the hurdle of overcoming familiarity with existing techniques. Finally, we list suggested guidelines for the optimal use of stereo displays. We provide a concise guide for medical practitioners who want to assess the potential benefits of stereo displays before adopting them.

Optic disk and cup segmentation from monocular color retinal images for glaucoma assessment

Automatic retinal image analysis is emerging as an important screening tool for early detection of eye diseases. Glaucoma is one of the most common causes of blindness. The manual examination of optic disk (OD) is a standard procedure used for detecting glaucoma. In this paper, we present an automatic OD parameterization technique based on segmented OD and cup regions obtained from monocular retinal images. A novel OD segmentation method is proposed which integrates the local image information around each point of interest in multidimensional feature space to provide robustness against variations found in and around the OD region. We also propose a novel cup segmentation method which is based on anatomical evidence such as vessel bends at the cup boundary, considered relevant by glaucoma experts. Bends in a vessel are robustly detected using a region of support concept, which automatically selects the right scale for analysis. A multi-stage strategy is employed to derive a reliable subset of vessel bends called r-bends followed by a local spline fitting to derive the desired cup boundary. The method has been evaluated on 138 images comprising 33 normal and 105 glaucomatous images against three glaucoma experts. The obtained segmentation results show consistency in handling various geometric and photometric variations found across the dataset. The estimation error of the method for vertical cup-to-disk diameter ratio is 0.09/0.08 (mean/standard deviation) while for cup-to-disk area ratio it is 0.12/0.10. Overall, the obtained qualitative and quantitative results show effectiveness in both segmentation and subsequent OD parameterization for glaucoma assessment.

Computational quantification of complex fundus phenotypes in age-related macular degeneration and Stargardt disease

PURPOSE

To describe an automated method of quantification of specific fundus phenotypes and evaluate its performance in differentiating drusen, the hallmark lesions of age-related macular degeneration (AMD), from similar-looking bright lesions, the pisciform deposits or flecks typical of Stargardt disease (SD).

METHODS

Fundus macular images of 30 eyes of 30 subjects were studied. Fifteen subjects had a clinical diagnosis of AMD with at least 10 intermediate and/or 1 large drusen, and the other 15 had SD. As a test of bright-lesion separation, AMD and SD subjects were chosen from the heterogeneous phenotypes of each disorder, to be as visually similar as possible. Drusen and fleck properties were quantified from the color images by using an automated method, and a shape classifier was used to divide the images as characteristic of either AMD or SD. Image identification performance was quantified by using the area under the receiver operating characteristic curve (AUC).

RESULTS

All SD subjects demonstrated at least one disease-associated variant of the ABCA4 gene. The method achieved an AUC of 0.936 for differentiating AMD from SD.

CONCLUSIONS

Automated quantification of fundus phenotypes was achieved, and the results show that the method can differentiate AMD from SD, two distinctly different genetically associated disorders, by quantifying the properties of the bright lesions (drusen and flecks) in their fundus images, even when the images were visually selected to be similar. Quantification of fundus phenotypes may allow recognition of new phenotypes, correlation with new genotypes and may measure disease-specific biomarkers to improve management of patients with AMD or SD.

Comparison of automated analysis of Cirrus HD OCT spectral-domain optical coherence tomography with stereo photographs of the optic disc

OBJECTIVE

To evaluate a new automated analysis of optic disc images obtained by spectral-domain optical coherence tomography (SD OCT). Areas of the optic disc, cup, and neural rim in SD OCT images were compared with these areas from stereoscopic photographs to represent the current traditional optic nerve evaluation. The repeatability of measurements by each method was determined and compared.

DESIGN

Evaluation of diagnostic technology.

PARTICIPANTS

One hundred nineteen healthy eyes, 23 eyes with glaucoma, and 7 glaucoma suspect eyes.

METHODS

Optic disc and cup margins were traced from stereoscopic photographs by 3 individuals independently. Optic disc margins and rim widths were determined automatically in SD OCT. A subset of photographs was examined and traced a second time, and duplicate SD OCT images also were analyzed.

MAIN OUTCOME MEASURES

Agreement among photograph readers, between duplicate readings, and between SD OCT and photographs were quantified by the intraclass correlation coefficient (ICC), by the root mean square, and by the standard deviation of the differences.

