Maximum value projection produces better en face OCT angiograms than mean value projection

: Optical coherence tomography angiography (OCTA) images rely on en face data projections for both qualitative and quantitative interpretation. Both maximum value and mean value projections are commonly used, and many researchers consider them essentially interchangeable approaches. On the contrary, we find that maximum value projection achieves a consistently higher signal-to-noise ratio and higher image contrast across multiple vascular layers, in both healthy eyes and for each disease examined.


Study po
the Institution e tenants of the quisition cular scans we (Avanti RTV the fast transve repeated B-sca d in the slow sampling is 3 cans at each correlation (S improve the s orthogonal sca MCT). The p ifacts [21]. with nine or ten ence Universit 3 scans of eac dividuals, 1-2 s ated (9 individ oard/Ethics Co of Helsinki. ith a 70-kHz, vue Inc.) using direction, 304 ired at each lat ection. Axial r ral OCT data OCTA data w rithm [6,28]. ratio of image ons, registered olved OCTA e projection te n slabs that w capillary plexu d bidirectiona projection as a ong an A-scan tained within ated as 1 = D N 1 2 , , , )  N D D D  [29].
As seen in Fig. 1, slabs are of differing thickness and contain different relative amounts of signal and noise. Concisely, the task for en face projection is to determine what decorrelation value ( , ) D x y within a certain slab will lead to the most informative and interpretable en face image. While many algorithms interpret en face data, the specifics of various analytic techniques are not in general conserved, which means that the "best" metric for determining image quality may be context dependent. We therefore limited our analysis to metrics that are of broad interest and essentially agnostic with respect to analytic detail; specifically, the signal-to-noise ratio (SNR) and root-mean-square contrast (RMS contrast). We calculated both of these for en face projection images obtained by either mean or maximum value projection. The former, SNR, can be conveniently defined in OCTA as where parafovea D and FAZ D represent means of the decorrelation in the parafovea and foveal avascular zone (FAZ), respectively, and 2 FAZ σ is the variance in decorrelation in the FAZ.
Since, in healthy subjects, the FAZ is avascular 2 FAZ σ serves as an excellent estimate of the background in OCTA images. In the present study we manually demarcated a region within the FAZ to serve this purpose (Fig. 3). RMS contrast is a similarly simple metric obtained by calculating the RMS value across all pixels ( , ) x y in a region of interest A: In the present study we take as the region of interest the entire en face projection. An image may have high SNR but low RMS contrast, for instance if the signal strength is strong but within signal regions pixel values are similar; alternatively, a low SNR image with high contrast may have sharp variation between pixels, yet this variation may be due in part or totally to noise. RMS contrast and SNR are thus complimentary measures, and in conjunction they provide a versatile description of image quality.
To investigate the impact of the projection technique on clinically relevant metrics, we also calculated vessel density (VD) from en face projections. VD is quantified as the percentage VD = (vascular pixels)/(total pixels) × 100. To identify vascular vs. avascular pixels we employed an Otsu threshold. The Otsu threshold is a common approach that determines a threshold value by minimizing the intraclass variance. From this threshold an effectiveness metric can be calculated. The effectiveness metric runs from 0 to 1, with 0 indicating completely ineffective segmentation (as would result from attempting segmentation on an image in which all pixels shared an identical value) and 1 indicating perfect segmentation (for instance from two totally distinguishable pixel populations, i.e. a binary image).
We conducted this research with custom software written in Matlab 2018a (Mathworks, Natick, MA).

