bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020.

The copyright holder for this preprint

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

1 Title (<120 characters): Measuring Compound Eye Optics with Microscope and MicroCT Images

2 Article Type: Tools and Resources

3 Author Names and Affiliations: John Paul Currea1, Yash Sondhi2, Akito Y. Kawahara3, and Jamie
4 Theobald2

5 1
Department of Psychology, Florida International University, Miami, FL 33199, U.S.A.

6 2
Department of Biological Sciences, Florida International University, Miami, FL 33199, U.S.A.

7 3
Florida Museum of Natural History, University of Florida, Gainesville, FL, 32611, U.S.A.

8 Abstract (<150 words):

9 The arthropod compound eye is the most prevalent eye type in the animal kingdom, with an impressive
10 range of shapes and sizes. Studying its natural range of morphologies provides insight into visual
11 ecology, development, and evolution. In contrast to the camera-type eyes we possess, external
12 structures of compound eyes often reveal resolution, sensitivity, and field of view if the eye is
13 spherical. Non-spherical eyes, however, require measuring internal structures using imaging
14 technology like MicroCT (µCT). Thus far, there is no efficient tool to automate characterizing
15 compound eye optics. We present two open-source programs: (1) the ommatidia detecting algorithm
16 (ODA), which automatically measures ommatidia count and diameter, and (2) a µCT pipeline, which
17 calculates anatomical acuity, sensitivity, and field of view across the eye by applying the ODA. We
18 validate these algorithms on images, images of replicas, and µCT scans from eyes of ants, fruit flies,
19 moths, and a bee.

1
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

20 There are about 1.3 million described arthropod species, representing about 80% of described animal
21 species (Zhang 2013). These arthropod species range vastly in size, lifestyle, and the habitat that they
22 live in. With body lengths ranging from the 85 µm ectoparasitic crustacean Tantulacus dieteri
23 (Mohrbeck, Arbizu, and Glatzel 2010) to the 0.5 m green rock lobster, Sagmariasus verreauxi
24 (Holthuis 1991; Kensler 1967), arthropods inhabit nearly every ecological niche. Along with their
25 diversity in body size and ecology is an array of eye architectures, the most common type being the
26 compound eye (Cronin et al. 2014; Land and Nilsson 2012). In addition, recent progress in
27 understanding compound eye development facilitates studying the selective pressures placed on the
28 development of the compound eye (Casares and McGregor 2020; Currea, Smith, and Theobald 2018;
29 Friedrich 2003; Gaspar et al. 2019; Harzsch and Hafner 2006). Because the compound eye found
30 among arthropods is the most prevalent eye type in the animal kingdom and includes an array of
31 morphologies that emerge following a relatively well-understood development, it is a prime subject in
32 the study of vision (Cronin et al. 2014; Land and Nilsson 2012).

33 Studying eye morphology is fundamental to understanding how animals see because eye structure
34 physically limits visual capacity (Land and Nilsson 2012). Depending on available light and image
35 motion, an optimal eye architecture exists to maximize the ability to differentiate brightness levels and
36 resolve spatial details (Land 1997; Snyder, Laughlin, and Stavenga 1977; Snyder, Stavenga, and
37 Laughlin 1977). Because of the critical role of eye morphology in vision, studying the natural diversity
38 of compound eye morphologies provides insight into visual ecology, development, and evolution. We
39 offer two open-source programs written in Python to support the effort of characterizing compound
40 eyes: (1) the ommatidia detecting algorithm (ODA), which automatically detects the individual facets,
41 called ommatidia, in 2D images, and (2) a multi-stage µCT pipeline, which applies the ODA to
42 segment individual crystalline cones and characterize the visual field.

43 In contrast to the camera-type eyes we possess, many structures that limit optical quality in compound
44 eyes are externally visible. For example, a compound eye is composed of many individual units called
45 ommatidia, each of which has a lens and a crystalline cone that direct light onto photoreceptors below.
46 Ommatidia usually represent the individual pixels of an image, so their number limits the total number
47 of images the eye can form, or its spatial information capacity, and they can be counted in microscope
48 images. They range from ~20 ommatidia in one of the smallest flying insects, the fairyfly Kikiki huna
49 (body length=158µm; Huber and Beardsley 2000; Huber and Noyes 2013) to >30,000 in large
50 dragonflies, which have enhanced vision for hunting prey (Cronin et al. 2014).

