Repeatability of choroidal thickness measurements with Spectralis OCT images

Introduction

The choroid, a thin elastic vascular structure between the retina and the sclera of the eye, is known to play several roles. Apart from providing oxygen and nutrients to the outer retina, it is also important in thermoregulation,1 drainage of aqueous humour in the uveoscleral outflow,2 adjustment of the retinal position,3 4 secretion of growth factors5 and, possibly, ocular elongation (see review by Nickla and Wallman).6

Myopic defocus imposed with a positive lens leads to an increase in the choroidal thickness, whereas hyperopic defocus with a negative lens results in thinning in both animal7 and human8–10 studies. Such changes in choroidal thickness result in anterior or posterior displacement of the retina towards the image plane and lead to decreased or increased growth of the eyeball in chicks.7 The changes in choroidal thickness may be associated with the changes in axial length in an approximate antiphase relationship.8

For many years, the ultrasound biometer has been used to measure axial length,11 anterior segment structures12 and posterior segment components.13 However, the use of advanced technology, in the form of optical coherence tomography (OCT), provides greatly improved images and more accurate measurements, so is the instrument of choice in research. A typical 10-MHz ultrasound instrument provides an axial resolution of approximately 200 µm and a transverse resolution of 500 µm. In contrast, a spectral-domain OCT with a 870 nm super luminescent diode (Spectralis OCT; Heidelberg Engineering, Heidelberg, Germany) provides 40 000 A-scans per second and cross-sectional images with axial and lateral resolutions of 3.9 and 11 µm, respectively. The use of ultrasound biometry also requires topical local anaesthetic which may induce ocular adverse events, such as contact dermatitis and subconjunctival haemorrhages,14 and possible measurement errors due to misalignment of the probe15 which make it less valuable for measuring axial length and choroidal thickness in children. OCT, on the other hand, does not have such limitations.

Spaide et al 16 described enhanced depth imaging OCT (EDI-OCT) which involves placing the instrument closer to the eye, which reduces the displacement of deeper structures from zero delay in order to obtain better visualisation of images. This technique has been widely used in the field for obtaining choroidal thickness17–21 and assessing ocular diseases.22–27 The Spectralis OCT also provides eye-tracking and image averaging functions which allow for evaluation of localised changes in retinal and choroidal thickness at follow-up visits.

In most published studies,28–33 choroidal thickness was determined manually due to the unavailability of built-in functions. It has been shown that the coefficient of repeatability (CR) of the interimage repeatability (in which images were captured consecutively at the same visit) was about 35 µm.30 34 With the help of image processing and analysis software, Boonarpha et al 35 measured choroidal thickness using manually marked boundaries and showed an overall CR of 54 µm in choroids of different contours and shapes. However, manual procedures involved for choroidal segmentation requires longer execution time36 and subjective variation in determination of chorioscleral interface (CSI) could lead to substantial measurement errors and bias. To reduce the error of subjective discrepancies involved in the procedures, automated methods have been proposed. Alonso-Caneiro et al 37 developed a software for automatic segmentation of choroidal thickness with a smooth spline-fit function based on graph theory which facilitates the detection of the choroidal boundaries with minimal subjective judgements. They demonstrated agreement between manual and their proposed algorithm with a limit of agreement (LoA) of +35.37 to −30.79 µm for data from both children and adults. Twa et al 38 also used a software using analysis of graph theory and dynamic programming and produced a comparable repeatability of choroidal thickness to manual segmentation (where the LoA of manual and automatic measurements were ±15 and±14 µm, respectively). However, the repeatability of using the proposed algorithm in measuring choroidal thickness in children, especially those receiving myopic control treatment, is yet uncertain.

With increasing prevalence of myopia worldwide39 40 and associated pathologies with high myopia,41–43 many optical and pharmacological interventions have been investigated in controlling myopia in children (summarised by Huang et al).44 Orthokeratology (ortho-k) uses reverse-geometry rigid gas permeable lenses to affect ocular changes to slow myopia progression45–47 and is a popular optical method used for myopia control, especially in East Asian countries.48 A number of studies have reported about 50% slower axial elongation in children wearing ortho-k compared with control subjects wearing single-vision (SV) spectacles or contact lenses.45 49–51 Although two studies52 53 have shown subfoveal choroidal thickening in subjects wearing ortho-k lenses for 1–6 months and suggested that this may be negatively correlated with some changes in axial length in early ortho-k treatment, another study54 did not observe any choroidal thickening after 9-month ortho-k treatment. Furthermore, central flattening and midperipheral steepening of the cornea after ortho-k lens wear55 may affect the magnification of OCT images and also choroidal thickness measurement. Thus, it is important to validate the repeatability of measurements in ortho-k treated children before confirmation of these changes.

Moreover, a few studies16 31 56 have examined the repeatability of choroidal thickness measurement using correlation coefficients, but Bland and Altman57 reported this could be misleading and inappropriate and suggested using LoA and CR as repeatability parameters.

