Accuracy of fundus autofluorescence imaging for the diagnosis and monitoring of retinal conditions: a systematic review

Age-related macular degeneration

Age-related macular degeneration is categorised as either early AMD or late AMD. In early AMD the vision is initially unaffected, but there are risk features for late AMD. Late AMD is classified into two types: wet AMD, which is also known as exudative or neovascular AMD (the term neovascular is used in this report); and dry AMD, which, in advanced cases, is also known as GA. AMD is the commonest cause of visual impairment, and results in progressive, irreversible damage to the macula, the retinal pigment epithelium and retinal photoreceptors. 4 Clinical and phenotypic variations have been identified, and increasing age and smoking are risk factors for the condition. 5 Although AMD usually affects both eyes, it rarely results in complete blindness, as the peripheral vision is not usually affected.

Diabetic retinopathy

Diabetic retinopathy is the leading cause of vision loss in people aged under 50 years in the developed world. Its prevalence is increasing,19 and it is the commonest cause of blindness certification in working age adults. 2 Diabetic retinopathy is a term that describes any pathological features that occur at the retina due to diabetes. This can range from minimal non-sight-threatening vascular features such as microaneurysms, to proliferative retinopathy, which is the growth of delicate new blood vessels that can bleed easily and may result in intraocular haemorrhage with sudden loss of vision. 20,21 Proliferative retinopathy may be transient if the haemorrhage clears, but may have features of advanced disease with an associated retinal detachment. Other types of diabetic retinopathy include fluid collecting at the macula (diabetic macular oedema) or loss of blood vessels at the macula, known as macular ischaemia, which can be associated with severe central visual loss depending on the extent of the loss of vascular perfusion. 22

The most common symptom of diabetic macular oedema is blurred vision. Other symptoms include metamorphopsia (distortion of the visual image); floaters (moving spots seen in the field of view); loss of contrast sensitivity; photosensitivity (intolerance to light); changes in colour vision; scotomas (where part of the visual field is missing); vision becoming blurred, hazy or fluctuating, or with the appearance of black lines or dots; periods of temporary ‘blackness’ caused by retinal haemorrhage; poor peripheral vision; and poor depth perception. 23 People with diabetic macular oedema may also be affected by a difference in sight between their two eyes. 22 Laser treatment of macular disease or proliferative retinopathy can also lead to visual symptoms including field loss and photosensitivity.

Estimates of UK prevalence of diabetic retinopathy vary owing to the range of different sources of data. Based on data from 2010, there is an estimated prevalence of 748,000 cases of early diabetic retinopathy in the UK, 85,000 cases of more advanced retinopathy and 188,000 cases of diabetic maculopathy. 3 In a study of people with diabetes in England (which also estimated prevalence data for 2010), 7.12% had diabetic macular oedema in one or both eyes, of which 39% had clinically significant diabetic macular oedema. 24

Central serous chorioretinopathy

Central serous chorioretinopathy is the fourth most common retinopathy after AMD, diabetic retinopathy and retinal vein occlusion. 25 It is characterised by fluid accumulating as a result of dysfunction of retinal pigment epithelial cells and/or the choroid. Usually, if the patient experiences symptoms, the fluid will have accumulated under the central macula. 26 Spontaneous visual recovery can occur within 2 or 3 months in the acute type,25 but chronic CSC, which is more common in older people, can lead to persistent or episodic blurred vision. In some cases of chronic CSC, disease progression is more severe, leading to retinal pigment epithelial atrophy,27 which may result in permanent visual loss. 25 There is some evidence that risk factors for CSC include smoking, hypertension or sympathetic responses (nervous system activation in response to stressors),27 steroid medications28 or genetic susceptibility. Symptoms of the condition include image size disparity between the eyes, blurred central vision, altered colour vision and, in chronic disease with atrophic change, a loss of visual acuity. 25

Men are more often affected by CSC than women, in a ratio of up to 8 : 1. 26 Although predominantly an adult disease, some cases of CSC have been reported in children. 26 People aged over 50 years are more likely to have the condition in both eyes, may develop choroidal neovascularisation25 and may have coincidental polypoidal choroidopathy. CSC recurs in about one-third of patients within around 1 year. 25 The peak incidence is around 40–45 years in men, but later in women, at about 51 years. 26 Reliable incidence and prevalence data are not available for the UK. According to a study in the USA (based on data for 1980–2002), the incidence of CSC is about 9.9 per 100,000 in men and 1.7 per 100,000 in women. 25

Inherited retinal dystrophies

Inherited retinal dystrophies are a broad group of genetically determined disorders that affect the retina and lead to progressive visual loss. These disorders can be generally classified according to whether they affect primarily the centre of the vision, the periphery or both. Further classification depends on whether rods or cones are primarily affected, whether or not both central and peripheral systems are involved and which is more affected (referred to as rod–cone or cone–rod). 29 Macular dystrophies may only involve central vision without affecting the visual fields, and the rods and cones may not be affected. However, progression may occur so that, for example, a disorder that was classified as macular dystrophy may, in time, develop cone and rod dysfunction. Terms used to describe these conditions depend on the appearance of the retina, which cells are involved, and (more recently) if the gene mutation(s) is/are known. Retinal dystrophies include, among others, retinitis pigmentosa, macular dystrophy, Stargardt disease (which primarily develops in childhood and adolescence) and Best disease.

The age of onset of retinal dystrophies varies, with early and late onset forms having different inheritance patterns, and mutations in over 200 different genes are known to be involved in different types, including syndromic retinal dystrophies. 30 Symptoms and disease progression also vary within each named condition. Symptoms may be apparent from birth or may appear at any age, depending on the type and the specific genetic variant inherited. The degree of central visual loss, peripheral field loss and difficulty with seeing in the dark depends on how well each group of photoreceptor cells is working, and the speed of sight loss varies with different genetic forms and between individuals. 31 In cases of severe rod–cone dystrophy with early onset, macular involvement may lead to early decline of visual acuity, and these cases may be more challenging to diagnose. 29

Although some inherited retinal dystrophies affect few people, overall, these disorders are stated as being a cause in 15.8% of certified blindness registrations and 10% of partial sight registrations in England and Wales, and are considered to be a significant burden in the working age population. 2 The annual incidence of hereditary retinal disorders in the UK has been estimated at around 1500 cases, and about 100 children and 480 adults of working age in the UK are registered as blind or partially sighted as a result of these conditions. 2 A study published in 201232 estimated that (based on data from 2006 to 2008) the annual UK incidence of childhood-onset retinal dystrophies, is 1.4 in 100,000 children (aged 0–15 years) and the cumulative incidence by age 16 years is 22.3 in 100,000 children.

