Ординатура / Офтальмология / Английские материалы / Automated Image Detection of Retinal Pathology_Jelinek, Cree_2009

[40]Lee, S.C. and Wang, Y., A general algorithm for recognizing small, vague, and imager-alike objects in a nonuniformly illuminated medical diagnostic image, in 32nd Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA, 1998, vol. 2, 941–943.

[41]Lee, S.C., Wang, Y., and Tan, W., Automated detection of venous beading in retinal images, Proceedings of the SPIE, 4322, 1365, 2001.

[42]Larsen, N., Godt, J., Grunkin, M., et al., Automated detection of diabetic retinopathy in a fundus photographic screening population, Investigative Ophthalmology & Visual Science, 44(2), 767, 2003.

[43]Larsen, M., Godt, J., Larsen, N., et al., Automated detection of fundus photographic red lesions in diabetic retinopathy, Invest Ophthalmol Vis Sci, 44(2), 761, 2003.

[44]Sinthanayothin, C., Boyce, J.F., Williamson, T.H., et al., Automated detection of diabetic retinopathy on digital fundus images, Diabet Med, 19(2), 105, 2002.

[45]Delori, F.C. and Pflibsen, K.P., Spectral reflectance of the human ocular fundus, Applied Optics, 28(6), 1061, 1989.

[46]Preece, S.J. and Claridge, E., Monte Carlo modelling of the spectral reflectance of the human eye, Physics in Medicine and Biology, 47, 2863, 2002.

[47]Osareh, A., Mirmehdi, M., Thomas, B., et al., Classification and localisation of diabetic-related eye disease, in 7th European Conference on Computer Vision (ECCV2002), 2002, vol. 2353 of Lecture Notes in Computer Science, 502–516.

[48]Goatman, K.A., Whitwam, A.D., Manivannan, A., et al., Colour normalisation of retinal images, in Proceedings of Medical Image Understanding and Analysis, Sheffield, UK, 2003, 49–52.

[49]Cree, M.J., Gamble, E., and Cornforth, D.J., Colour normalisation to reduce inter-patient and intra-patient variability in microaneurysm detection in colour retinal images, in APRS Workshop on Digital Image Computing (WDIC2005), Brisbane, Australia, 2005, 163–168.

[50]Perreault, S. and Hebert,´ P., Median filtering in constant time, IEEE Transactions on Image Processing, 16, 2389, 2007.

[51]Streeter, L. and Cree, M.J., Microaneurysm detection in colour fundus images, in Proceedings of the Image and Vision Computing New Zealand Conference (IVCNZ’03), Palmerston North, New Zealand, 2003, 280–285.

[52]Sharp, P.F., Olson, J., Strachan, F., et al., The value of digital imaging in diabetic retinopathy, Health Technology Assessment, 7(30), 1, 2003.

[53]Bayes, T., An essay towards solving a problem in the doctrine of chances,

Philosophical Transactions of the Royal Society of London, 53, 370, 1763.

[54]Han, J. and Kamber, M., Data Mining Concepts and Techniques, Morgan Kaufman, 2001.

[55]Metz, C.E., ROC methodology in radiologic imaging, Investigative Radiology, 21, 720, 1986.

[56]Jelinek, H.F., Cree, M.J., Worsley, D., et al., An automated microaneurysm detector as a tool for identification of diabetic retinopathy in rural optometric practice, Clinical and Experimental Optometry, 89(5), 299, 2006.

[57]Lee, S.C. and Wang, Y., Automatic retinal image quality assessment and enhancement, in Medical Imaging 1999: Image Processing, 1999, vol. 3661 of Proceedings of the SPIE, 1581–1590.

[58]Fleming, A., Philip, S., Goatman, K., et al., Automated assessment of retinal image field of view, in Medical Image Understanding and Analysis (MIUA 2004), London, UK, 2004, 129–132.

[59]Newsom, R.S.B., Clover, A., Costen, M.T.J., et al., Effect of digital image compression on screening for diabetic retinopathy, British Journal of Ophthalmology, 85(7), 799, 2001.

[60]Basu, A., Kamal, A.D., Illahi, W., et al., Is digital image compression acceptable within diabetic retinopathy screening? Diabetic Medicine, 20(9), 766, 2003.

[61]Stellingwerf, C., Hardus, P.L.L.J., and Hooymans, J.M.M., Assessing diabetic retinopathy using two-field digital photography and the influence of JPEGcompression, Documenta Ophthalmologica, 108(3), 203, 2004.