RESULTS

Optic disc areas tended to be slightly larger when judged in photographs than by SD OCT, whereas cup areas were similar. Cup and optic disc areas showed good correlation (0.8) between the average photographic reading and SD OCT, but only fair correlation of rim areas (0.4). The SD OCT was highly reproducible (ICC, 0.96-0.99). Each reader also was consistent with himself on duplicate readings of 21 photographs (ICC, 0.80-0.88 for rim area and 0.95-0.98 for all other measurements), but reproducibility was not as good as SD OCT. Measurements derived from SD OCT did not differ from photographic readings more than the readings of photographs by different readers differed from each other.

CONCLUSIONS

Designation of the cup and optic disc boundaries by an automated analysis of SD OCT was within the range of variable designations by different readers from color stereoscopic photographs, but use of different landmarks typically made the designation of the optic disc size somewhat smaller in the automated analysis. There was better repeatability among measurements from SD OCT than from among readers of photographs. The repeatability of automated measurement of SD OCT images is promising for use both in diagnosis and in monitoring of progression.

Optimal filter framework for automated, instantaneous detection of lesions in retinal images

Automated detection of lesions in retinal images is a crucial step towards efficient early detection, or screening, of large at-risk populations. In particular, the detection of microaneurysms, usually the first sign of diabetic retinopathy (DR), and the detection of drusen, the hallmark of age-related macular degeneration (AMD), are of primary importance. In spite of substantial progress made, detection algorithms still produce 1) false positives-target lesions are mixed up with other normal or abnormal structures in the eye, and 2) false negatives-the large variability in the appearance of the lesions causes a subset of these target lesions to be missed. We propose a general framework for detecting and characterizing target lesions almost instantaneously. This framework relies on a feature space automatically derived from a set of reference image samples representing target lesions, including atypical target lesions, and those eye structures that are similar looking but are not target lesions. The reference image samples are obtained either from an expert- or a data-driven approach. Factor analysis is used to derive the filters generating this feature space from reference samples. Previously unseen image samples are then classified in this feature space. We tested this approach by training it to detect microaneurysms. On a set of images from 2739 patients including 67 with referable DR, DR detection area under the receiver-operating characteristic curve (AUC) was comparable (AUC=0.927) to our previously published red lesion detection algorithm (AUC=0.929). We also tested the approach on the detection of AMD, by training it to differentiate drusen from Stargardt's disease lesions, and achieved an AUC=0.850 on a set of 300 manually detected drusen and 300 manually detected flecks. The entire image processing sequence takes less than a second on a standard PC compared to minutes in our previous approach, allowing instantaneous detection. Free-response receiver-operating characteristic analysis showed the superiority of this approach over a framework where false positives and the atypical lesions are not explicitly modeled. A greater performance was achieved by the expert-driven approach for DR detection, where the designer had sound expert knowledge. However, for both problems, a comparable performance was obtained for both expert- and data-driven approaches. This indicates that annotation of a limited number of lesions suffices for building a detection system for any type of lesion in retinal images, if no expert-knowledge is available. We are studying whether the optimal filter framework also generalizes to the detection of any structure in other domains.

Automated segmentation of optic disc region on retinal fundus photographs: Comparison of contour modeling and pixel classification methods

The automatic determination of the optic disc area in retinal fundus images can be useful for calculation of the cup-to-disc (CD) ratio in the glaucoma screening. We compared three different methods that employed active contour model (ACM), fuzzy c-mean (FCM) clustering, and artificial neural network (ANN) for the segmentation of the optic disc regions. The results of these methods were evaluated using new databases that included the images captured by different camera systems. The average measures of overlap between the disc regions determined by an ophthalmologist and by using the ACM (0.88 and 0.87 for two test datasets) and ANN (0.88 and 0.89) methods were slightly higher than that by using FCM (0.86 and 0.86) method. These results on the unknown datasets were comparable with those of the resubstitution test; this indicates the generalizability of these methods. The differences in the vertical diameters, which are often used for CD ratio calculation, determined by the proposed methods and based on the ophthalmologist's outlines were even smaller than those in the case of the measure of overlap. The proposed methods can be useful for automatic determination of CD ratios.