Results
We examined a total of 46 OCTA images from healthy eyes, calculating the SNR and RMS contrast values for both maximum and mean value en face projections in 3 different retinal slabs, the SVC, ICP and DCP. To elucidate the results, we plot binned (12 points/bin, average) SNR and RMS contrast by signal strength index (SSI), a proprietary metric output by the RTVue-XR instrument. SSI is based on the average tissue reflectance amplitude of the OCT image on a logarithmic scale; it is broadly indicative of OCTA image quality and runs on a scale of 1 (poor) to 100 (excellent).
In this data set of healthy eyes maximum value projection achieves on average a better SNR and RMS contrast across every slab and SSI value. The difference between maximum value and mean value SNR and RMS contrast is greatest in the SVC and least in the ICP; not coincidentally, these are the thickest and thinnest slabs examined in Fig. 2 (mean thickness 52.9 and 31.8 µm, respectively). In thick slabs the maximum value will make up proportionally less, and in thin slabs proportionally more, of the measurements used to construct the mean value projection. We expect, then, that as slab thickness decreases the correlation between maximum and mean value projection will increase. In the limiting case where a slab is just a single pixel thick the two measurements coalesce.
This suggests that the relative advantage of the en face projection techniques depends on the thickness of the region being projected. To investigate this hypothesis, we examined the image quality metrics across the nerve fiber layer plexus (NFLP) [30]. The NFLP is thickest in the peripheral retina but tapers near the FAZ, providing an opportunity to examine anatomically similar material with varying thickness. In Fig. 3 we show SNR and RMS contrast against slab thickness in the NFLP; the SNR and RMS contrast values were calculated for concentric regions extending increasingly further from the FAZ in 20 pixel intervals ( Fig. 3(a)). Here again the maximum value projection outperforms the mean value, and we also observe the convergence of SNR and RMS contrast for the projection as the slab thins. Yet even at a relatively thin slab (4.78 pixels) maximum value projection continues to achieve superior performance. And, perhaps more importantly, each projection technique yields metrics that are correlated with slab thickness (to two significant figures, mean projection: RMS contrast correlation coefficient = 0.89, SNR correlation coefficient = 0.93; maximum value projection: RMS contrast correlation coefficient = 0.97, SNR correlation coefficient = 1.00). Since the correlation is stronger for maximum projection in both cases, we can expect that the thicker the slab, the more maximum projection will outperform mean value projection; or, similarly, we would need extremely thin slabs (i.e. somewhat less than approximately 5 pixels thick) to expect mean value projection to perform as well as maximum value projection in similar tissue.
We also examined en face projections in eyes with diabetic retinopathy (DR). DR often degrades OCTA quality, as evidenced by lower average SSI scores. Unfortunately DR affects decorrelation signal in more acute and pernicious ways than simply the reducing image quality, since disruptions like hyperreflective materials associated with exudation, such as hard exudates and suspended scattering particles in cysts, can introduce extravascular decorrelation signal into even otherwise avascular regions like the FAZ. For this reason the FAZ no longer serves as an adequate estimator for background in DR-the decorrelation signal in the FAZ can approach and indeed sometimes exceed the average decorrelation signal in the parafovea. Therefore restricting our analysis to just RMS contrast, we continue to find that maximum value projection outperforms mean value-yet by a smaller increment. Even given these diminishing returns in RMS contrast, DR is an example in which specific techniques may benefit from maximum value projection: vessel dilation can be used to accurately differentiate normal and DR eyes [31], and the higher contrast in maximum value projection is well-suited to this purpose.  on methods by We chose to mple threshold between the p reshold to segm hy eye populat reported in T s are statistica value projecti least in this co m. It is also wo e slightly more ns of choroidal n senting in a thin l na, yellow: CNV s/internal limiting line), and Bruch's jection, (a3) mean ar layer.
indicate decorrela ) Mean value pro 10% of pixels will re preserved from y comparing th examine vesse ding and needn projection meth ment vascular tion from Fig.  Table 2, as we ally significant ion producing ontext, has a n orth noting that e confident in t neovascularization layer. (a1) B-scan V, red: CC. These membrane (violet membrane/retinal n value projection ation value. (a) 3 × ojection. For each l be black, and the m Fig. 1   Although we so far limited our analysis to just maximum and mean value projection, it should be noted that these are only the most common techniques used to produce en face images in the literature, and not the only projection techniques possible or employed. Averaging suprathreshold voxels only or averaging the top 10the percentile, say, are also reasonable (and similar) approaches that are sometimes encountered. Actually we can consider any projection technique that averages successively larger populations of voxels as lying along a spectrum from mean projection (in which all voxels are averaged) to maximum value projection (in which just top voxel is "averaged"). In Fig. 7 we show results of projection achieved by averaging over cohorts in an A-scan in 20th percentile increments; i.e., projection in which all of voxels are included, then the top 80th percentile, then top 60th, etc. As the figure indicates, SNR and RMS contrast increase monotonically as the population of averaged voxels is reduced. This monotonic increase indicates that in terms of these metrics it is always best to include fewer members of the A-scan to form the projection. The limit of this series, of course, being to take just the maximum valued voxel.