Cystoid macular oedema

Macular oedema is the accumulation of fluid within the retina at the macular area, secondary to various retinal conditions. Depending on its cause, acute oedema may resolve spontaneously. 33 Cystoid macular oedema is the result of chronic oedema that persists for ≥ 4 months, leading to the formation of cystic honeycomb-like spaces in the retina,34 and can occur as a consequence of retinal dystrophies,35 inflammatory diseases (uveitis, scleritis, birdshot chorioretinopathy, toxoplasmosis),36 retinal vascular disease (retinal vein occlusions, idiopathic retinal telangiectasia, radiation retinopathy),34 diabetic retinopathy,34 cataract or other eye surgery, injury to the eye, choroidal tumours, or may be drug induced, for example with prostaglandins such as latanoprost. 37 As with other macular conditions, the main symptoms of cystoid macular oedema are blurred or decreased central vision, but peripheral vision is unaffected. Estimates of the incidence and prevalence of cystoid macular oedema vary considerably depending on the cause of the oedema. For example, the incidence of clinically significant cystoid macular oedema caused by uncomplicated cataract surgery has been estimated across studies as 0.6–2.6%,38 whereas the prevalence of cystoid macular oedema in patients with retinitis pigmentosa may be as high as 32–49% (unilateral) or 18–44% (bilateral). 39,40

Fundus autofluorescence in relation to retinal health

The fundus of the human eye has a ‘natural’ or ‘background’ level of fluorescence, which is referred to as autofluorescence. This is caused by the presence of molecules with fluorescent properties (i.e. molecules that, when exposed to light of an appropriate wavelength, absorb the incident electromagnetic energy and re-emit this as light at wavelengths longer than those of the initial source). The fluorescent molecules, known as fluorophores, are potentially of clinical value in detecting age- or disease-related processes, since their density and distribution alters with ageing of the eye and with certain pathological conditions. Notably, lipofuscin is a fluorescent pigment that accumulates in the retinal pigment epithelium as a by-product of cell metabolism and can lead to the development of drusen. Lipofuscin deposition normally increases with age but its accumulation may also reflect cell dysfunction or metabolic abnormalities in the retinal pigment epithelium. Excessive lipofuscin deposition in the retinal pigment epithelium is considered pathologic and is associated with visual loss. 50 FAF imaging techniques have potential value as diagnostic tools, since the presence and distribution of fluorophores may correlate with disease activity and may provide an early indication of future disease development, progression or response to treatment. FAF imaging is also potentially valuable in detecting areas of reduced or absent autofluorescence (hypoautofluorescence), which can arise either through atrophy of retinal structures or blocking of FAF reflectance (e.g. by blood vessels).

Fundus autofluorescence imaging techniques

The intensity of light emitted during FAF is relatively weak (about two orders of magnitude lower than the background of a fluorescein angiogram at peak dye transit) and so specialist imaging techniques are required to enable FAF to be detected and mapped. FAF has been an area of interest in ophthalmic research for more than 40 years but it has only recently become clinically relevant, as a result of technological advances. 50 As outlined below, there are three main methods for assessing FAF: using a fundus camera, fundus spectrophotometry or confocal scanning laser ophthalmoscopy (cSLO).

The fundus camera uses a single flash to image the entire retinal area instantaneously. The acquired autofluorescence image is derived from all tissues in the light beam with fluorescent properties; light scattered anterior and posterior to the plane of interest can greatly influence the detected signal.

Fundus spectrophotometry was developed to measure FAF from small retinal areas (2 degrees of visual angular diameter). It incorporates an image intensifier diode array as a detector and beam separation in the pupil to minimise the contribution of autofluorescence from the crystalline lens. 51

In cSLO a focused low-power laser beam is swept across the fundus in a raster pattern (horizontal and vertical scanning across a grid) to provide the excitatory light source for fluorophores. 52 Different excitation wavelengths can be generated, depending on the type of laser used, including 488 nm (blue light) with a solid-state laser and 514 nm (green light) with an argon-ion laser. For near-infrared fundus autofluorescence (NIR-FAF), the excitation wavelength is 790 nm. The confocal nature of the optics reduces the detection of autofluorescence from structures anterior to the retina, such as the lens and cornea. Unlike fundus spectrophotometry, cSLO allows imaging of FAF over larger retinal areas [e.g. 55 degrees in Heidelberg Retina Angiograph-based systems (Heidelberg Engineering, Heidelberg, Germany)]. To reduce background noise and enhance image contrast, the mean image of several FAF images is obtained (usually based on up to 20 frames, after adjustments to correct for eye movement). In order to block the reflected light but permit autofluorescence light to pass, cSLO have barrier filters of 500 nm, 525 nm and 800 nm for excitation wavelengths of 488 nm, 514 nm and 790 nm, respectively. As well as the excitation and barrier filter wavelengths, image contrast and brightness settings also differ among cSLO devices and these differences must be taken into account when comparing the results of FAF imaging obtained from different cSLO devices. 16

Confocal scanning laser ophthalmoscopy is the most sensitive imaging approach for identifying autofluorescence that arises specifically from the fundus (i.e. minimising the detection of autofluorescence arising from other parts of the eye such as the lens). 16 According to clinical experts advising the review, cSLO is the current standard method employed for obtaining FAF images of retinal conditions, and is therefore the only type of scanning laser ophthalmoscopy permitted as an index test for assessing FAF in the current review.