Learning-based approach for the automatic detection of the optic disc in digital retinal fundus photographs

The optic disc is an important feature in the retina. We propose a method for the detection of the optic disc based on a supervised learning scheme. The method employs pixel and local neighbourhood features extracted from the ROI of a digital retinal fundus photograph. A support vector machine based classification mechanism is used to classify each image point as belonging to the cup and retina. The proposed method is evaluated on a sample image set of 68 retinal fundus images. The results show a high correlation (r>0.9) with the ground truth segmentation, with an overlap error of 6.02%, and found to be comparable to the inter-observer variability based on an independent second observer segmentation of the same data set.

Detecting the optic disc boundary in digital fundus images using morphological, edge detection, and feature extraction techniques

Optic disc (OD) detection is an important step in developing systems for automated diagnosis of various serious ophthalmic pathologies. This paper presents a new template-based methodology for segmenting the OD from digital retinal images. This methodology uses morphological and edge detection techniques followed by the Circular Hough Transform to obtain a circular OD boundary approximation. It requires a pixel located within the OD as initial information. For this purpose, a location methodology based on a voting-type algorithm is also proposed. The algorithms were evaluated on the 1200 images of the publicly available MESSIDOR database. The location procedure succeeded in 99% of cases, taking an average computational time of 1.67 s. with a standard deviation of 0.14 s. On the other hand, the segmentation algorithm rendered an average common area overlapping between automated segmentations and true OD regions of 86%. The average computational time was 5.69 s with a standard deviation of 0.54 s. Moreover, a discussion on advantages and disadvantages of the models more generally used for OD segmentation is also presented in this paper.

Automated segmentation of neural canal opening and optic cup in 3D spectral optical coherence tomography volumes of the optic nerve head

PURPOSE

To develop an automated approach for segmenting the neural canal opening (NCO) and cup at the level of the retinal pigment epithelium (RPE)/Bruch's membrane (BM) complex in spectral-domain optical coherence tomography (SD-OCT) volumes. To investigate the correspondence and discrepancy between the NCO-based metrics and the clinical disc margin on fundus photographs of glaucoma subjects.

METHODS

SD-OCT scans and corresponding stereo fundus photographs of the optic nerve head were obtained from 68 eyes of 34 patients with glaucoma or glaucoma suspicion. Manual planimetry was performed by three glaucoma experts to delineate a reference standard (RS) for cup and disc margins from the images. An automated graph-theoretic approach was used to identify the NCO and cup. NCO-based metrics were compared with the RS.

RESULTS

Compared with the RS disc margin, the authors found mean unsigned and signed border differences of 2.81 ± 1.48 pixels (0.084 ± 0.044 mm) and -0.99 ± 2.02 pixels (-0.030 ± 0.061 mm), respectively, for NCO segmentation. The correlations of the linear cup-to-disc (NCO) area ratio, disc (NCO) area, rim area, and cup area of the algorithm with the RS were 0.85, 0.77, 0.69, and 0.83, respectively.

CONCLUSIONS

In most eyes, the NCO-based 2D metrics, as estimated by the novel automated graph-theoretic approach to segment the NCO and cup at the level of the RPE/BM complex in SD-OCT volumes, correlate well with RS. However, a small discrepancy exists in NCO-based anatomic structures and the clinical disc margin of the RS in some eyes.

Automated quantification of inherited phenotypes from color images: a twin study of the variability of optic nerve head shape

PURPOSE

Discovery and description of heritable optic nerve head (ONH) phenotypes have been haphazard. In this preliminary study, the authors test the hypothesis that inheritable phenotypes can be discovered and quantified computationally by estimating three-dimensional ONH shape parameters from stereo color photographs from the Twins Eye Study in Tasmania and determining how much of the variability in ONH shape is accounted for by genetic influence.

METHODS

Three-dimensional ONH shape was estimated by an automated algorithm from stereoscopic optic disc color photographs of a random sample of 172 subjects (344 eyes, 45 pairs of monozygotic [MZ] and 41 dizygotic [DZ] twins). Shape resemblances between eyes were quantified with a distance metric. The heritability of the shape resemblance was determined both through the distribution of the discongruence indices and through structural equation modeling techniques (ACE model).