51 In each ommatidium, the size of the lens restricts its aperture, which then limits the ability to discern
52 luminance levels. High optical sensitivity is important for animals when photons are limited, such as
53 under dim light or fast image motion (Snyder, Laughlin, et al. 1977; Snyder, Stavenga, et al. 1977;
54 Theobald 2017; Warrant 1999). Compound eyes can be divided into two structural groups, apposition
55 and superposition eyes. In apposition eyes, light from each ommatidium is separated by light-absorbing
56 pigment cells and is restricted to a single rhabdom, such that lens size directly limits their optical
57 sensitivity (Figure 1A). In superposition eyes, light is directed through a clear zone, so that many facets
58 contribute to each point in the image (Figure 1B), thereby multiplying the final optical sensitivity
59 (although still subject to the diffraction limit of the individual apertures).

2
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

60 Thus far, there has been little effort to develop tools and workflows to calculate ommatidia count and
61 diameter, even though they can be derived from 2D images. Several algorithms and software plugins
62 have been proposed to help extract ommatidia but require user input for each image and underestimate
63 ommatidia counts (Woodman, Todd, and Staveley 2011), overestimate ommatidial diameter (Schramm
64 et al. 2015), or were not validated against manual measurements or measurements in the literature
65 (Diez-Hermano et al. 2015; Iyer et al. 2016; Vanhoutte, Michielsen, and Stavenga 2003). In addition,
66 none of the programs were tested on more than one species, over a substantial range of eye sizes, or on
67 different media. To support this effort, we offer an open-source program written in Python called the
68 Ommatidia Detecting Algorithm (ODA). We test the reliability and validity of this technique on single
69 images of 6 eye molds of 5 different ant species ranging dramatically in size, microscope images of 39
70 fruit flies (Drosophila melanogaster), and processed images of MicroCT (µCT) scans of 2 moths
71 (Manduca sexta and Deilephila elpenor) and one bee (Bombus terrestris). The method is widely
72 accessible because it only requires microscope imaging.

73 For spherical eyes, the lens diameter measurement provided by the ODA can be combined with
74 measurements of eye radius (using the luminous pseudopupil technique, for instance) to measure the
75 angular separation of ommatidia, called the interommatidial angle (IO angle; Figure 1 Δφ). The IO ). The IO
76 angle inversely limits spatial acuity or resolution (Land 1997; Snyder, Laughlin, et al. 1977; Snyder,
77 Stavenga, et al. 1977). High spatial acuity affords many behaviors, such as prey, predator, and mate
78 detection, and perceiving small changes in self-motion (Land 1997; Land and Nilsson 2012). For
79 spherical eyes, the IO angle is approximately: Δφ). The IO = D/R, where D is the ommatidial lens diameter and
80 R is the radius of curvature, assuming the axes of all ommatidia converge to a central point. Because of
81 the limits imposed by photon noise and diffraction, smaller eyes face increasing constraints on IO angle
82 and ommatidial size (Snyder, Stavenga, et al. 1977), generally resulting in homogeneous, spherical
83 eyes. Likewise, because optical superposition eyes form an image by superposing light from many
84 ommatidia on the retina below (Figure 1B), they often require a roughly spherical design (Land and
85 Nilsson 2012). Thus, many compound eyes are safely approximated by the spherical model.

86 However, many compound eyes are non-spherical, having ommatidial axes askew to the eye surface.
87 Skewed ommatidia sacrifice sensitivity by reducing their effective aperture as a function of the
88 skewness angle and refraction. Further, skewness can improve acuity at the expense of field of view
89 (FOV) by directing more ommatidia over a small visual field, decreasing the effective IO angle (Figure
90 1C). Alternatively, skewness can increase FOV at the expense of acuity by spreading a few ommatidia
91 over a large visual field (Figure 1D).

92 MicroCT (µCT) is a 3D X-Ray imaging technique that provides a rich and valuable resource to study
93 arthropod vision. Thus far, it has been used on arthropods to study muscles (Walker et al. 2014), brains
94 (Smith et al. 2016), ocelli (Taylor et al. 2016), and eyes (Brodrick et al. 2020; Gaspar et al. 2019;
95 Taylor et al. 2018, 2020). By visualizing internal structures in 3D, µCT affords calculating visual
96 parameters for non-spherical compound eyes, measuring anatomical IO angles at very high spatial
97 resolution, and segmenting different tissues, such as visual neuropils, within the same dataset.

98 Our second proposed method, the µCT pipeline, extracts crystalline cones from a µCT image stack
99 (using the ODA) to measure the size, orientation, and spatial distribution of ommatidia. As mentioned

3
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

100 above, we validate it on µCT scans of the spherical eyes of two moth species (M. sexta and D. elpenor)
101 collected by us and an open access scan of a non-spherical, oval bumblebee eye (B. terrestris) used in
102 Taylor et al. (2018). For the bumblebee eye, we further demonstrate how the µCT pipeline can test for
103 features of an oval eye; measure regional changes in skewness, spatial acuity, and sensitivity; and
104 project onto world-referenced coordinates to more accurately measure FOV.