Discussion and conclusion
Parenthetically, it should be noted that even the list of projection techniques just provided is not exhaustive. Vascular pixels need not be colored according to decorrelation value after projection at all. Depth mapping is a common approach in which vascular pixels are colored according to location along an A-scan (possibly with pixel brightness still indicative of decorrelation value). In this particular instance the location of the brightest voxel along the Ascan is physically meaningful, in contrast to something akin to a "center of mass" (i.e. a weighted sum of decorrelation values) that might select a location that does not correspond to any real feature in the slab. Depth mapping then is best performed as a sort of pseudomaximum value projection, in which the location of the maximum decorrelation voxel-rather than its value-is used to color the image. Of course pixels can convey more information that even just depth or decorrelation value, but other such pixel encodings are contextual and difficult to address in a systematic fashion.
Projections that select the value of a pixel from an operation over the voxel values are amenable to a general comparison, though. The en face OCTA image quality in terms of SNR and RMS contrast is clearly superior with maximum value projection across the tissue slabs and the disease states examined. Since both of these metrics are important for most if not all image analysis tasks, the maximum value en face projection should be used in absence of competing concerns.
Yet SNR and RMS contrast are just two quantitative metrics out of many that can and are used to characterize images. And, furthermore, as noted above, we are ultimately more interested in diagnostic parameters such as vessel density. Quantifying how the choice of projection techniques affects any specific calculation is beyond the scope of this report, but our results for vessel density give some indication that maximum value projection is also producing images in which we can be slightly more confident of further quantification. Examining just this single metric is suggestive but far from conclusive. We can, however, still take note of several qualitative features that may be important for image analysis. Qualitative features also usually lend themselves to maximum value projection. For instance, as indicated by the higher RMS contrast value captured by maximum value projection, capillaries are more easily demarcated from surrounding tissue-a considerable advantage for maximum value projection. Furthermore, capillaries appear to have their full length interrupted less by apparent noise fluctuations, indicative of a more accurate rendering of vasculature. And, finally, even when maximum value projection does incorporate noise in avascular regions into the flow signal, this noise is in general still easily discernable from the true flow signal (c.f. noisy pixels in the FAZ, Fig. 2 and 4).
Such qualitative concerns cannot be conclusively addressed in absence of concrete metrics to interrogate. One question is whether there exists a regime in which mean value projection may be useful or even outperform maximum value projection. Actually, we know that in individual images mean value projection can achieve superior SNR; it is only after averaging over populations that maximum value projection emerges a clear winner. But such pathological cases where the mean value projection obtains better SNR are typically low quality images anyway. From a logical standpoint, one should only be tempted by mean value projection in cases where most, if not all, of the A-scan is sampling just signal in the region of interest-else, using the mean projection only deteriorates signal quality by conflating it with measurements of noise. Alternatively, if the region of interest includes significant background (for instance when the vasculature is difficult to segment precisely) OCTA image quality will not be jeopardized by intrusions from randomly large decorrelation signal. Both of these conditions can pertain in CNV cases, were vasculature can be dense and background sampling of the CC (which abuts the CNV region of interest) can be problematic. And the CC itself may be an even better example-signal is plentiful and the anatomical layer is not thick. Still, at least in terms of RMS contrast, mean value projection underperforms maximum value projection even for imaging CNV or the CC, and, furthermore, features in both are just as easily discernable with maximum value projection. Actually, visual inspection reveals little intelligible variation between maximum and mean projection in the CC. In absence of a concrete qualitative difference between the approaches, the advantage of maximum projection in terms of RMS contrast offers a means to decide between otherwise seemingly equivalent choices. Ultimately, our results indicate that ideal conditions for mean projection do not pertain frequently enough in OCTA imaging to warrant much consideration, and even when they do, they do not render an obviously superior OCTA image.
Our investigation used the decorrelation signal obtained from the SSADA algorithm to construct en face angiograms, but SSADA is of course not the only OCTA algorithm available. A natural question is whether our results are transferrable to other OCTA images constructed using different algorithms (for instance optical microangiography, OMAG [7]). Although we did not investigate these alternative algorithms directly, the underlying logic that supports our conclusions is unaltered, and so we expect that maximum value projection should perform better in general, independent of the specific OCTA algorithm. It would be interesting to test this hypothesis in future work.
Since so many OCTA algorithms and diagnostic measures are concerned in the end with the vasculature it is paramount in most cases to construct the representation that can most readily distinguish between flow and non-flow pixels. SNR and RMS contrast are both excellent metrics to characterize this goal, and both indicate that maximum intensity projection is simply best suited to this task.

Disclosures
Oregon Health & Science University (OHSU), David Huang, and Yali Jia have a significant financial interest in Optovue, Inc. These potential conflicts of interest have been reviewed and managed by OHSU.