RESULTS

Significantly different discongruence indices were found for MZ (1.0286; 95% CI, 0.9872-1.0701) and DZ twins (1.4218; 95% CI, 1.2631-1.5804); larger indices for DZ twins indicated that variability was substantially determined by genetic factors. The standardized variances of the A(dditive genetic), C(ommon environmental), and (nonshared) E(nvironmental) components were 0.80, 2.00 × 10(-15) and 0.20, respectively, for all OD, and 0.79, 3.24 × 10(-14), and 0.21 for all OS.

CONCLUSIONS

This preliminary study shows that quantitative phenotyping of the ONH shape from color images leads to phenotypes that can be measured and are largely under genetic control. The association of these inherited phenotypes with genotypes deserves confirmation and further study.

Automated volumetric evaluation of stereoscopic disc photography

PURPOSE

To develop a fully automated algorithm (AP) to perform a volumetric measure of the optic disc using conventional stereoscopic optic nerve head (ONH) photographs, and to compare algorithm-produced parameters with manual photogrammetry (MP), scanning laser ophthalmoscope (SLO) and optical coherence tomography (OCT) measurements.

METHODS

One hundred twenty-two stereoscopic optic disc photographs (61 subjects) were analyzed. Disc area, rim area, cup area, cup/disc area ratio, vertical cup/disc ratio, rim volume and cup volume were automatically computed by the algorithm. Latent variable measurement error models were used to assess measurement reproducibility for the four techniques.

RESULTS

AP had better reproducibility for disc area and cup volume and worse reproducibility for cup/disc area ratio and vertical cup/disc ratio, when the measurements were compared to the MP, SLO and OCT methods.

CONCLUSION

AP provides a useful technique for an objective quantitative assessment of 3D ONH structures.

Intelligent fusion of cup-to-disc ratio determination methods for glaucoma detection in ARGALI

Glaucoma is a leading cause of permanent blindness. ARGALI, an automated system for glaucoma detection, employs several methods for segmenting the optic cup and disc from retinal images, combined using a fusion network, to determine the cup to disc ratio (CDR), an important clinical indicator of glaucoma. This paper discusses the use of SVM as an alternative fusion strategy in ARGALI, and evaluates its performance against the component methods and neural network (NN) fusion in the CDR calculation. The results show SVM and NN provide similar improvements over the component methods, but with SVM having a greater consistency over the NN, suggesting potential for SVM as a viable option in ARGALI.

Retinal imaging and image analysis

Many important eye diseases as well as systemic diseases manifest themselves in the retina. While a number of other anatomical structures contribute to the process of vision, this review focuses on retinal imaging and image analysis. Following a brief overview of the most prevalent causes of blindness in the industrialized world that includes age-related macular degeneration, diabetic retinopathy, and glaucoma, the review is devoted to retinal imaging and image analysis methods and their clinical implications. Methods for 2-D fundus imaging and techniques for 3-D optical coherence tomography (OCT) imaging are reviewed. Special attention is given to quantitative techniques for analysis of fundus photographs with a focus on clinically relevant assessment of retinal vasculature, identification of retinal lesions, assessment of optic nerve head (ONH) shape, building retinal atlases, and to automated methods for population screening for retinal diseases. A separate section is devoted to 3-D analysis of OCT images, describing methods for segmentation and analysis of retinal layers, retinal vasculature, and 2-D/3-D detection of symptomatic exudate-associated derangements, as well as to OCT-based analysis of ONH morphology and shape. Throughout the paper, aspects of image acquisition, image analysis, and clinical relevance are treated together considering their mutually interlinked relationships.