105 Because there are so many arthropod species and compound eye morphologies, it is challenging but
106 valuable to characterize them in a meaningful, fast manner. Our proposed methods minimize the
107 substantial labor that is typically required in characterizing optical performance in compound eyes and
108 therefore facilitate understanding their role in visual ecology, development, and evolution.

4
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

109 Methods

110 Microscope Images

111 Eye Molds: Ants

112 Microscope images of eye molds or replicas were taken previously on five ant species. These species
113 were: Notoncus ectatommoides (N=2 images) and a Golden tail sugar ant (Camponotus aeneopilosus)
114 of the Formicinae subfamily, a jumper ant (Myrmecia nigrocincta) and a bull ant (M. tarsata) of the
115 Myrmeciinae subfamily, and Rhytidoponera inornata of the Ectatomminae subfamily. Images of the
116 first two species were obtained from Palavalli-Nettimi and Narendra (2018) and the latter three from
117 Palavalli-Nettimi et al. (2019). Manual measurements of ommatidial counts and diameters were taken
118 from their respective articles to compare with our automated measurements.

119 Direct Images: Fruit Flies

120 Microscope images of the eye of 39 fruit flies (Drosophila melanogaster) were obtained from the study
121 of Currea et al. (2018). Ommatidial counts and diameters were measured by hand and compared to our
122 automated measurements. Ommatidia were counted individually and diameters were measured as the
123 average of 9 different sets of 9 sequential diameters around the center of the eye. Three sets were
124 measured for each of the three major axes of the ommatidial array.

125 µCT

126 Moths

127 Micro-computed tomographs (µCTs) of two moth species, the tobacco hornworm (Manduca sexta) and
128 the elephant hawkmoth (Deilephila elpenor), were collected. Vouchered specimens from the Florida
129 Natural History Museum collections were stored in -20°C in 95% ethanol. Heads were sliced using
130 sterilized scalpels and left to soak in the staining solution in Eppendorf vials or falcon tubes. The
131 antennae were removed to allow for better stain penetration. Specimens were subjected to I2+KI
132 (created by mixing in equal proportions 1.25% I2 and 2.5% KI solutions) staining to increase soft
133 tissue contrast for 48–36 hours.

134 A fixed X-ray source was used to capture µCTs of the moth heads. The heads were placed in between
135 the x-ray source and a detector and were rotated by small angles to get a complete 360 degree view of
136 the sample. This generates a 3D X-ray image. Since Iodine is a heavy element, X-rays are absorbed
137 depending on the amount of stain a portion of the sample has absorbed, creating differences in intensity
138 in the final image stack.

139 The Manduca sexta specimen was scanned using a Phoenix V|Tome|X M system with a 180kv X-ray
140 tube containing a diamond-tungsten target, with a tube voltage 80 kV and current of 110 µA, a source
141 object distance of 17.8 mm and an object-detector distance of 793 mm and detector capture time
142 adjusted for each scan to maximize absorption range for the specimen. The acquisition consisted of
143 2300 projections of 8 sec each. Raw x-ray data were processed using GE’s proprietary datos|x r
144 software version 2.3 to produce a series of tomogram images and volumes, with final voxel resolution

5
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

145 of 4.50074 Sm. The resulting µCT volume files were imported into VG StudioMax version 3.3.3
146 (Volume Graphics, Heidelberg, Germany), the eyes isolated using VG StudioMax’s suite of
147 segmentation tools, and then exported as cleaned Tiff stacks.

148 The Deilephila elpenor specimen was scanned using a Zeiss Xradia 520 Versa (Carl Zeiss Microscopy
149 GmbH, Jena, Germany) using X-ray µCT, with a tube voltage 80 kV and current of 88 µA filtering of
150 the low energy (the filter composition and thickness is not disclosed by the manufacturer); a source
151 object distance of 22.5 mm and a object-detector distance of 210 mm; an indirect detector comprising a
152 scintillator, a 0.392x optical lens and a charge-coupled device camera. The acquisition was performed
153 using the software Scout-and-Scan (Carl Zeiss Microscopy GmbH), and consisted of 3201 projections
154 of 8 sec each and using the adaptive motion correction option. The tomographic reconstruction was
155 performed automatically generating a 32-bit txrm set of tomograms with an isotropic voxel size of
156 3.3250 µm. Finally, the data were converted to a stack of 16-bit tiff file using the XRM controller
157 software (Carl Zeiss Microscopy GmbH).

158 Bumblebee

159 A bumblebee (Bombus terrestris) eye was scanned previously using a synchrotron X-ray source. CT
160 imaging using a synchrotron source uses phase contrast in the samples between two slightly shifted X-
161 ray beams to identify the different densities present in a sample (Baird and Taylor 2017). This scan was
162 used with the permission of the author and it and other scans are available at
163 https://www.morphosource.org/Detail/ProjectDetail/Show/project_id/646. See Taylor et al., (2018) for
164 the details on the scans and specimens.