Fundus autofluorescence for diagnosis and monitoring of disease

The quantitative accuracy of FAF imaging for the diagnosis and/or monitoring of retinal conditions (i.e. its sensitivity and specificity) is unclear. However, FAF imaging is considered helpful in a number of conditions to help establish a diagnosis and monitor treatment without the need for angiography,1 is relatively easy to accomplish and requires little time. 53 Studies have suggested the potential diagnostic value of FAF imaging as a more sensitive marker of retinal pathology than existing examinations alone, for example, in GA,54 choroidal neovascularisation development in AMD,55 retinal pigment epithelium alterations,56 Best disease,57 cystoid macular oedema58 and CSC. 59 Clinical advisors to the current review also suggested that FAF is useful for diagnosing any type of inherited retinal dystrophy, and has the potential to show pathologic features earlier than on fundoscopy. FAF imaging also appears promising for monitoring changes in a number of retinal conditions, either natural progression or responses to therapy. FAF imaging has been used, for example, for monitoring changes in GA;60–66 retinal pigment epithelium tear or loss;67–69 retinitis pigmentosa;70,71 Stargardt disease,72 CSC,73 diabetic macular oedema74 and retinal vasculitis. 75 Increasingly, FAF has been specified as an end point in clinical research studies of therapies for retinal conditions, for example, with antivascular endothelial growth factor drugs,68,76–78 sirolimus,65 lampalizumab,66 finasteride,73 photodynamic therapy79 and vitrectomy. 80

Number and type of studies included

The eight included research papers describe eight unique primary studies. 83,84,96–101 Full details of these studies, including our assessment of their methodological rigour, are given in the data extraction forms (see Appendix 5).

Characteristics of the included studies

Key characteristics of the included studies are summarised in Table 2.

Although FAF imaging may be used to detect autofluorescence in a wide range of retinal diseases (see Chapter 1), the eight included studies83,84,96–101 cover only three broad retinal conditions. These are AMD (four studies96,97,99,101) cystoid macular oedema (two studies83,100) and diabetic macular oedema (two studies84,98). Of four studies on AMD, three studies specifically investigated FAF imaging in the detection of reticular pseudodrusen96,97,101 and one study investigated FAF imaging in the diagnosis of choroidal neovascularisation. 99

The eight studies included in the systematic review provide information on diagnostic accuracy only (see Table 2). No studies of the use of FAF imaging for monitoring the natural progression of retinal conditions, or for monitoring the response of retinal conditions to therapy, met the inclusion criteria.

Four of the studies involved the analysis of retrospective data,84,97,100,101 two involved prospective recruitment of patients,96,99 one appeared to prospectively recruit patients, although this was not explicitly reported,83 and it is unclear whether the remaining study was retrospective or prospective98 (see Table 2). Four of the studies selected patients consecutively (i.e. in the chronological order in which they first presented to clinicians)84,97,98,100 and the remaining studies did not report whether selection was consecutive. 83,96,99,101 Retrospective studies, and those lacking consecutive patient recruitment, may be at greater risk of selection bias, and this is considered further below in relation to the overall assessment of the quality of evidence (see Quality of research available).

Most of the studies were conducted in Europe. The study by McBain and colleagues100 was exclusively in the UK and the study by Waldstein and colleagues84 also appears to have been conducted in the UK, although this was not explicitly reported. Two studies were conducted in multiple countries, which included the UK. Hogg and colleagues96 conducted studies in Italy, Portugal and the UK, and Smith and colleagues101 in the USA and UK. The remaining studies were conducted in Italy,98 Portugal,99 Turkey83 and Japan. 97 The studies were published between 2006 and 2014. The earliest reported participant recruitment was in 2004100 and the latest in November 201084,97 but dates of recruitment were not reported in three studies. 96,99,101

Both men and women were included in the studies. Where reported (in seven studies83,84,96–100), the proportion of men ranged from 50% to 69%. Age was not reported in one study on AMD,101 but in the remaining studies participants were older in the AMD studies (mean age 74–76 years)96,97,99 than in the studies on cystoid macular oedema (mean age 59–62 years)83,100 and diabetic macular oedema (mean age 49–67 years)84,98 (see Table 3).

As indicated in Table 3, three of the four studies on AMD excluded participants with ocular comorbidities,96,97,99 but this was not reported in the fourth study. 101 One study on cystoid macular oedema excluded participants with ocular comorbidities,83 while the other did not report exclusion criteria. 100 For diabetic macular oedema, one study excluded patients with significant media opacities98 while the other excluded patients with any macular comorbidity. 84

In all studies the unit of analysis was the eye, and the number of eyes included ranged from 34 to 263 (see Table 3). As shown in Table 4, the studies differed according to whether they included one eye per patient,96,99,100 both eyes,97 or a mixture of single eyes and both eyes. 83,84,98,101 In the study on AMD by Smith and colleagues,101 the results of two case series are reported (obtained from two databases – one in the USA and one in the UK) each with different inclusion criteria, one of which included both eyes, while the other included single eyes. However, the data on test accuracy from the study by Smith and colleagues101 relevant to the current review are only available for the pooled population.

Descriptions of the retinal conditions varied considerably in detail and the studies were generally heterogeneous in terms of the conditions they included (see Table 4). The three studies on reticular pseudodrusen in AMD96,97,101 differed according to the stage of AMD and whether it affected one or both eyes. The two studies on cystoid macular oedema83,100 differed in the primary conditions from which the cystoid macular oedema developed. The studies on diabetic macular oedema84,98 differed on whether they included ocular comorbidities.

In keeping with the inclusion criteria for the current review, all FAF imaging tests were conducted using confocal scanning ophthalmoscopy (Table 5). The studies (except one that did not specify the image acquisition equipment101) used variants of the Heidelberg Retina Angiograph (Heidelberg Engineering, Heidelberg, Germany). Where reported (six studies83,84,97–100), the excitation wavelength was 488 nm (blue) and, where reported (five studies83,97–100), the detection wavelength was > 500 nm. Two of the studies included an additional FAF imaging test with a higher excitation wavelength; these were at 514 nm (green) in a study of diabetic macular oedema84 and at 790 nm (near infrared) in a study of reticular pseudodrusen. 97 The field of view, which was specified in five studies,83,84,96,99,100 was 30 degrees and the number of frames from which the final (mean) FAF image was calculated (specified in all studies except one101) ranged from 9 to 20.