[62]Anagnoste, S.R., Orlock, D., Spaide, R., et al., Digital compression of ophthalmic images: Maximum JPEG compression undetectable by retina specialists, Investigative Ophthalmology & Visual Science, 39, S922, 1998.

[63]Lee, M.S., Shin, D.S., and Berger, J.W., Grading, image analysis, and stereopsis of digitally compressed fundus images, Retina, 20(3), 275, 2000.

[64] Baker, C.F., Rudnisky, C.J., Tennant, M.T.S., et al., JPEG compression of stereoscopic digital images for the diagnosis of diabetic retinopathy via teleophthalmology, Canadian Journal of Ophthalmology, 39(7), 746, 2004.

[65]Hansgen,¨ P., Undrill, P.E., and Cree, M.J., The application of wavelets to retinal image compression and its effect on automatic microaneurysm analysis,

Computer Methods and Programs in Biomedicine, 56(1), 1, 1998.

[66]Cree, M.J. and Jelinek, H.F., The effect of JPEG compression on automated detection of microaneurysms in retinal images, in Electronic Imaging Symposium: Image Processing: Machine Vision Applications, San Jose, CA, 2008, vol. 6813 of Proceedings of the SPIE, 68 130M–1.

[67]Hanley, J.A. and McNeil, B.J., The meaning and use of the area under a receiver operating characteristic (ROC) curve, Radiology, 143, 29, 1982.

[68]Hoover, A., Kouznetsova, V., and Goldbaum, M., Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response,

IEEE Transactions on Medical Imaging, 19(3), 203, 2000.

[69]Staal, J., Abramoff, M.D., Niemeijer, M., et al., Ridge-based vessel segmentation in color images of the retina, IEEE Transactions on Medical Imaging, 23(4), 501, 2004.

[70]Hellstedt, T. and Immonen, I., Disappearance and formation rates of microaneurysms in early diabetic retinopathy, British Journal of Ophthalmology, 80, 135, 1996.

[71]Baudoin, C., Maneshi, F., Quentel, G., et al., Quantitative-evaluation of fluorescein angiograms-microaneurysm counts, Diabetes, 32, 8, 1983.

[72]Jalli, P.Y., Hellstedt, T.J., and Immonen, I.J., Early versus late staining of microaneurysms in fluorescein angriography, Retina, 17, 211, 1997.

7.2Early Description of Retinal Vascular Changes

Cardiovascular disease remains the leading cause of death in developed countries. Clinicians have long been relying on a spectrum of “traditional” cardiovascular risk factors, such as the presence of diabetes, hypertension, hyperlipidemia and cigarette smoking to help identify, monitor, and treat high-risk individuals [1–4]. However, there is a belief that these “traditional” risk factors are inadequate in explaining a substantial proportion of cardiovascular morbidity and mortality [5–7]. To pursue further refinement in cardiovascular risk stratification, there has been increasing interest in searching for other biomarkers of predictors that may provide additional information regarding a person’s cardiovascular risk [8–10].

The human retina may provide an ideal platform to hunt for these variables as the retinal vasculature, which can be visualized directly, shares similar anatomical and physiological properties with the cerebral and coronary circulation [11–14]. Retinal vascular damage in essence reflects damage from chronic hypertension, diabetes, and other processes [12; 15–19], and may therefore be a biomarker for cardiovascular disease risk [11; 14].

Recognition of this potential dates back a century ago [20–23] when Keith, Wagener, and Barker showed that the severity of retinal vascular changes was predictive of mortality in patients with hypertension [24]. Subsequent researchers described associations of various retinopathy signs with different cardiovascular diseases and mortality [25–32]. Nevertheless, since the 1960s there has been less interest in this field of research for a number of reasons. First, the association between retinal vascular changes and cardiovascular disease was not consistently demonstrated in all studies [25; 26; 33–38], and the majority of studies had significant limitations, such as lack of control for potential confounders and the use of clinicor hospital-based patients not representative of the general population. Second, early studies were conducted in populations with untreated hypertension, in which the more severe retinal abnormalities were observed (e.g. Keith, Wagener, and Barker Grade III and IV retinopathy [24].). This degree of retinopathy was felt to be uncommon in the modern setting with better blood pressure control [39–43]. Third, despite many attempts to improve previous grading systems, there was no general consensus regarding a clinically meaningful and standardized classification of retinal vascular signs [44]. Finally, the detection of retinal vascular abnormalities using the direct ophthalmoscope was found to be subjective and unreliable [36–38; 45–47].