Retinal imaging and image analysis

Many important eye diseases as well as systemic diseases manifest themselves in the retina. While a number of other anatomical structures contribute to the process of vision, this review focuses on retinal imaging and image analysis. Following a brief overview of the most prevalent causes of blindness in the industrialized world that includes age-related macular degeneration, diabetic retinopathy, and glaucoma, the review is devoted to retinal imaging and image analysis methods and their clinical implications. Methods for 2-D fundus imaging and techniques for 3-D optical coherence tomography (OCT) imaging are reviewed. Special attention is given to quantitative techniques for analysis of fundus photographs with a focus on clinically relevant assessment of retinal vasculature, identification of retinal lesions, assessment of optic nerve head (ONH) shape, building retinal atlases, and to automated methods for population screening for retinal diseases. A separate section is devoted to 3-D analysis of OCT images, describing methods for segmentation and analysis of retinal layers, retinal vasculature, and 2-D/3-D detection of symptomatic exudate-associated derangements, as well as to OCT-based analysis of ONH morphology and shape. Throughout the paper, aspects of image acquisition, image analysis, and clinical relevance are treated together considering their mutually interlinked relationships.

Segmentation of the optic disc in 3-D OCT scans of the optic nerve head

Glaucoma is the second leading ocular disease causing blindness due to gradual damage to the optic nerve and resultant visual field loss. Segmentations of the optic disc cup and neuroretinal rim can provide important parameters for detecting and tracking this disease. The purpose of this study is to describe and evaluate a method that can automatically segment the optic disc cup and rim in spectral-domain 3-D OCT (SD-OCT) volumes. Four intraretinal surfaces were segmented using a fast multiscale 3-D graph search algorithm. After surface segmentation, the retina in each 3-D OCT scan was flattened to ensure a consistent optic nerve head shape. A set of 15 features, derived from the segmented intraretinal surfaces and voxel intensities in the SD-OCT volume, were used to train a classifier that can determine which A-scans in the OCT volume belong to the background, optic disc cup and rim. Finally, prior knowledge about the shapes of the cup and rim was incorporated into the system using a convex hull-based approach. Two glaucoma experts annotated the cup and rim area using planimetry, and the annotations of the first expert were used as the reference standard. A leave-one-subject-out experiment on 27 optic nerve head-centered OCT volumes (14 right eye scans and 13 left eye scans from 14 patients) was performed. Two different types of classification methods were compared, and experimental results showed that the best performing method had an unsigned error for the optic disc cup of 2.52+/-0.87 pixels (0.076+/-0.026 mm) and for the neuroretinal rim of 2.04+/-0.86 pixels (0.061+/-0.026 mm). The interobserver variability as indicated by the unsigned border positioning difference between the second expert observer and the reference standard was 2.54+/-1.03 pixels (0.076+/-0.031 mm for the optic disc cup and 2.14+/-0.80 pixels (0.064+/-0.024 mm for the neuroretinal rim. The unsigned error of the best performing method was not significantly different (p > 0.2) from the interobserver variability.

A critical discussion of the rates of progression and causes of optic nerve damage in glaucoma: International Glaucoma Think Tank II: July 25-26, 2008, Florence, Italy

The International Glaucoma Think Tank II brought together glaucoma clinicians and researchers from all over the world to discuss current practices in glaucoma diagnosis and management, and the neurobiology of glaucoma. The meeting focused on several themes, including rates of deterioration in glaucoma patients, mechanisms of optic nerve damage, and implications for treatment. Issues such as how to measure and integrate progression information into clinical practice, screening protocols, or trials were discussed, as were promising new technologies and limitations of currently available measurement tools. Clinical applications for genetic testing were considered. Study of the neurobiology of glaucoma continues to inform our understanding of underlying degenerative processes, as well as to introduce possibilities for early detection or prevention. Many questions regarding glaucoma pathophysiology and best treatment practices remain unanswered, but with continued research and discussion, we will advance our understanding of this disease and ensure that patients receive optimal care.

Fast detection of the optic disc and fovea in color fundus photographs

A fully automated, fast method to detect the fovea and the optic disc in digital color photographs of the retina is presented. The method makes few assumptions about the location of both structures in the image. We define the problem of localizing structures in a retinal image as a regression problem. A kNN regressor is utilized to predict the distance in pixels in the image to the object of interest at any given location in the image based on a set of features measured at that location. The method combines cues measured directly in the image with cues derived from a segmentation of the retinal vasculature. A distance prediction is made for a limited number of image locations and the point with the lowest predicted distance to the optic disc is selected as the optic disc center. Based on this location the search area for the fovea is defined. The location with the lowest predicted distance to the fovea within the foveal search area is selected as the fovea location. The method is trained with 500 images for which the optic disc and fovea locations are known. An extensive evaluation was done on 500 images from a diabetic retinopathy screening program and 100 specially selected images containing gross abnormalities. The method found the optic disc in 99.4% and the fovea in 96.8% of regular screening images and for the images with abnormalities these numbers were 93.0% and 89.0% respectively.