The reference standards reported in the primary studies were: fluorescein angiography for diagnosing choroidal neovascularistion99 and cystoid macular oedema;83,100 colour fundus photography101 or multiple imaging modalities96,97 for diagnosing reticular pseudodrusen; and either retromode scanning laser ophthalmology98 or spectral-domain optical coherence tomography84 for diagnosing diabetic macular oedema (see Table 5). In the two studies that employed multiple imaging modalities,96,97 the reference standard is somewhat unusual since diagnosis required a positive result on a specified minimum number of modalities. The clinical relevance of the reference standards is discussed further in our assessment of study quality (see Quality of research available).

Six of the studies reported that patients underwent an ophthalmic examination, which included assessments of visual acuity, intraocular pressure and/or slit lamp biomicroscopy83,84,96–99 but they did not specify whether or not information obtained from these examinations was necessary to assist any of the diagnostic decisions made using the index test or reference standard. Three studies employed further retinal imaging tests in addition to their index test and reference standard83,98,99 (these are listed as ‘comparators’ in Appendix 5). Two of these, by Cachulo and colleagues99 on choroidal neovascularisation (where the additional tests were colour fundus photography, indocyanine green angiography, optical coherence tomography and a retinal leakage analysis) and by Dinc and colleagues83 on cystoid macular oedema (where the additional test was optical coherence tomography), did not explain whether or not any information from the additional tests was taken into account when making diagnostic decisions using FAF imaging or the reference standard (fluorescein angiography). The third study, by Vujosevic and colleagues98 on diabetic macular oedema (where the additional tests were time-domain optical coherence tomography and fluorescein angiography), mentioned that images were graded in a masked fashion, suggesting that the additional tests would not have influenced any diagnostic decisions made using FAF imaging or the reference standard (retromode scanning laser ophthalmoscopy). One of the studies on diabetic macular oedema which employed both 488 nm and 514 nm wavelengths of FAF imaging used the results of both tests to calculate macular pigment optical density maps. 84 This approach was reported to have lower sensitivity and specificity for detecting diabetic macular oedema than either of the individual FAF imaging tests on which it was based (see Appendix 5) and as such is not considered as a separate index test in the current report.

The diagnostic criteria on FAF imaging for each of the retinal conditions were qualitative (i.e. descriptive) in all studies (see Table 5), and only one of the studies reported an objective (quantitative) approach for determining how abnormal (hypo or hyper) autofluorescence was defined. 101 The three studies on reticular pseudodrusen all specified that the diagnostic criterion was a reticular pattern of hypoautofluorescence, although there were slight differences in how this was described in each study. 96,97,101 The two studies on cystoid macular oedema83,100 both appeared to diagnose the condition as round or oval patterns of autofluorescence at the fovea, although this was not stated explicitly in one study. 83 The two studies on diabetic macular oedema84,98 only mentioned the diagnostic criteria briefly and differed in how these were described. Despite the subjective nature of the diagnostic criteria, only three of the eight studies investigated intergrader agreement in the classification of retinal images (see Table 2). 84,97,98 Limited information was provided about the expertise of the image graders, who were described as two independent ophthalmologists,97 two retinal specialists trained in image grading98 and two independent graders. 84

Quality assessment of diagnostic accuracy studies question 1: applicability of the patient spectrum

The populations included in the primary studies were considered unlikely to be representative of those who would present in NHS clinical practice (the patient spectrum was considered ‘probably not representative’ in seven studies83,84,96,97,99,100,101 and unclear in one study98) (see Table 6). In the studies by Cachulo and colleagues,99 Hogg and colleagues96 and Smith and colleagues101 the case mix was considered atypical and unlikely to represent patients in the NHS as the retinal condition was required to differ between the two eyes. In the studies by McBain and colleagues,100 Dinc and colleagues,83 Ueda-Arakawa and colleagues97 and Waldstein and colleagues84 the case mix also appears to have limited relevance to current NHS practice, as McBain and colleagues100 and Dinc and colleagues83 required cystoid macular oedema to be specific to certain causal conditions, while Ueda-Arakawa and colleagues97 and Waldstein and colleagues84 excluded patients with ocular comorbidities. In addition, the studies by McBain and colleagues,100 Ueda-Arakawa and colleagues97 and Waldstein and colleagues84 selected patients retrospectively, which may further limit their applicability to current practice (as the case mix is potentially selective). 102 The study by Vusojevic and colleagues98 included patients with any stage of treated or untreated diabetic retinopathy and permitted ocular comorbidities and, as such, appears potentially relevant to the NHS; however, the study was conducted in Italy and it is unclear whether patients were selected prospectively or retrospectively.

Quality assessment of diagnostic accuracy studies question 2: applicability of reference standards

In five of the studies83,84,99–101 the reference standards employed for diagnosing the target retinal condition were judged to be appropriate in principle as they are accepted standard methods: fluorescein angiography was used for diagnosing choroidal neovascularisation99 and cystoid macular oedema;83,100 colour fundus photography was used for diagnosing reticular pseudodrusen;101 and spectral-domain optical coherence tomography was used for diagnosing diabetic macular oedema. 84 However, there are caveats around this interpretation because clinical experts advised that it is unlikely that individual imaging tests would be used in isolation, as appeared to be the case in these primary studies. For instance, fluorescein angiography would typically be employed in conjunction with optical coherence tomography when diagnosing cystoid macular oedema and diabetic macular oedema. 34,38 For the detection of reticular pseudodrusen, although colour fundus photography has been commonly used and is often considered a standard approach, this may not capture all the reticular pseudodrusen that could be identifiable using other imaging methods. 11,103 As noted in Table 6, for three studies we judged the ability of the reference standard to diagnose the target retinal condition as unclear. 96–98 This was either because the studies employed a suite of imaging modalities and required a positive result on a specified minimum number of these96,97 (i.e. a non-standard diagnostic approach) or because the reference standard employed in the primary study (retromode scanning laser ophthalmology) has not been previously explored as a diagnostic tool for the specified condition (diabetic macular oedema). 98 As such, these three studies would not be reflective of standard imaging approaches employed in the NHS.