These issues have been largely addressed in the last decade with advancements in the field of retinal photography and image-analysis techniques, enabling objective quantification of a range of retinal vascular parameters [48] and large populationbased studies to determine clinical significance [49]. Furthermore, more subtle changes in the retinal vasculature (e.g., degree of generalized retinal arteriolar narrowing) are now quantifiable from high-resolution digitized photographs [49–53]. These new approaches to study retinal vascular characteristics have renewed the interest in using digital retinal photography as a potential tool for the prediction of systemic cardiovascular disorders.

Retinal Vascular Changes as Biomarkers of Systemic Cardiovascular Diseases 187

7.3Retinal Vascular Imaging

7.3.1Assessment of retinal vascular signs from retinal photographs

Modern digital imaging systems have revolutionized the assessment of retinal photographs. Using these newly developed quantitative methods, several studies have demonstrated excellent reproducibility for the detection of well-defined retinopathy signs (kappa values have ranged from 0.80–0.99 for microaneurysms and retinal hemorrhages) and good reproducibility for more subtle retinal arteriolar lesions (0.40–0.79 for focal retinal arteriolar narrowing and arteriovenous nicking) [49; 50].

In addition, objective measurement of historically subjective retinal vascular changes, such as generalized retinal arteriolar narrowing, is now possible with the semi-automated retinal image-analysis software. Parr, Hubbard, and their associates developed formulae to generate summarized measures of retinal arteriolar (central retinal arteriolar equivalent [CRAE]) and venular (central retinal venular equivalent [CRVE]) diameters, as well as their dimensionless quotient (arteriovenous ratio [AVR]) [49; 54; 55]. These retinal vascular indices have been used in several large-scale epidemiological studies [49; 50; 52; 53; 56], which demonstrated substantial reproducibility for these vessel measurements (intra-class correlation coefficient ranged from 0.80–0.99), providing further evidence that retinal photography offers a more sensitive and precise means of assessing architectural changes in the retinal vascular network.

Based on these new standardized assessments of photographs, retinal vascular signs have been found to be fairly common in adult people 40 years and older, even in those without a history of diabetes or hypertension. Prevalence [40; 43; 57] and incidence [41; 58] rates ranging from 2 to 15% have been reported for various retinal vascular lesions.

Other emerging computer imaging systems have features such as automated detection of optic disc, identification and tracking of arterioles and venules, batch processing of retinal images (i.e., analysis of multiple images at a time), and measurement of retinal vessel diameters with greater precision and reproducibility [52; 59]. Whether these systems are superior to existing techniques in cardiovascular risk prediction remains to be seen.

7.3.2Limitations in current retinal vascular imaging techniques

While retinal image analysis posts exciting possibilities [48], its applicability in clinical settings is yet to be established, partly due to a number of methodological issues concerning its use.

First, the formulae utilized to combine individual retinal vascular diameters into summarized indices are based on theoretical and empiric models. The Parr [54; 55] and Hubbard [49] formulae for CRAE and CRVE were derived from examination of a large number of retinal images with branching points, calculating the relationship between individual trunk vessels and their respective branch vessels using a root mean square deviation model that best fit the observed data. Although used

widely in many epidemiological studies of cardiovascular and ocular diseases, there are some drawbacks in using these formulae. Knudtson and associates made an important observation that the Parr-Hubbard formulae were dependent on the number of retinal vessels measured [60]. In addition, since the formulae contained constants within the equations, they were also dependent on the units with which the vessels were measured. Knudtson and colleagues therefore developed a modified formula for summarizing retinal vascular caliber, and demonstrated clear superiority of their formula over the previously used Parr-Hubbard formulae [60]. It is likely that there will be further refinement. Recently, for example, some investigators suggested a revised formula for more accurate estimation of arteriolar branch coefficient [61]. This formula used a linear regression model that incorporated a relationship to the asymmetry index of the vessel branches being measured. However, whether this or other newer formulae can indeed improve the predictive ability to detect associations with systemic cardiovascular diseases is unclear.