Automated segmentation of the cup and rim from spectral domain OCT of the optic nerve head

PURPOSE

To evaluate the performance of an automated algorithm for determination of the cup and rim from close-to-isotropic spectral domain (SD) OCT images of the optic nerve head (ONH) and compare to the cup and rim as determined by glaucoma experts from stereo color photographs of the same eye.

METHODS

Thirty-four consecutive patients with glaucoma were included in the study, and the ONH in the left eye was imaged with SD-OCT and stereo color photography on the same day. The cup and rim were segmented in all ONH OCT volumes by a novel voxel column classification algorithm, and linear cup-to-disc (c/d) ratio was determined. Three fellowship-trained glaucoma specialists performed planimetry on the stereo color photographs, and c/d was also determined. The primary outcome measure was the correlation between algorithm-determined c/d and planimetry-derived c/d.

RESULTS

The correlation of algorithm c/d to experts 1, 2, and 3 was 0.90, 0.87, and 0.93, respectively. The c/d correlation of expert 1 to 2, 1 to 3, and 2 to 3, were 0.89, 0.93, and 0.88, respectively.

CONCLUSIONS

In this preliminary study, we have developed a novel algorithm to determine the cup and rim in close-to-isotropic SD-OCT images of the ONH and have shown that its performance for determination of the cup and rim from SD-OCT images is similar to that of planimetry by glaucoma experts. Validation on a larger glaucoma sample as well as normal controls is warranted.

Variance owing to observer, repeat imaging, and fundus camera type on cup-to-disc ratio estimates by stereo planimetry

OBJECTIVE

To determine and compare variance components in linear cup-to-disc ratio (LCDR) estimates by computer-assisted planimetry by human experts, and automated machine algorithm (digital automated planimetry).

DESIGN

Prospective case series for evaluation of planimetry.

PARTICIPANTS

Forty-four eyes of 44 consecutive patients from the outpatient Glaucoma Service at University of Iowa with diagnosis of glaucoma or glaucoma suspect were studied.

METHODS

Six stereo pairs of optic nerve photographs were taken per eye: 3 repeat stereo pairs using simultaneous fixed-stereo base fundus camera (Nidek 3Dx) and another 3 repeat stereo pairs using sequential variable-stereo base fundus camera (Zeiss). Each optic disc stereo pair was digitized and segmented into cup and rim by 3 glaucoma specialists (computer-assisted planimetry) and using a computer algorithm (digital automated planimetry), and LCDR was calculated for each segmentation (either specialist or algorithm). A linear mixed model was used to estimate mean, SD, and variance components of measurements.

MAIN OUTCOME MEASURES

Average LCDR, interobserver, interrepeat, intercamera coefficients of variation (CV) of LCDR and their 95% tolerance limits.

RESULTS

There was a significant difference in LCDR estimates among the 3 glaucoma specialists. The interobserver CV of 10.65% was larger than interrepeat (6.7%) or intercamera CV (7.6%). For the algorithm, the LCDR estimate was significantly higher for simultaneous stereo fundus images (Nidek, mean: 0.66) than for sequential stereo fundus images (Zeiss, mean: 0.64), whereas interrepeat CV for Nidek (4.4%) was lower than Zeiss (6.36%); the algorithm's interrepeat and intercamera CV were 5.47% and 7.26%, respectively.

CONCLUSIONS

Interobserver variability was the largest source of variation for glaucoma specialists, whereas their interrepeat and intercamera variability is comparable with that of the algorithm. DAP reduces variability on LCDR estimates from simultaneous stereo images, such as the Nidek 3Dx.

Automated diagnosis of retinopathy by content-based image retrieval

PURPOSE

To describe a novel computer-based image analysis method that is being developed to assist and automate the diagnosis of retinal disease.