Quality assessment of diagnostic accuracy studies question 3: test interval

The time period between the index and reference tests (QUADAS question 3) was unclear (not reported) in five of the studies83,96,97,99,101 and hence the risk of disease progression bias in these studies is unclear. In one study on cystoid macular oedema100 and both studies on diabetic macular oedema84,98 the time interval was considered short enough that both tests would have been assessing the same disease state and so these studies would be considered at low risk of disease progression bias.

Quality assessment of diagnostic accuracy studies question 4: population receiving the reference standard

In all studies except one96 all the patients included in each study received the reference standard. In the study by Hogg and colleagues96 on reticular pseudodrusen, multiple imaging methods were used, but it is unclear, according to the reported sample sizes, whether all participants received all tests and some data were then excluded, or whether some participants did not receive all tests.

Quality assessment of diagnostic accuracy studies question 5: same reference standard for all patients

In six studies all the patients included in each study received the same reference standard. 83,84,98–101 In the two studies where this was not the case, conducted by Hogg and colleagues96 and Ueda-Arakawa and colleagues,97 the combination of imaging modalities that contributed to a diagnosis (at least one of five tests96 or at least two of seven tests97) could have differed between the patients.

Quality assessment of diagnostic accuracy studies question 6: independence of index test and reference standard

In six studies the index text was independent of the reference standard. 83,84,98–101 In the two studies where this was not the case, conducted by Hogg and colleagues96 and Ueda-Arakawa and colleagues,97 the index tests were among five tests96 or seven tests97 that made up the reference standard.

False positives are not possible where FAF imaging is also part of the reference standard and, as would be expected, no false positives were reported for FAF imaging by Hogg and colleagues. 96 However, Ueda-Arakawa and colleagues did report false positives (9/217 for FAF imaging and 5/136 for near-infrared FAF imaging). These false positives appear to be an artefact of the diagnostic criterion for reticular pseudodrusen, which required a positive result on two or more of seven imaging modalities. The ‘false positives’ in this case indicate that, in nine and five eyes, FAF imaging and near-infrared FAF imaging, respectively, were the only tests out of seven in the reference standard that detected reticular pseudodrusen.

Quality assessment of diagnostic accuracy studies question 7: interpretation of reference standard independent of index test results

In all studies except one98 it was unclear (not reported) whether or not investigators interpreting the results of the reference standard test might have been aware of the results of the index test. In the study by Vusojevic and colleagues98 the order of tests was not reported but it was stated that images were independently graded in a masked fashion.

Quality assessment of diagnostic accuracy studies question 8: interpretation of index test independent of reference standard results

In six studies it was unclear (not reported) whether or not investigators interpreting the results of the index test might have been aware of the results of the reference standard. 84,96,97,99–101 In the study by Dinc and colleagues83 interpretation of the index test was independent of the reference standard as the index test was conducted first. In the study by Vusojevic and colleagues,98 the order of tests was not reported but it was stated that images were independently graded in a masked fashion.

Quality assessment of diagnostic accuracy studies question 9: clinical information available for test interpretation

None of the studies explicitly stated whether or not their index tests or reference standards were interpreted in conjunction with other clinical information (i.e. the results from standard ophthalmic examinations or other imaging tests) and this casts considerable doubt over whether or not any of the reference standard tests as employed in the primary studies would properly represent how the tests would be employed in clinical practice.

Quality assessment of diagnostic accuracy studies question 10: reporting of uninterpretable or intermediate test results

Five of the studies (two on reticular pseudodrusen,96,101 two on diabetic macular oedema84,98 and one on cystoid macular oedema83) did not report whether uninterpretable or intermediate test results were considered either in the calculation of diagnostic accuracy or as exclusion criteria, so it is unclear whether or not image quality might have limited the utility of any of the imaging tests in these studies. Three studies reported poor quality or uninterpretable images. 97,99,100 Cachulo and colleagues99 mentioned that the pattern of autofluorescence could not be determined for two eyes (4%) because of poor-quality images, but it is unclear whether or not these were included in the diagnostic accuracy calculation. Ueda-Arakawa and colleagues97 reported that 84 eyes (38%) were excluded from analysis because of poor image quality. McBain and colleagues100 stated that nine eyes (9%) were excluded because of poor FAF images.

Quality assessment of diagnostic accuracy studies question 11: study withdrawals explained

In four studies there were no withdrawals and so this question was answered as ‘not applicable’. 83,84,98,101 In three studies the authors provided explanations for the withdrawals. 97,99,100 Cachulo and colleagues99 reported these as death, withdrawal of informed consent, hospitalisation and loss to follow-up. McBain and colleagues100 reported these as no FAF, more than 2 weeks between tests, FAF imaging < 4 days after the reference standard (no explanation provided), poor FAF images related to media opacities and cystoid macular oedema related to other diseases (e.g. previous branch vein occlusion, diabetic retinopathy or AMD). Ueda-Arakawa and colleagues97 reported these as due to phthisis bulbi or poor image quality in three or more imaging modalities. One study, by Hogg and colleagues,96 explained that images were excluded from analysis because of poor image quality but did not provide this information specifically for the index test.

