Discussion
We have described the 3D morphology of macular hole using a novel and validated automated 3D segmentation algorithm. The algorithm is robust and was able to accurately segment the full consecutive series of 104 OCTs included in the study, including when there was VMT present. We used a high-density scanning protocol with 30-μm line spacing and averaging 16 A scans per line, reducing noise and meaning that the scan lines were more likely to include the maximum hole dimensions.11 We have previously shown that the 3D methodology can very accurately segment out the macular hole boundaries as compared with a human observer, and can therefore be regarded as providing a ground truth for macular hole dimensions and shape. Macular holes are shown to be complex shapes with significant asymmetry, meaning that conventionally acquired clinician measurements fail to represent their key parameters accurately. For example, we found that the XZ meridian of the minima of the MA was only within 10° of the conventionally measured horizontal X axis in 10% of cases, and differed from the human-measured MLD by a mean of nearly 50 μm and up to 200 μm. Similarly, the true maximum BD varied from the mean of the human measurements by 87 μm or 12%. The resultant differences led to a reclassification in size using the International Vitreomacular Traction Study Group classification in a quarter of the patients.5 16 17 This has significant implications for studies using macular hole measurements to predict outcomes and to act as cut-off points for deciding on treatments.
The human measurements had a consistent tendency to overestimate the widths of the holes. To measure a macular hole MLD, a human observer must first accurately locate the scan line with the greatest dimensions and then pick the minimum hole dimension, avoiding the area of the operculum if present. The minimal dimension is typically measured parallel to the RPE. Measuring macular hole using a horizontal line scanning protocol relies on the macular hole being symmetric, but we show that the holes were significantly asymmetric in all dimensions. There was a mean difference of 55 μm in maximum and minimum dimensions of the MA and 87 μm for the same measures of the BA. These differences concur with those found by Philippakis et al17 using en face SDOCTs to measure macular hole dimensions, although they did not comment on the orientation of the maximum/minimum measurements. The minima of the MA were typically approximately 90° to the horizontal, while the maximum of the BA was predominantly horizontal. The holes were therefore oval with their maximum dimension in the XZ axis at the horizontal meridian. Interestingly this corresponds to asymmetries found in the foveal avascular zone (FAZ), where previous studies have found an approximate 30-μm difference, with the horizontal diameter being widest.18 It is known that FAZ size is closely related to foveal floor size, and a recent study has suggested an association between macular hole size and foveal floor width.19 20 The clinically acquired measurement of MLD is thus typically measured in an axis that does not coincide to its true minimal dimension. Indeed, using an algorithm-derived horizontal dimension of the minimal area, corresponding more to the human MLD measurement, there was no significant difference in the measurements.
Although the holes were generally vertical, the centre point of the MA and BA was misaligned by over 150 μm in 70% of the eyes. This therefore adds to the measurement error of human graders who have tended to measure the MLD and maximum BD on the same SDOCT slice when in reality this occurrence will rarely occur. These asymmetries further explain the human measurement error compared with the true measurements found by the algorithm, as well as interobserver variability. The 95% limits of agreement between the two observers for MLD were −140 to 146, which is in broad agreement with the values found by Banerjee et al.21 We asked observer 2 to record the height above the RPE at which they measured the MLD, and although it was not significantly different from the height the algorithm measured the minimal area at, it varied from the algorithm by more than 40 μm in 29% of eyes, which is likely another source of error. The two observers were both experienced in measuring macular holes and from the same institution, and it is likely that less experienced observers, with different training, may have had even greater differences between them. If different scanning protocols and OCT machines were added, then the differences would be greater again. We did not assess intravisit variability, which would have increased the variability further.
It is thought that most macular holes are formed by the effects of anteroposterior vitreoretinal traction and VMT. Of the holes in this series 26% had VMT, which is in keeping with previous figures from the same population area.22 We found no significant association between the presence of VMT and any of the size parameters measured, which is in keeping with the findings of Philippakis et al.23
We found that age and gender were not significantly associated with the algorithm and human-measured values, but preoperative vision was, as other authors have found.24 25 The strongest relationships were all those derived from the algorithm as opposed to the human observers. However, the preoperative visions were checked without a protocol refraction and exact relationships are uncertain.
Three other approaches have been suggested to evaluate macular hole shape and dimensions beyond human measurements from standard OCT line scans. Philippakis et al17 elegantly demonstrated the use of en face reconstruction to measure macular hole minimal area and dimensions. The technique however had a high technical failure rate of ~50%, often had to be manually adjusted when VMT was present and was unable to measure other hole parameters. Problems may also be encountered where holes are misaligned vertically as we have already observed above. Geng et al26 used a manual segmentation technique combined with Matlab to produce a 3D representation of the hole from which 3D parameters could be measured, but involves a time-consuming manual mark-up. Xu et al27 have described an approach of automatically measuring macular hole dimensions based on the sum of 2D images. In comparison, our algorithm considers the overall 3D geometry of the hole and is significantly faster. We have also validated the accuracy of our system against human segmentation in a set of 30 eyes and showed very high accuracy.
Our study has several limitations. We did not correct the measurements for axial length, although we restricted the entry criteria to eyes with axial lengths between 22 and 25 mm. Furthermore, inaccuracies introduced by doing this would only affect absolute measurements, not the differences in dimension we describe nor differences from the human measures. We used a specified scanning protocol by one OCT manufacturer, which also limits the applicability of our technique and interpretation of our findings. Similarly, although a consecutive cohort, our sample was restricted to patients undergoing surgery in one centre, which may not be representative of all idiopathic macular holes or other ethnicities and populations.
In conclusion we have previously described a 3D automated macular hole segmentation system that is able to accurately segment out a macular hole from its constituent cross-sectional 2D scans. We now present the detailed findings from a cohort of 104 consecutive macular holes, with description of several clinically relevant 2D and 3D dimensions derived from the 3D macular hole shape extracted. We show that the measurements are significantly different from those measured by experienced human graders. Macular hole size is known to be one of the strongest predictors of surgical success both anatomically and functionally. Evaluation of the measurements generated from this automated system in a prospectively collected data set of eyes undergoing surgery with outcomes analysis will be of great interest.