METHODS

Content-based image retrieval is the process of retrieving related images from large database collections using their pictorial content. The content feature list becomes the index for storage, search, and retrieval of related images from a library based upon specific visual characteristics. Low-level analyses use feature description models and higher-level analyses use perceptual organization and spatial relationships, including clinical metadata, to extract semantic information.

RESULTS

We defined, extracted, and tested a large number of region- and lesion-based features from a dataset of 395 retinal images. Using a statistical hold-one-out method, independent queries for each image were submitted to the system and a diagnostic prediction was formulated. The diagnostic sensitivity for all stratified levels of age-related macular degeneration ranged from 75% to 100%. Similarly, the sensitivity of detection and accuracy for proliferative diabetic retinopathy ranged from 75% to 91.7% and for nonproliferative diabetic retinopathy, ranged from 75% to 94.7%. The overall purity of the diagnosis (specificity) for all disease states in the dataset was 91.3%.

CONCLUSIONS

The probabilistic nature of content-based image retrieval permits us to make statistically relevant predictions regarding the presence, severity, and manifestations of common retinal diseases from digital images in an automated and deterministic manner.

Quantitative depth analysis of optic nerve head using stereo retinal fundus image pair

Depth analysis of the optic nerve head (ONH) in the retinal fundus is important for the early detection of glaucoma. In this study, we investigate an automatic reconstruction method for the quantitative depth measurement of the ONH from a stereo retinal fundus image pair. We propose a technique to obtain the depth value from the stereo retinal fundus image pair, which mainly consists of five steps: 1. cutout of the ONH region from the stereo retinal fundus image pair, 2. registration of the stereo image pair, 3. disparity measurement, 4. noise reduction, and 5. quantitative depth calculation. Depth measurements of 12 normal eyes are performed using the stereo fundus camera and the Heidelberg Retina Tomograph (HRT), which is a confocal laser-scanning microscope. The depth values of the ONH obtained from the stereo retinal fundus image pair were in good accordance with the value obtained using HRT (r=0.80+/-0.15). These results indicate that our proposed method could be a useful and easy-to-handle tool for assessing the cup depth of the ONH in routine diagnosis as well as in glaucoma screening.

Automated assessment of the optic nerve head on stereo disc photographs

PURPOSE

To develop automated software for optic nerve head (ONH) quantitative assessment from stereoscopic disc photographs and to evaluate its performance in comparison with human expert assessment.

METHODS

A fully automated system, including three-dimensional ONH modeling, disc margin detection, cup margin detection, and calculation of stereometric ONH parameters, was developed and tested. One eye each from 54 subjects (23 healthy, 17 suspected glaucoma, and 14 glaucoma) was enrolled. The majority opinion of three experts defined disc and cup margins on the disc photographs was used for comparison. Seven ONH parameters, disc area, rim area, rim volume, cup area, cup volume, cup-to-disc (C/D) area ratio, and vertical C/D ratio, were computed based on both machine- and expert-defined margins and compared between the methods.

RESULTS

All automated ONH measurements showed good correlation with the expert defined margins (Pearson r = 0.90, disc area; 0.56, rim area; 0.78, rim volume; 0.88, cup area; 0.93, cup volume; 0.69, C/D area ratio; and 0.67, vertical C/D ratio; all P <or= 0.0001). No statistically significant difference was found in the glaucoma-discriminating ability of all seven ONH parameters (P >or= 0.21). The mean or median of automatically defined disc and cup areas was significantly higher than the subjective assessment (disc area P = 0.0001, t-test; cup area P = 0.036, Wilcoxon signed ranks test), although they had high correlation coefficients. The software failed to detect the disc margin for all the disc photographs with peripapillary atrophy.

CONCLUSIONS

The automated ONH analysis method provides an objective and quantitative ONH evaluation using widely available stereo disc photographs.

Evaluation of a system for automatic detection of diabetic retinopathy from color fundus photographs in a large population of patients with diabetes

OBJECTIVE

To evaluate the performance of a system for automated detection of diabetic retinopathy in digital retinal photographs, built from published algorithms, in a large, representative, screening population.