Second, existing retinal vascular research has largely focused on differences in retinal vascular changes between groups of people (e.g., people with smaller retinal arteriolar diameter are more likely to develop cardiovascular disease than people with relatively larger arteriolar diameter). To allow the use of retinal vessel measurement as a potential risk stratification tool in a clinic setting, retinal image analysis must produce results that enable an assessment of absolute risk in individual patients. The measurement of absolute retinal vascular caliber, for example, is critical to this development [53; 62]. This requires addressing the issue of magnification effect from retinal photography, either by incorporating an adjusted measurement to compensate for this effect or using dimensionless measurements. While there are already a few methods to adjust for magnification using ocular biometric data (e.g., axial length) [48], most were designed for telecentric cameras. For nontelecentric cameras, Rudnicka and colleagues have described a method to adjust for magnification using plan films [63], but its applicability on digitized retinal photographs is unknown.

To account for magnification effects and allow for comparison of measurements of retinal topographical changes between individuals, studies have sought alternative geometric attributes of retinal vasculature that are dimensionless in nature. These include the retinal AVR, junctional exponents, vascular bifurcation angles, vascular tortuosity, and length-to-diameter ratio [48]. Among these, the AVR has been the most commonly used measure. It is important to note, however, that the AVR has significant limitations, including the inability to capture separately the information of the individual arteriolar and venular caliber component measurements [64; 65]. For example, both narrower arterioles and wider venules may produce a smaller AVR. Thus, AVR cannot differentiate between changes in arteriolar and venular caliber. There is increasing evidence that this differentiation is important, as different systemic diseases appear to be associated with specific caliber changes in arterioles and venules. While smaller retinal arteriolar caliber is associated with hypertension, and may even precede clinical hypertension development, larger retinal venular caliber has been associated with inflammation, smoking, hyperglycemia, obesity and dyslipidemia [56]. These observations suggest that changes in retinal arteriolar and venular caliber may reflect different pathophysiological processes underlying the as-

Retinal Vascular Changes as Biomarkers of Systemic Cardiovascular Diseases 189

sociated systemic diseases. Combining these two components into one estimate, the AVR, without consideration of separate arteriolar or venular caliber measurements, therefore masks these associations.

Third, researchers have recently discovered another important concept in retinal vascular imaging analysis: the need to adjust for retinal arteriolar caliber in analysis of retinal venular caliber, and vice versa [66; 67]. Liew and colleagues showed that the high correlation between retinal arteriolar and venular caliber means that individuals with narrower arterioles are more likely to have narrower venules [66; 67]. As a result, the confounding effect of arteriolar and venular caliber for individual measurements should be taken into consideration. This is clearly demonstrated in the Rotterdam Study, in which a counterintuitive association between retinal venular narrowing and hypertension was initially reported [68], but after the use of the new analytical approach, modeling retinal arteriolar and venular calibers simultaneously, retinal venular narrowing was shown to have no association with hypertension [69].

Finally, despite the vast amount of data on retinal vessel measures in numerous population-based studies, there is a lack of knowledge about the normative data for these measurements. Defining what is normal and abnormal is crucial for development of a clinical tool. One of the challenges in deriving normative data using studies in the adult population was that it was difficult to completely control for the confounding effect of systemic (e.g., hypertension, diabetes, smoking, medications) and ocular (e.g., diabetic retinopathy, glaucoma) disease processes on retinal vessel measurements. Studying retinal vascular caliber in healthy children, who are generally free of these influences, may provide a better understanding of the reference data for this important vascular variable [70]. Several studies of retinal vascular measurements in children are currently underway.

7.4Retinal Vascular Changes and Cardiovascular Disease

One of the major advances in retinal vascular imaging research in the last decade has been the clear demonstration that physiological and pathological alterations in the retinal vascular network are associated with a variety of cardiovascular diseases, including hypertension, stroke, coronary heart disease (CHD) and congestive heart failure (Table 7.1).

7.4.1Hypertension

It has long been known that hypertension exerts a profound effect on the retinal vasculature. The association of retinal vascular changes with blood pressure, in particular, is strong, graded, and consistently seen in both adult [40; 41; 43; 56; 57; 71; 72; 74; 96–98] and child populations [91].

The Beaver Dam Eye Study in Wisconsin has reported increases in both prevalence [40] and 5-year incidence [41] of various retinal arteriolar abnormalities in hypertensive individuals compared to those without hypertension. Among these,

CHD = coronary heart disease; AVR = arteriole-to-venule ratio; WESDR = Wisconsin Epidemiological Study of Diabetic Retinopathy.