Therefore, the aim of this study was to determine and compare the repeatability of choroidal thickness measurement in spectacle-wearing and ortho-k wearing children using the semiautomated segmentation software on EDI-OCT images.

Methods

The EDI-OCT scans were retrieved from two longitudinal ortho-k studies which were approved by the Departmental Research Committee of the School of Optometry at The Hong Kong Polytechnic University and registered at ClinicalTrial.gov (NCT02643875 and NCT02643342). Procedures performed were in accordance with the Declaration of Helsinki and informed consent was obtained from the parents after full explanation and disclosure of the objectives of the ortho-k research.

Subjects were Chinese aged 6 and 10 years (inclusive), with low myopia (0.50–4.00 D) and low astigmatism (≤1.25 D) in both eyes, and low anisometropia (≤1.50 D). Their eyes were symmetrical in corneal topography (<2.00 D difference). They had no prior myopia control experience nor any systemic or ocular conditions that could affect refractive development. A total of 40 healthy subjects were randomly selected, 20 wearing SV spectacles and 20 undergoing ortho-k treatment for 1–4 weeks.

The Spectralis OCT was used for capturing chorioretinal images. For each subject, three series of EDI-OCT images were taken, using the high-speed star scanning protocol, consisting of six 30° long line scans, each separated by 30°, radially centred at the fovea. All the images were captured under EDI mode for enhanced visualisation of choroidal tissues. The automatic real-time mean was set to allow using the average of 30 B-scans. The first image taken was treated as the reference image and served as the registration link for the following images. The ‘Auto Rescan’ function was applied in order to achieve sequential images of the same retinal location tracked by the confocal scanning laser ophthalmoscopy. The first two images captured for either eye (randomly selected) were used for data analysis.

The EDI-OCT images were then exported to custom written software for choroidal thickness segmentation.37 The measurements of choroidal thickness were performed using the horizontal line scans of the EDI-OCT images. Choroidal thickness was automatically determined as the thickness between the outer retinal pigment epithelium/Bruch’s membrane complex and the inner CSI. An experienced observer, masked to the treatment received by each subject, then manually corrected any segmentation errors in the boundary outlines generated by the software and determined the centre of the foveal pit, before the software derived the choroidal thickness. Figure 1 shows the outlines of different layers of the retina and the choroid as generated by the software and the centre of foveal pit determined by the observer. The measurements of the choroidal thickness for the second images were performed 4–6 weeks later to minimise subjective bias and the learning effect on the correction of segmentation errors in the images.

Request permission

The interimage repeatability was calculated using the measurements obtained from the two images. The distribution of the choroidal thickness measurements followed a normal distribution (Shapiro-Wilk tests, p>0.05) and therefore parametric statistical analyses were used. The two measurements were compared using paired t-tests and the relationship between the mean and the difference of measurements was checked using Pearson’s correlation test. The mean interimage differences and LoA of the differences were determined and plotted as suggested by Bland and Altman.57 58 Since the repeatability of the same method was being tested on two consecutive OCT scans, the mean difference should be mimimal. The CR can be calculated as twice the SD of the difference between two measurements.57 The average of two choroidal thickness measurement was also reported. A probability (p) value of less than 0.05 was chosen to indicate statistical significance. All statistical analyses were performed using SPSS data analysis software (SPSS V.22).

Results

The SV group comprised 5 males and 15 females with a mean (±SD) age of 8.7±1.1 years, whereas the ortho-k group consisted of 7 males and 13 females aged 9.4±1.0 years. The baseline age of SV subjects were younger than those of ortho-k subjects (unpaired t-test, p=0.043). However, the baseline refractive errors, including sphere, cylinder, and spherical equivalent refraction, were not significantly different between two groups (unpaired t-tests, p>0.05; table 1). The mean (±SD) choroidal thicknesses obtained were not significantly different between groups (unpaired t-test, p>0.05; SV: 227.3±42.2 µm; ortho-k: 251.1±54.4 µm).

The choroidal thickness measurements from the two images were not significantly different for both groups of subjects (paired t-tests, p>0.05). There were no significant correlations between the differences and the means of choroidal thickness measurements in either group (−0.38<Pearson’s r<0, p>0.05). The mean interimage differences were −0.99±3.54 and −1.14±5.03 µm for the SV and ortho-k subjects, respectively. For the SV group, the LoA was +5.96 to −7.93 µm, and for the ortho-k group, the LoA was +8.72 to −11.00 µm. The CRs were 7.08 and 10.06 µm for the SV and ortho-k groups, respectively. Figure 2 shows the Bland and Altman plots of the interimage differences and their means.