165 Ommatidia Detecting Algorithm (ODA)

166 The 2D Fast Fourier Transform (FFT) allows for the detection of 2D periodic patterns

167 The Fourier transform decomposes a function into a set of sinusoidal functions, each with its own
168 frequency, amplitude, and phase, allowing detection of periodic patterns. The sinusoidal elements of a
169 2D Fourier transform are called gratings, and have an additional parameter of orientation, the angle
170 perpendicular to the grating’s wavefront. For digital images, the 2D Fast Fourier Transform (FFT)
171 generates a 2D array of the same size, containing the component gratings, represented by their contrast,
172 frequency, phase, and orientation (Figure 2 A).

173 Because the FFT is invertible, operations can be applied to the frequency representation, called the
174 reciprocal space, then inverse transformed to generate a frequency filtered version of the original
175 image. In particular, if we want to target a periodic, grating-like pattern in a picture, we can filter the
176 relevant range of the spatial frequencies in the reciprocal space to amplify that pattern.

177 Ommatidia centers can be detected via a low-pass filter

178 The typical hexagonal arrangement of ommatidia has 3 major axes (Figure 2B blue, green and red).
179 Each axis is approximated by a grating in the reciprocal space. By filtering out frequencies
180 significantly higher than the three fundamental gratings, we can amplify the contribution of these
181 ommatidial axes. The inverse FFT generates a smooth version of the original image, with peaks at the

6
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

182 approximate center of each ommatidium. Ommatidia centers are found near the local maxima of this
183 inverse FFT image by using a local maxima detection algorithm.

184 ODA Python Module

185 A module written in Python was developed to apply the algorithm described above. A single image can
186 be imported and processed through the ODA: (1) 2D FFT the image, (2) find the three fundamental
187 gratings as local maxima closest to the center of the reciprocal image, (3) pass frequencies lower than
188 those of the fundamental gratings, (4) inverse 2D FFT the filtered reciprocal space, and (5) find local
189 maxima in the now smoothed image (Figure 2 B). Optionally, the user can view the generated
190 reciprocal image with the local maxima superimposed to check that the program worked appropriately.
191 The program stores a list of ommatidia coordinates in the image and calculates an average ommatidial
192 diameter as the average distance between neighboring coordinates around the center of the image.
193 Optionally, a binary mask (saved as a white silhouette on a black background) can be imported to avoid
194 counting false positives outside of the eye. This program is available at
195 www.github.com/jpcurrea/fly_eye/, where you can download the Python package and find some basic
196 examples on how to use it.

197 Measuring Ommatidia using µCT

198 The ODA was applied to a pipeline for processing 3D, specifically µCT, data. Here we describe the
199 fundamental elements of this pipeline.This program is available at
200 www.github.com/jpcurrea/compound_eye_tools/, where you can download the Python package and
201 find some basic examples on how to use it.

202 A. Import Stack

203 First, the large 3D dataset is reduced to coordinates found predominantly in the crystalline cones
204 (Figure 3 A). These filtered coordinates should be saved as a stack of images to be imported by our
205 program. Our program allows one to pre-filter data by choosing a range of density values in a graphical
206 user interface, ideally isolating the crystalline cones but often including other structures. These images
207 can be imported into an image editing program like ImageJ (Schneider, Rasband, and Eliceiri 2012) to
208 clean up the individual slides, deleting irrelevant pixels. Once imported, our program allows a preview
209 of the whole dataset (Figure 4 A).

210 B. Get Cross-Sectional Shells

211 Next, the coordinates are projected onto 2D images that can be processed through the ODA (Figure 3
212 B). Our program leverages the fact that the layer of crystalline cones is approximately a curved surface
213 with little variation normal to the surface of the eye. A best-fitting sphere is found for the whole dataset
214 using ordinary least squares regression and the sphere’s center is saved. Then, the coordinates are
215 transformed into spherical coordinates in reference to the sphere’s center, representing each point by its
216 radius, elevation, and azimuth.

217 In spherical coordinates, a continuous surface is fitted to the whole dataset, using linear interpolation to
218 model the points’ radii as a function of their elevation and azimuthal angles. With this fitted surface,

7
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

219 the coordinates are partitioned into two shells of points: those inside of the fitted surface (radii <
220 surface radii) and those outside the surface (radii >= surface radii).