Summary of quality assessment in relation to risks of bias

Seven of the eight included studies appear to be at high risk of spectrum bias (QUADAS question 1)83,84,96,97,99–101 since their patient populations appeared not to be representative of those likely to be encountered in clinical practice, and in the remaining study98 the risk of spectrum bias is unclear. Although most (five) of the studies appear to have employed reference standards that are commonly used and therefore might be considered acceptable (QUADAS question 2),83,84,99–101 these studies appeared to employ single imaging tests as the reference standard whereas diagnostic decisions in clinical practice would more likely employ a combination of imaging tests. For this reason the risk of verification bias was considered as being unclear for these five studies and high for the remaining three studies96–98 where the reference standard was unlikely to have accurately diagnosed the target retinal condition. The risk of disease progression bias (assessed in QUADAS question 3) is largely unclear, except for one study on cystoid macular oedema100 and both studies on diabetic macular oedema,84,98 which were judged to be at low risk. The risk of differential verification bias, that is, bias arising as a result of participants not all receiving the same reference standard (assessed in QUADAS questions 4 and 5), was considered to be low for most (six)83,84,98–101 of the studies but high for two studies96,97 on reticular pseudodrusen owing an unorthodox design of reference standard (where diagnosis required a positive result on a specified minimum number of imaging modalities). These same two studies on reticular pseudodrusen were considered at high risk of incorporation bias (assessed in QUADAS question 6) since the index test was not independent of the reference standard; the remaining six studies were judged to have low risk of this bias. 83,84,98–101 The risk of diagnostic review bias, that is, the index test result influencing interpretation of the reference standard result (assessed in QUADAS question 7) was unclear for most (seven) studies but judged to be low for one study on diabetic macular oedema. 98 The risk of test review bias, that is, where results of the reference standard influence interpretation of the index test result (assessed in QUADAS question 8) was unclear in most (six) studies84,96,97,99–101 but judged to be low for one study on cystoid macular oedema83 and for one study on diabetic macular oedema. 98 The risk of clinical review bias (assessed in QUADAS question 9) is unclear since none of the included studies reported whether their index tests were interpreted together with other clinical information, such as the results of routine ophthalmic examinations or results of other imaging modalities they employed. The risk of bias from uninterpretable or intermediate test results (QUADAS question 10) is generally unclear. Risk of attrition bias (QUADAS question 11) was judged to be low in five studies since either there were no exclusions,83,84,98,101 or the exclusions were clearly explained and appeared unlikely to influence the analysis. 97 Two studies explained their withdrawals but presented discrepancies in the numbers they analysed99,100 (see Appendix 5), while one study did not explain withdrawals;96 these three studies were judged to have an unclear risk of attrition bias. Overall, taking into consideration the different risks of bias mentioned above, it is not possible to identify individual studies, or specific retinal conditions, for which the quality of the clinical evidence would be regarded as being adequately robust to directly inform clinical practice.

An aspect of study quality not captured by the QUADAS criteria is that studies that included both eyes of a patient did not take into account adjustment for the correlation between the patient’s eyes, and this might lead to underestimation of standard errors for sensitivity and specificity. 95 This applies to two studies of reticular pseudodrusen in AMD by Ueda-Arakawa and colleagues97 and Smith and colleagues,101 and to both studies of diabetic macular oedema, by Vusojevic and colleagues98 and Waldstein and colleagues. 84

Choroidal neovascularisation in neovascular age-related macular degeneration

One study, by Cachulo and colleagues,99 assessed the diagnostic accuracy of FAF imaging (488 nm) against fluorescein angiography in detecting choroidal neovascularisation in AMD. The study authors reported sensitivity of 93% and specificity of 37%. However, using the data available from Cachulo and colleagues,99 we calculated sensitivity of 88% and specificity of 34% (Table 7) (for calculations see Appendix 5). This discrepancy might result from differences in categorising two (out of 52) eyes for which the pattern of FAF could not be determined because of poor-quality images. Confidence intervals for the estimates of sensitivity and specificity suggest moderate uncertainty around these values (Figure 3), but were not reported for the estimates calculated by the study authors.

FIGURE 3.

Diagnostic accuracy of FAF imaging for choroidal neovascularisation in neovascular AMD. CI, confidence interval; FN, false negative; FP, false positive; TN, true negative; TP, true positive.

Cachulo and colleagues99 did not report any information on interobserver or intraobserver agreement in image grading, test acceptability to patients or clinicians, or adverse events that might influence the interpretation of diagnostic outcomes.

As noted above (see Quality of research available), the single available study on diagnosis of choroidal neovascularisation in AMD99 was judged to be at high risk of spectrum bias (owing to an unrepresentative case mix); in addition the risks of several types of bias that could arise in relation to the sequence and timing of the index and reference tests, as well as the risk of bias in the interpretation of these tests, were unclear. Taking these considerations into account, it is not possible to draw a rigorous clinically relevant conclusion about the diagnostic accuracy of FAF 488 nm imaging as compared against fluorescein angiography in this retinal condition.

Reticular pseudodrusen in age-related macular degeneration (different stages)

Three studies, by Hogg and colleagues,96 Smith and colleagues101 and Ueda-Arakawa and colleagues,97 assessed the accuracy of FAF imaging for detecting reticular pseudodrusen in AMD.

Two of the studies did not report the FAF imaging excitation wavelength,96,101 while Ueda Arakawa and colleagues97 used FAF 488 nm and infrared FAF 790 nm (NIR-FAF) as index tests. The studies used various reference standards. Smith and colleagues101 used colour fundus photography, Hogg and colleagues96 required a positive result on one or more of five imaging modalities, and Ueda-Arakawa and colleagues97 required a positive result on two or more of seven imaging modalities. Each study aimed to detect reticular pseudodrusen in a different stage of AMD (see Table 4). Hogg and colleagues96 studied fellow eyes of patients with neovascular AMD in the non-study eye, Ueda-Arakawa and colleagues97 studied both eyes of patients newly diagnosed with early AMD, neovascular AMD or GA, and Smith and colleagues101 included a mixture of both eyes of patients without any evidence of choroidal neovascularisation and fellow eyes of patients with unilateral choroidal neovascularisation in the non-study eye.

By taking together the appropriate reference standards for diagnosing reticular pseudodrusen (i.e. colour fundus photography and optical coherence tomography) and the other reference standards nominated in the primary studies, data are available for six diagnostic accuracy comparisons for FAF imaging (Table 8). Although the reference standard reported by Hogg and colleagues96 involves a positive result on one or more tests, individual comparisons of FAF imaging against colour fundus photography and FAF imaging against optical coherence tomography were also reported.

FAF imaging based on a 488 nm excitation wavelength had relatively high (86–100%) sensitivity for detecting reticular pseudodrusen when compared against the various reference standards (Figure 4). However, near-infrared FAF imaging had low (32%) sensitivity (see Figure 4). In contrast, both FAF and NIR-FAF imaging had high (84–100%) specificity compared against the various reference standards, except for a lower specificity (67%) when FAF imaging was compared against colour fundus photography in the study by Hogg and colleagues. 96 A major limitation of the results for near-infrared FAF imaging reported by Ueda-Arakawa and colleagues97 is that 38% of the eyes were excluded owing to poor image quality.

FIGURE 4.