Request permission

Discussion

This study showed that using a written graph-theory-based software to segment EDI-OCT images produced a repeatability of choroidal thickness measurement of approximately 7 and 10 µm in children wearing SV spectacles and ortho-k lenses, respectively. Previous studies30 34 35 38 have segmented the choroidal thickness manually, while others38 59 have applied algorithm-based software for automatic segmentation and measurement of choroidal thickness. While the CR estimate is useful in setting an assessment tool of minimal detectable change,60 61 our results of a CR of approximately 10 µm indicated that automatic measurement of choroidal thickness was more repeatable than other studies solely using manual segmentation (CR of about 35 µm).30 34 Our results were comparable to the interobserver repeatability results obtained by Li et al,53 whose CRs of SV and ortho-k groups were 9.12 and 6.58 µm, respectively. However, the challenges involved in fully manual segmentation include weak signals resulting from deeper structures, the vascular nature of the choroidal tissue and varying thickness of choroid.30 37 62 With the introduction of an automated segmentation technique, researchers require less time for segmentation procedures and this reduces subjective variation in determining the choroidal boundaries and therefore the measurement of its thickness.

Researchers have developed different algorithms for automated segmentation of the choroid. The software used in this study, previously described by Alonso-Caneiro et al,37 uses a combination of an edge filter and a directional weighted map penalty to detect the inner choroidal boundary and a dual brightness probability gradient for the detection of the outer choroidal boundary.37 Other studies have applied different approaches, including dynamic programming,62 63 graph search method,62 statistical modelling,63 64 estimated morphological modelling,64 active contour modelling,65 Gaussian mixture modelling66 and a combination of polarisation and birefringence67 to differentiate the boundaries. Compared with the software developed by Twa et al,38 whose LoA was ±14 µm, the performance of this software (LoA of +5.96 to −7.93 and +8.72 to −11.00 µm for the SV and ortho-k groups, respectively) was slightly better. Different algorithms could result in discrepancies in the detection of choroidal boundaries due to variations in graph analysis methods. Further investigation of the repeatability of different automatic segmentations of choroidal thickness may warrant their use in detecting subtle choroidal changes, such as diurnal variation, early detection of ocular diseases or responses to optical defocus.

Read et al’s study,59 which investigated the longitudinal changes in choroidal thickness in children, reported that a semiautomated method of choroidal thickness (automatic detection for the inner choroidal boundary and manual segmentation for the outer boundary) produced an excellent interimage repeatability for the subfoveal choroidal thickness with an LoA of approximately ±4 µm, which was almost the axial resolution of the instrument (3.9 µm). They used high-resolution acquisition mode, which could produce double the number of A-scans (1536 scans) compared with the 768 scans we captured under high-speed mode. However, the difference in the number of A-scans would only affect the lateral resolution and have minimal influence on subfoveal choroidal thickness measurement. A study which compared the two acquisition modes for measurement of the retinal nerve fibre layer thickness found no significant difference between the results.68 However, the high-resolution mode requires longer acquisition time and some children might not be able to maintain steady fixation during the scanning procedures.59 The variability in choroidal thickness obtained using these two modes is yet to be confirmed. Another possible reason for the discrepancy in measurements are the subjective procedures involved in segmenting or correcting the choroidal boundaries. Different levels of experience of the observer in choroidal segmentation could lead to possible errors in the measurement.

The CR for interimage analysis was approximately 10 µm. Although measurement error from scan misalignment was minimised by using the Auto Rescan function for tracking the same retinal location, a possible lateral measurement error of 11 µm could still result due to the instrument’s limitation, which might lead to inconsistent measurement of the same exact location. Errors can also result from subjective judgement in correcting segmentation errors and determination of the centre of the foveal pit.

A larger variability in choroidal thickness measurement was observed in our ortho-k children. As previously reported, central flattening and midperipheral steepening were observed in corneas after wearing ortho-k lenses.55 These changes result in hyperopic shifts in the peripheral refractive powers, but only at some midperipheral regions of the cornea.69 Therefore, the correcting lens in the Spectralis OCT may not be able to provide completely in-focus images. In Li et al’s study,53 they found that manual segmentation of EDI-OCT images to be slightly more repeatable for ortho-k subjects compared with SV children. However, in both studies, the differences in CR between SV and ortho-k groups were about 3 µm, which was less than the axial resolution (3.9 µm) of the instrument and this difference may not be significant.

Our study used images of consecutive scans which eliminates possible diurnal variation of the choroid and the influence of other ocular components on its thickness within the day.70 71 A strength of the study was that children were included. While different populations may show different repeatability errors,37 our results helped to identify minimal detectable changes of choroidal thickness in ortho-k treated children. Apart from indicating that all children were able to successfully undergo OCT scans under the high-speed scanning mode, the involvement of ortho-k treated children in our study help to identify real changes in choroidal thickness, despite the influence on the optics of the eye from corneal reshaping. A limitation of the study was that the association between the effects of changes in corneal shape on OCT images and choroidal thickness measurement was not studied. Further investigation on the repeatability of choroidal thickness measurement in subjects with distorted or reshaped optical components, such as keratoconus and LASIK-treated subjects, may help to identify the possible reasons of slightly larger variation in choroidal thickness measurement in the ortho-k children.

Conclusion

In conclusion, our results show that using an automated algorithm produces a good repeatability of choroidal thickness measurement in Chinese children. A subfoveal choroidal change of more than 10 µm is likely to be a real detectable change in ortho-k wearing children.