RESEARCH DESIGN AND METHODS

We conducted a retrospective analysis of 10,000 consecutive patient visits, specifically exams (four retinal photographs, two left and two right) from 5,692 unique patients from the EyeCheck diabetic retinopathy screening project imaged with three types of cameras at 10 centers. Inclusion criteria included no previous diagnosis of diabetic retinopathy, no previous visit to ophthalmologist for dilated eye exam, and both eyes photographed. One of three retinal specialists evaluated each exam as unacceptable quality, no referable retinopathy, or referable retinopathy. We then selected exams with sufficient image quality and determined presence or absence of referable retinopathy. Outcome measures included area under the receiver operating characteristic curve (number needed to miss one case [NNM]) and type of false negative.

RESULTS

Total area under the receiver operating characteristic curve was 0.84, and NNM was 80 at a sensitivity of 0.84 and a specificity of 0.64. At this point, 7,689 of 10,000 exams had sufficient image quality, 4,648 of 7,689 (60%) were true negatives, 59 of 7,689 (0.8%) were false negatives, 319 of 7,689 (4%) were true positives, and 2,581 of 7,689 (33%) were false positives. Twenty-seven percent of false negatives contained large hemorrhages and/or neovascularizations.

CONCLUSIONS

Automated detection of diabetic retinopathy using published algorithms cannot yet be recommended for clinical practice. However, performance is such that evaluation on validated, publicly available datasets should be pursued. If algorithms can be improved, such a system may in the future lead to improved prevention of blindness and vision loss in patients with diabetes.

Detection of anatomic structures in human retinal imagery

The widespread availability of electronic imaging devices throughout the medical community is leading to a growing body of research on image processing and analysis to diagnose retinal disease such as diabetic retinopathy (DR). Productive computer-based screening of large, at-risk populations at low cost requires robust, automated image analysis. In this paper we present results for the automatic detection of the optic nerve and localization of the macula using digital red-free fundus photography. Our method relies on the accurate segmentation of the vasculature of the retina followed by the determination of spatial features describing the density, average thickness, and average orientation of the vasculature in relation to the position of the optic nerve. Localization of the macula follows using knowledge of the optic nerve location to detect the horizontal raphe of the retina using a geometric model of the vasculature. We report 90.4% detection performance for the optic nerve and 92.5% localization performance for the macula for red-free fundus images representing a population of 345 images corresponding to 269 patients with 18 different pathologies associated with DR and other common retinal diseases such as age-related macular degeneration.

Automated detection and differentiation of drusen, exudates, and cotton-wool spots in digital color fundus photographs for diabetic retinopathy diagnosis

PURPOSE

To describe and evaluate a machine learning-based, automated system to detect exudates and cotton-wool spots in digital color fundus photographs and differentiate them from drusen, for early diagnosis of diabetic retinopathy.

METHODS

Three hundred retinal images from one eye of 300 patients with diabetes were selected from a diabetic retinopathy telediagnosis database (nonmydriatic camera, two-field photography): 100 with previously diagnosed bright lesions and 200 without. A machine learning computer program was developed that can identify and differentiate among drusen, (hard) exudates, and cotton-wool spots. A human expert standard for the 300 images was obtained by consensus annotation by two retinal specialists. Sensitivities and specificities of the annotations on the 300 images by the automated system and a third retinal specialist were determined.

RESULTS

The system achieved an area under the receiver operating characteristic (ROC) curve of 0.95 and sensitivity/specificity pairs of 0.95/0.88 for the detection of bright lesions of any type, and 0.95/0.86, 0.70/0.93, and 0.77/0.88 for the detection of exudates, cotton-wool spots, and drusen, respectively. The third retinal specialist achieved pairs of 0.95/0.74 for bright lesions and 0.90/0.98, 0.87/0.98, and 0.92/0.79 per lesion type.

CONCLUSIONS

A machine learning-based, automated system capable of detecting exudates and cotton-wool spots and differentiating them from drusen in color images obtained in community based diabetic patients has been developed and approaches the performance level of retinal experts. If the machine learning can be improved with additional training data sets, it may be useful for detecting clinically important bright lesions, enhancing early diagnosis, and reducing visual loss in patients with diabetes.