221 C. Find Ommatidial Clusters

222 Next, an image of each shell is formed by ignoring differences in radii and taking a 2D histogram of
223 the coordinates’ elevation and azimuth (Figure 3 C). However, to avoid extreme spherical warping, the
224 coordinates are processed in 90°x 90° segments, each recentered before forming the histogram image.
225 The ODA is applied to each segment, determining the approximate center of each lens within that
226 segment. Then all of the centers are recombined onto the original spherical coordinate system by
227 finding the nearest groups of points between each shell. Finally, each point in the dataset is labeled by
228 its nearest center in each shell. By applying these labels to the original rectangular coordinates, the
229 dataset has been effectively segmented into clusters of points representing separate crystalline cones.

230 Skewed Ommatidia

231 Steps B and C work for spherical eyes but fail if ommatidia are substantially skewed. For instance, if
232 ommatidia are skewed by 45°, the offset between inside and outside shells can conflict with our ability
233 to match points from the same cluster. When a cluster in one shell fails to find a pair in the other shell
234 within a reasonable distance, that cluster is excluded from the count. To account for this, our program
235 has an option specifically for non-spherical ommatidia. Instead of the two shells in step B, we extract
236 one shell of points around the fitted surface containing 50% of the residuals (Figure 4 B). Then, as in
237 step C, we partition the eye into 90° x 90° segments and find the centers of the clusters within this shell
238 by applying the ommatidia finding algorithm to 2D histograms of the polar angles (Figure 4 C top and
239 middle). We use the number of ommatidia from the algorithm to apply the spectral clustering algorithm
240 to the total collection of points (Figure 4 C bottom). This method was successful for our bumblebee eye
241 but was prohibitively time consuming on the larger moth scans.

242 D. Measure Visual Parameters

243 The number of clusters approximates the ommatidia count, which limits the total number of images the
244 eye can form or its spatial information capacity. However, consider the number of ommatidial clusters
245 at various stages of the pipeline. The initial count occurs at the beginning of step C, when applying the
246 ODA, but is reduced if it fails to match the clusters of both shells for spherical eyes or if the clusters
247 fail to be detected by the clustering algorithm for non-spherical eyes. Determining which count to use
248 depends on the quality and resolution of the scan and can be settled by visual inspection of the cluster
249 centers superimposed on the shells’ 2D histograms.

250 The average distance between adjacent cluster centroids approximates the ommatidial lens diameter,
251 which ultimately limits the eye's sensitivity, or ability to discern luminance levels. The ideal
252 ommatidial axis is derived from the planes formed by triplets of centers near the cluster. By averaging
253 the normal vector for each plane, we can approximate the vector orthogonal to the surface of the eye at
254 each cluster. The anatomical ommatidial axis is derived by using singular value decomposition to find
255 the longest semi-axis of the ellipsoid formed by the cluster. By finding the angular difference between
256 the ideal and anatomical axes, we can measure the anatomical skewness of each ommatidium.

8
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

257 Ideally, the anatomical IO angles would be derived from the anatomical axes of neighboring
258 ommatidia. However, such results were highly variable, especially in the bee scan (for instance, many
259 of the resulting angles were negative). This was likely because the length of each cluster is small
260 compared to the distance to the point of intersection, so that angular deviations in either axis have
261 magnified effects on their intersection and thus interior angle. Getting such measurements to work
262 might require increasing the resolution of the scan or using other structures, like the rhabdom, to extend
263 our axis approximations closer to the intersection point.

264 Instead, anatomical IO angles are approximated by partitioning the coordinates into evenly spaced
265 vertical and horizontal sections (Figure 4 D). For each section, all clusters within that section are
266 projected onto a parallel plane. For instance, for vertical sections, all clusters within a range of x values
267 are considered and the 2D clusters formed by their y and z values are used to determine their vertical
268 IO angle component.

269 Then, the angle of the projected ommatidial axes are modeled as a function of their position, using
270 nonlinear least squares regression. The program fits a polynomial function of degree between 2 and 5,
271 only accepting a higher degree if it reduces the sum of the squared residuals by at least 5%. The
272 function is constrained to having a positive first derivative, based on the assumption that the IO angles
273 were never negative. This is repeated independently for each of the vertical and horizontal sections.

274 Once finished, the process approximates a horizontal and vertical angle for each cluster pair and we
275 calculate the total angle as their hypotenuse. We call this hypotenuse the anatomical IO angle. While
276 this method assumes one regression function per section, it allows for different regression models per
277 section and independent approximations for horizontal and vertical IO angles across the eye (Stavenga
278 1979). In addition, by keeping track of the orientation of the cluster pairs along the surface of the eye,
279 we can calculate the IO angles from horizontal and vertical IO angles using the two-dimensional lattice
280 proposed by Stavenga (1979; Figure 6C).