Diagnostic accuracy of FAF imaging for reticular pseudodrusen in AMD. CFP, colour fundus photography; CI, confidence interval; FN, false negative; FP, false positive; OCT, optical coherence tomography; TN, true negative; TP, true positive. a, Versus two or more imaging modalities.

For most of the studies the 95% confidence intervals around the specificity estimates are relatively narrow. However, confidence intervals for specificity reported by Hogg and colleagues,96 and those reported for sensitivity by all studies are wider, meaning that these estimates may be less reliable (see Figure 4). However, it is possible that the width of the confidence intervals reported by Smith and colleagues101 and Ueda-Arakawa and colleagues97 might have been underestimated, as the analyses did not account for any intrapatient correlations in cases where both eyes of the same patient were analysed.

Two of the studies, by Hogg and colleagues96 and Ueda-Arakawa and colleagues,97 provided information on intergrader agreement in diagnosing reticular pseudodrusen using FAF imaging. Kappa values indicate actual agreement compared with that which would occur by chance. A value of 1 indicates perfect agreement and a value of 0 indicates agreement equivalent to chance. Kappa values were 0.563 (n = 35, p < 0.001) (84.2% agreement)97 and 0.94 (sample size, p-value and percentage agreement not reported). 96 These kappa values could be interpreted as ‘moderate’ and ‘almost perfect’ agreement, respectively. 104 Although intergrader agreement was almost perfect in Hogg and colleagues’ study,96 it was reported for only one of their three study centres and so it is unclear whether or not the ‘best’ agreement was selectively presented or whether or not the reported agreement is reflective of that across all the study centres. None of the studies on reticular pseudodrusen provided any information on intragrader variation in diagnosis, the acceptability of the imaging modalities to patients or clinicians, or on whether any adverse events of any tests were observed.

In summary, the diagnostic outcomes suggest that FAF imaging with an excitation wavelength of 488 nm could potentially contribute to the diagnosis of reticular pseudodrusen but near-infrared FAF imaging (790 nm) does not appear adequately sensitive to be of diagnostic value. However, there is considerable uncertainty around these findings, owing to the limited number of studies and comparisons available and shortcomings in the study designs. As noted above (see Quality of research available), the three available studies on diagnosis of reticular pseudodrusen in AMD96,97,101 were judged to be at high risk of spectrum bias (owing to an unrepresentative case mix), while the risks of several types of bias that could arise in relation to the sequence and timing of the FAF imaging tests in relation to those of the reference standards, and in the interpretation of these tests, were unclear. In addition, in the studies by Hogg and colleagues96 and Ueda-Arakawa and colleagues,97 the FAF imaging index tests also formed part of the reference standard, which complicates interpretation of any diagnostic outcomes. Furthermore, Ueda-Arakawa and colleagues97 had to exclude over one-third of the eyes from their analysis of near-infrared FAF imaging because good image quality could not be obtained with this method. Taking these considerations into account, it is not possible to draw a rigorous clinically relevant conclusion about the diagnostic accuracy of FAF imaging for diagnosing reticular pseudodrusen in AMD.

Cystoid macular oedema secondary to various conditions

Two studies, by Dinc and colleagues83 and McBain and colleagues,100 reported the diagnostic accuracy of 488 nm FAF imaging for identifying cystoid macular oedema secondary to various conditions. Both studies employed fluorescein angiography as the reference standard, which would partly reflect current clinical practice (optical coherence tomography and possibly also colour fundus imaging would be used in addition to fluorescein angiography to reach a diagnosis in clinical practice). The study by Dinc and colleagues83 indicated high (98%) sensitivity with a narrow confidence interval suggesting good precision of the estimate, while that of McBain and colleagues100 indicated moderate sensitivity (81%) but with high uncertainty (Table 9 and Figure 5). However, the former study had 0% specificity (reflecting a lack of true negatives in the data) and hence, as the likelihood ratios suggest, the study is not diagnostically informative. The latter study, by McBain and colleagues,100 is limited by the uncertainty around its sensitivity and specificity estimates which, according to the lower confidence limits, could have been as low as 58% and 39%, respectively.

Neither of the two studies on cystoid macular oedema83,100 reported any information on interobserver or intraobserver agreement in image grading, test acceptability to patients or clinicians, or adverse events of any tests that might influence the interpretation of diagnostic outcomes.

FIGURE 5.

Diagnostic accuracy of FAF imaging for cystoid macular oedema. CI, confidence interval; FA, fluorescein angiography; FN, false negative; FP, false positive; TN true negative; TP, true positive.

As noted above (see Quality of research available), both studies on diagnosis of cystoid macular oedema83,100 were judged to be at high risk of spectrum bias, owing to an unrepresentative case mix. Although Dinc and colleagues83 interpreted results of the index test (FAF imaging) without knowledge of the results of the reference standard (fluorescein angiography), indicating a low risk of test review bias, the risks of diagnostic review bias and clinical review bias were unclear in both studies. Taking these considerations into account, noting that one study appears to be diagnostically uninformative83 while the other has high uncertainty of its diagnostic outcomes,100 it is not possible to draw a rigorous clinically relevant conclusion about the accuracy of FAF imaging for diagnosing cystoid macular oedema.

Diabetic macular oedema

Two studies, by Vujosevic and colleagues98 and Waldstein and colleagues,84 reported the diagnostic accuracy of FAF imaging for diabetic macular oedema. Vujosevic and colleagues98 compared 488 nm FAF imaging against a reference standard of retromode scanning laser ophthalmoscopy. Waldstein and colleagues84 compared 488 nm and 514 nm FAF imaging against a reference standard of spectral-domain optical coherence tomography. For diabetic macular oedema, optical coherence tomography would be a clinically relevant reference standard (as acknowledged by Vusojevic and colleagues98), although, according to clinical experts consulted during the present review, it would likely be combined in clinical practice with colour fundus imaging and, if macular ischaemia is suspected, also with fluorescein angiography.