*Odds ratio or relative risk < 1:5 (+), 1.5–2.0 (++), > 2:0 (+++) and quantified using other measures (–).

generalized retinal arteriolar narrowing has long been regarded as an early characteristic sign of hypertension [12; 44; 45]. Using a new semi-automated computerbased image-analytical technique, the Atherosclerosis Risk in Communities (ARIC) study in the United States reported that retinal arteriolar diameter is strongly and inversely related to higher blood pressure levels [71], a finding subsequently confirmed in four other population-based studies [68; 73; 74; 96]. In the Beaver Dam Eye Study, each 10 mmHg increase in mean arterial blood pressure was associated with a 6 mm (or 3%) decrease in retinal arteriolar diameter, even after adjusting for age, gender, diabetes, smoking, and other vascular risk factors [73]. While these data support the strong link between generalized arteriolar narrowing and hypertension, the subtle degree of arteriolar narrowing also suggests that a clinical examination based on ophthalmoscopy is unlikely to be capable of detecting such small changes. While earlier studies have predominantly used smaller AVR as the only measure of generalized retinal arteriolar narrowing, subsequent studies in both adults [67; 92] and children [91] evaluating retinal arteriolar and venular calibers separately have validated the strong association of hypertension with arteriolar narrowing.

In the last few years, researchers have attempted to answer a key research question: are retinal vascular changes markers of cumulative, long-term blood pressure damage, or do they only reflect a transient effect of acutely elevated blood

Retinal Vascular Changes as Biomarkers of Systemic Cardiovascular Diseases 191

pressure? Several studies addressed this question by analyzing the association of specific retinopathy signs with both concurrent and past blood pressure levels [71; 74; 86]. These studies found that generalized arteriolar narrowing and AV nicking were independently related to past blood pressure levels, indicating that these signs may reflect persistent structural arteriolar changes from long-term hypertension [71; 74; 86]. In contrast, studies show that focal arteriolar narrowing, retinal hemorrhages, microaneurysms, and cotton wool spots were related to only concurrent but not past blood pressure levels, and therefore may be related more to fleeting, possibly physiological, changes caused by acute blood pressure elevation [71; 74].

Prospective data from four population-based studies have provided new understanding into the relation of generalized retinal arteriolar narrowing to subsequent development of systemic hypertension [75; 88; 89; 92]. In the ARIC study, normotensive participants who had generalized retinal arteriolar narrowing at baseline were 60% more likely to be diagnosed with hypertension over a three year period than individuals without arteriolar narrowing (relative risk 1.62, 95% CI, 1.21 to 2.18) [75]. The severity of arteriolar narrowing also appeared to correlate positively with the degree of change in blood pressure, independent of preexisting blood pressure, body mass index, and other known hypertension risk factors. Similar results were generated from the Beaver Dam, Rotterdam, and the Blue Mountains Eye studies [88; 89; 92], providing strong evidence that generalized arteriolar narrowing, possibly mirroring similar peripheral arteriolar changes elsewhere in the body, is a preclinical marker of hypertension. These observations support the concept that the microcirculation is critical to the pathogenesis of hypertension [99].

There is increasing evidence, albeit from a limited number of clinical reports, that hypertensive retinopathy signs may regress with improved blood pressure control [100–103]. For example, it has been reported that hypertensive retinopathy caused by accelerated hypertension can improve after normalization of blood pressure [100]. In addition, in a clinical study of 51 hypertensive patients, Pose-Reino and colleagues presented objective and quantitative evidence of generalized retinal arteriolar narrowing regression associated with six months of hypertension treatment with losartan, an angiotensin II inhibitor [103]. In another small study of mildly hypertensive patients randomized to treatment with enalapril (an angiotensin converting enzyme inhibitor) or hydrocholorthiazide, retinal arteriolar wall opacification (although not other retinopathy signs) was significantly reduced after 26 weeks of enalapril treatment while hydrochlorothiazide did not seem to have any effect on the retinopathy signs [101]. As yet, however, due to the lack of controlled clinical trials, the important clinical question of whether regression of hypertensive retinopathy signs is also associated with a concurrent reduction in a person’s cardiovascular risk remains unanswered.

7.4.2Stroke and cerebrovascular disease

The retinal and cerebral vasculature share similar embryological origin, anatomical features, and physiological properties [104]. This concept provides strong biological rationale for the use of retinal image analysis to indirectly study the cerebral microvasculature and related diseases. In support of this theory is the strong and