281 Finally, two spreadsheets are saved: (1) for each crystalline cone cluster, the Cartesian and spherical
282 coordinates of the centroid, the number and location of the points within its cluster, the approximate
283 lens diameter, and the ideal and anatomical axis direction vectors are saved per row; and (2) for each
284 pair of adjacent crystalline cones, the cones’ indexes from spreadsheet 1, the center coordinates for the
285 two cones, the resulting orientation in polar coordinates, and the anatomical IO angle between the two
286 are saved per row. With these metrics, researchers can approximate how spatial acuity, optical
287 sensitivity, and the eye parameter (Snyder 1979) vary across the eye (Taylor et al. 2018).

9
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

288 Results

289 Microscope Images

290 Eye Molds: Ants

291 We tested the ODA on images of flattened eye molds of six ants from 5 different species (Figure 5A).
292 Ommatidia number and lens diameter measurements from the algorithm compared well to those taken
293 manually on the same images. The automated counts were on average 97% (95.6 – 98.7%) of those
294 taken by hand and they shared a strong and significant correlation (r=.99, df=4, p≪.0001). The .0001). The
295 automated lens diameters were on average 99% (92.1 – 106.7%) of those taken by hand and they also
296 shared a strong and significant correlation (r=.82, df=4, p=.02).

297 Direct Images: Fruit Flies

298 We also tested the ODA on microscope images of the eyes of 39 fruit flies (D. melanogaster; Figure
299 5B and C), validating the algorithm’s performance on direct images of compound eyes and
300 comparisons within a species. Ommatidia counts from the algorithm compared well to those taken
301 manually on the same images. Automated counts shared a strong and significant correlation with
302 manual counts (r=.93, df=37, p≪.0001). The .0001). Automated counts were 845±87 (mean±standard deviation)
303 while hand counts were 810±76, with automated counts equalling 104±4% of those taken by hand.
304 Though the difference in counts was not significant (t=1.8, df=37, p=0.07), automated counts may
305 exceed the hand counts because of partial ommatidia that were excluded during the manual counting
306 process but not the algorithm.

307 Lens diameter measurements from the algorithm also compared well to those taken manually on the
308 same images. Automated diameters shared a strong and significant correlation with manual diameters
309 (r=.82, df=37, p≪.0001). The .0001). Automated diameters were 16.3±1µm while manual diameters were
310 15.6±.8µm, with automated diameters equaling 105±3% of those taken by hand and is close to the
311 16µm taken previously (Franceschini 1972). Automatic diameters were greater than manual diameters
312 (t=4, df=37, p<.001) possibly because manual diameters were calculated by taking the average over
313 multiple sequences of ommatidia. This assumed the ommatidia centers are arranged in perfectly
314 straight rows, though they curve slightly due to the curvature of the eye and can be interrupted by
315 lattice abnormalities (Ready, Hanson, and Benzer 1976). By measuring individual diameters, the
316 automatic diameters can account for irregularities in the ommatidial lattice and should equal at least
317 those taken by averaging over a straight distance.

318 µCT

319 We tested our µCT pipeline on scans of compound eyes of two moth species (M. sexta and D. elpenor)
320 collected by us and an publicly available scan of a bumblebee (B. terrestris) used in Taylor et al. (2018;
321 Figure 7). The resulting measurements can be found in the table of Figure 7B. In addition, we tested
322 methods specifically for non-spherical eyes on the bumblebee scan (Figure 8), including testing for
323 features of an oval eye, measuring regional changes in aperture and anatomical IO angle, and
324 projecting onto world-referenced coordinates.

10
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

325 Manduca sexta

326 We found that the M. sexta eye was composed of 27,756 ommatidia initially (µCT pipeline step B) but
327 the dataset was reduced to 25,159 after later processing (step C). Visual inspection suggested that the
328 first count is more accurate because the stack images were thoroughly cleaned before processing. The
329 two counts are 103% and 93% of the previous approximation of 27,000 (White 2003) and 108% and
330 98% of the previous count of 25,641 taken on the same species (Stöckl, O’Carroll, and Warrant 2017).
331 These ommatidia had lens diameters of 32.5±1.5µm, which is within range of previous measurements
332 on the same species: 30µm (White 2003), 31µm (Stöckl et al. 2017), and 34µm (Theobald, Warrant,
333 and O’Carroll 2010). The ommatidial axes were skewed by absolute angles of 5.7±4.4°
334 (median±interquartile range), which are minor deviations potentially due to error in estimating the
335 ommatidial axes or deformations of the eye during the µCT preparation procedure. We can calculate
336 the minimum effect of this skewness on lens diameter by multiplying by the cosine of the skew angle,
337 though the true angle is further skewed by refraction of the image space (Stavenga 1979). The
338 adjustment without accounting for refraction resulted in a reduction of 0.16±.25µm or 0.5±.8%, which
339 is minimal and consistent with a spherical eye. The ommatidia were separated by an anatomical IO
340 angle of 0.94±.33°, which is close to our spherical approximation of .88±.16° and within the range of
341 previous measurements of 0.91° (Stöckl et al. 2017) and 0.96° (Theobald et al. 2010) on the same
342 species.