Diagnostic outcomes are shown in Table 10 and Figure 6. Sensitivity of FAF imaging (488 nm) for diagnosing diabetic macular oedema was highest (92%) when compared against retromode scanning laser ophthalmoscopy, although this is unlikely to be used routinely in clinical practice. Sensitivity of FAF imaging when compared against spectral-domain optical coherence tomography was moderate (81%) for the shorter-wavelength imaging (488 nm) but relatively low (55%) for the longer-wavelength method (514 nm). In contrast, specificity was relatively high for both wavelengths of FAF imaging when compared against spectral-domain optical coherence tomography (90% to 95%) but lower when 488 nm FAF imaging was compared against retromode scanning laser ophthalmoscopy (85%). The estimates of sensitivity and specificity have relatively narrow confidence intervals, suggesting they are reasonably precise, although it is possible that the width of the confidence intervals for both studies might have been underestimated, since the analyses did not account for any intrapatient correlations where both eyes of the same patient were analysed.

FIGURE 6.

Diagnostic accuracy of FAF imaging for diabetic macular oedema. CI, confidence interval; FN, false negative; FP, false positive; RM-SLO, retromode scanning laser opthalmoscopy; SD-OCT, spectral domain optical coherence tomography; TN, true negative; TP, true positive.

Vusojevic and colleagues98 did not report intergrader agreement on interpreting FAF images. Waldstein and colleagues84 reported kappa values (but not sample size, p-value or percentage reviewer agreement) for intergrader agreement in interpretation of the FAF images, which were 0.84 for 488 nm FAF imaging and 0.63 for 514 nm FAF imaging. These could be interpreted as ‘almost perfect agreement’ and ‘substantial agreement’, respectively. 104 Neither study reported intraobserver agreement for image grading, test acceptability to patients or clinicians, or adverse events of any tests that might influence the interpretation of diagnostic outcomes.

As noted above (see Quality of research available), the study by Waldstein and colleagues84 was judged to be at high risk of spectrum bias (owing to an unrepresentative and retrospectively selected case mix) while the risks of various types of bias associated with the timing of the index test and reference standard, and interpretation of these tests, were judged to be unclear. The study by Vujosevic and colleagues98 was considered to be at unclear risk of spectrum bias because although the case mix appears potentially reflective of patients likely to receive retinal imaging in clinical practice (diabetic retinopathy patients with comorbidities permitted) it is unclear whether the patients were selected prospectively or retrospectively. This study98 was the only one of the eight included in the systematic review that was deemed to be at low risk of both test review bias and diagnostic review bias. However, as noted above, the reference standard employed by Vusojevic and colleagues98 is unlikely to reflect clinical practice. Taking these considerations into account, it is not possible to draw a rigorous clinically relevant conclusion about the accuracy of FAF imaging for diagnosing diabetic macular oedema, although results from the study by Waldstein and colleagues84 suggest that the longer-wavelength FAF imaging method (514 nm) was less sensitive for diagnosing diabetic macular oedema than the standard wavelength approach (488 nm).

Summary of diagnostic accuracy assessment

The eight studies83,84,96–101 included in the systematic review suggest qualitatively that FAF imaging may have a potential role in supporting the diagnosis of choroidal neovascularisation in AMD, reticular pseudodrusen in AMD, cystoid macular oedema or diabetic macular oedema. However, after taking into account a number of quality issues concerning the study designs, none of the studies were found to provide convincing quantitative information about the diagnostic accuracy of FAF imaging of relevance to current clinical practice. In two cases where long and short excitation wavelengths of FAF imaging could be compared, the longer wavelength appeared less diagnostically sensitive: 719 nm was less sensitive than 488 nm for detecting reticular pseudodrusen97 while 514 nm was less sensitive than 488 nm for detecting diabetic macular oedema. 84 However, these differences are based on only two studies that had methodological limitations and, as such, are illustrative only.

Methodological challenges

A challenge in the current review was to identify studies of the diagnostic accuracy of FAF imaging based on screening titles and abstracts. It was difficult to determine, based on titles and abstracts alone, whether or not studies reported sensitivity and specificity, or data from which these could be calculated. As a result it was necessary to retrieve and check 206 full-text records. Nevertheless, only eight studies finally met the inclusion criteria. 83,84,96–101 A difficulty is that most of these eight studies that provided relevant diagnostic outcomes were not primarily diagnostic studies and as such they were rated relatively poorly against the QUADAS criteria for risks of bias. Ideally, future studies that wish to explore the diagnostic accuracy of FAF imaging should be based on appropriate designs and methods compatible with the paradigm of diagnostic accuracy assessment,108,109 and should include a clear statement in the abstract as to the availability of quantitative diagnostic outcomes. 110

Patient and public involvement

This review did not formally involve patients or the general public. However, the draft protocol and report were provided to the advisory group, which included the RNIB, for comment. The RNIB did not identify any major issues of equitability in the use of FAF imaging from the perspective of patients and the general public. However, it should be noted that the technology is relatively new and the extent to which patients have access to the technology is currently unclear.

Strengths

• The review was based on a prespecified peer-reviewed protocol.

• A comprehensive literature search was conducted based on a wide range of prespecified evidence sources.

• All steps of the systematic review process were conducted by at least two reviewers, minimising the risks of errors and bias.

• The primary evidence was assessed critically using accepted criteria for the critical appraisal of test accuracy studies.

• All steps of the evidence synthesis are reported transparently; worksheets for study selection and data extraction were pilot tested and are presented with this report; all studies excluded at full-text screening are listed in Appendix 4, stating reasons for exclusion.

• An independent advisory group informed the review.

Limitations

• Interpretation of the primary research is hampered by clinical heterogeneity among the studies and limitations in methodological rigour, which in all studies led to judgements of high risk of bias. Owing to the small number of studies available it was not possible to assess the impact of methodological rigour on observed diagnostic outcomes.

• The primary research studies had an inappropriate (or in one study unclear98) patient spectrum (i.e. a case mix unlikely to be reflective of that presenting for retinal imaging in current NHS practice).

• Some studies included both eyes of patients in analyses, which (because of intrasubject correlations) may have led to underestimation of standard errors for sensitivity and specificity (and hence also underestimation of the width of the confidence intervals). 95 As there is no formally agreed approach for analysing such data in evidence syntheses,95 the current review does not investigate their impact on the precision of diagnostic outcomes. However, such cases are clearly identified in the review and any inaccuracy of confidence intervals in these cases would not have markedly altered the review conclusions.

• The systematic review was limited to English-language studies.