343 Deilephila elpenor

344 We found that the D. elpenor eye was composed of 13,801 ommatidia initially (step B) but the dataset
345 was reduced to 12,172 after later processing (step C). Visual inspection suggests that there are
346 substantial false positives in the first count because of noise around the boundary of the eye that we
347 failed to remove in pre-processing. The two counts are 120% and 106% of the count of the previous
348 approximation of 11,508 on the same species (Stöckl et al. 2017). These ommatidia had lens diameters
349 of 28.8±1.9µm (mean±s.d.), which is consistent with previous measurements of 28µm (Stöckl et al.
350 2017) and 29µm (Johnsen 2006; Theobald et al. 2010) on the same species. The ommatidial axes were
351 skewed by an absolute angle of 6.6±5.1° (median±IQR). The effect of this skewness on effective lens
352 diameter was a reduction of .19±.31µm or .7±.11%, which again is minimal and consistent with a
353 spherical eye. The ommatidia were separated by an anatomical IO angle of 1.49±0.74°, which is close
354 to the spherical approximation of 1.27±0.35° and previous measurements of 1.12° (Stöckl et al. 2017)
355 and 1.31° (Theobald et al. 2010) on the same species.

356 Bombus terrestris

357 We found that the B. terrestris worker bee eye was composed of 4,765 ommatidia initially (step B) but
358 the dataset was reduced to 3,562 after later processing (step C*). Visual inspection of the 2D
359 histograms and the superimposed ommatidial centers suggested that the first count is more accurate
360 than the second, though it missed many ommatidia. The large hole in the eye map of Figures 4 and 6 is
361 mostly due to differential exposure across the eye, resulting in many crystalline cones that were
362 excluded at the pre-filter stage. The two counts are 85% and 64% of the 5,597 density-based count of
363 the same scan (Taylor et al. 2018), and 85% of the previous count of 5,624±369 (mean±s.d.) for

11
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

364 conspecific worker bees counted manually (Streinzer and Spaethe 2014).

365 Despite missing a substantial number of ommatidia, our program provided some valid insights. The
366 ommatidia had an average lens diameter of 26.1±2.1µm (mean±s.d.), which is consistent with the
367 previous manual measurement of 25.6±2.5µm on the same scan (Taylor et al. 2018) and 25.1±1.9µm
368 for worker bees of the same species (Streinzer and Spaethe 2014). The ommatidial axes were skewed
369 by an absolute angle of 29.2±18.0° (median±IQR), which is consistent with previous anatomical
370 measurements of skewness angles up to 50° (Baumgärtner 1928). The effect of this skewness on
371 adjusted lens diameter was a reduction of 3.30±3.92µm or 12.7±15.4%, which is substantial and
372 consistent with a non-spherical eye. The ommatidia were separated by an anatomical IO angle of
373 2.29±1.65°, which is greater than the spherical approximation of 0.74±.27°, the previous measurement
374 of 1.6±0.6° using orthogonal ommatidial axes on the same scan (Taylor et al. 2018), and the previous
375 measurement of 1.9° using the pseudopupil technique on a similarly sized conspecific (Spaethe 2003).

376 World-Referenced Visual Field

377 Above we used spherical coordinates for the azimuth and elevation about a single center point (step B).
378 For the bee eye, this mostly accounted for curvature along the vertical axis but vastly underestimated
379 curvature along the horizontal axis. To get a better measurement of their visual field, we projected the
380 ommatidial axes onto a sphere outside of the eye, much like the world-referenced projection of Taylor
381 et al. (2018; Figure 7C). However, instead of using the center of the head as our center, we used the
382 center found in step B, which is near but not exactly the center of the head. We also chose a radius of
383 10cm based on their visual fixation behavior (Wehner and Flatt 1977). This world-referenced visual
384 field has a FOV of about 140° horizontally and 140° vertically, agreeing exactly with previous
385 measurements on the same scan, which did not account for skewness (Taylor et al. 2018).

386 Horizontal and Vertical Spatial Resolution

387 The bee eye is a non-spherical, oval eye, in which ommatidial axes intersect at different points for
388 horizontally and vertically oriented IO pairs (Ih and Iv in Figure 6C). Anatomical IO angles therefore
389 should be separated into independent horizontal and vertical components (Δφ). The IO h and Δφ). The IO v in Figure 6C).
390 To apply this lattice configuration we converted our measurements of horizontal and vertical subtended
391 angles from Method 6.4 into Δφ). The IO h and Δφ). The IO v and tested the accuracy of this oval eye model. In particular,
392 if the eye is oval, (1) the horizontal angle for horizontal IO pairs (0° orientation) equals 2Δφ). The IO h while the
393 vertical angle equals 0; (2) the horizontal and vertical angles for diagonal IO pairs (±60° orientations)
394 are Δφ). The IO h and Δφ). The IO v; and (3) the proportion Δφ). The IO v /Δφ). The IO h should equal 1/ √ 3. (Stavenga 1979)

395 In our bee eye dataset, IO pair orientations followed a trimodal distribution with modes at -59°, 0°, and
396 61° (K Means algorithm with 3 means; N=11,626, mean distance=12°). We selected IO pairs within
397 15° of these centroids to more accurately measure the horizontal and vertical angles (Figure 7A). For
398 horizontal pairs (orientation=0.0±15° ), the horizontal subtended angle was 2.94±1.08° while the
399 vertical angle was 0.08±.10° (N=2,675). For negative diagonal pairs (orientation=-59±15°), the
400 horizontal angle was 1.90±.81° while the vertical angle was .97±0.50° (N=2,435). For positive
401 diagonal pairs (orientation=61±15°), the horizontal angle was 1.36±0.60° while the vertical angle
402 was .98±0.49° (N=2,393). As a result, our dataset fits several of the predictions of an oval eye: (1) the

12
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

403 horizontal angle for horizontal IO pairs, 2.94°, is nearly twice the horizontal angle for diagonal pairs,
404 1.90°+1.36°=3.26° while the vertical angle, 0.08°, is close to 0; and (2) the vertical angles for diagonal
405 IO pairs, .97° and .98°, are nearly equal. However, the horizontal angles for diagonal IO pairs, 1.90°
406 and 1.36°, which should be equal, differ by about 28%.

407 By combining the horizontal angle of diagonal pairs and half the horizontal angle of horizontal pairs,
408 we approximate the horizontal anatomical IO angle component, Δφ). The IO h=1.48±.88° (N=7,503). This is
409 close to the range of 1.8°–3.3° measured previously (Spaethe 2003). By combining the vertical angles
410 of diagonal pairs we approximate Δφ). The IO v=.95±0.60° (N=4,828), which is within the range of 0.6°–1.4°
411 measured in Spaethe (2003). Using these two anatomical IO angle components, the total anatomical IO
412 angle Δφ). The IO = 1.76°, is closer to the previous measurements of 1.6° and 1.9° mentioned above. Also, the
413 proportion Δφ). The IO v/Δφ). The IO h=.64 is close to 1/ √ 3=0.58, as predicted for an oval eye.

414 The anatomical IO angle components also varied as a function of elevation and azimuth angles (Figure
415 7B and C). To measure the change in these parameters with respect to azimuth and elevation, we
416 binned values into 15° intervals. Δφ). The IO v decreases horizontally from a maximum of 1.1±0.33° at -37.5°
417 azimuth (bin N=571) to a minimum of 0.80±1.02° at 7.5° azimuth (N=705), and then increases to a
418 local maximum of .94±0.33° at 37.5° azimuth (N=614). This corresponds to a weak negative
419 correlation (r=-.14, df=2174, p≪.0001). The 0.0001) from -37.5° to 7.5° azimuth followed by an insignificant
420 positive correlation up to 37.5° (r=0.01, df=1,414, p=0.60; Figure B, top left). To our knowledge, this
421 has not been measured previously but is consistent with the observation that the greatest overall acuity
422 is in the central visual field (Baumgärtner 1928).

423 Δφ). The IO v decreases vertically from a maximum of 1.78±.74° at -67.5° elevation (N=53) to a minimum of
424 0.51±0.63° at 22.5° azimuth (N=1294), and then increases to a local maximum of 1.21±0.67° at 52.5°
425 azimuth (N=57). This corresponds to a strong negative correlation (r=-0.53, df=4870, p≪.0001). The 0.0001)
426 followed by a moderate positive correlation (r=0.29, df=2207, p≪.0001). The 0.0001; Figure B, top right). This
427 agrees partially with previous measurements showing a decrease from about 2.4° ventrally to about 1°
428 centrally followed by a gradual increase to about 2° dorsally. However, the previous measurement
429 found a dramatic increase to 4° at the dorsal extreme (Baumgärtner 1928), which we did not measure.
430 Differences in these measurements are likely due to technical differences: their bee species is
431 unspecified and may differ from B. terrestris used here; their specific bee specimen may be smaller and
432 thus have a larger Δφ). The IO v minimum than ours; and our eye scan lacks a substantial dorsal portion where
433 they measured most of the increase in Δφ). The IO v. (Baumgärtner 1928)

434 Δφ). The IO h decreases from a maximum of 2.23±2.25° at -67.5° azimuth (N=52) to a relatively constant range
435 from 1.40±0.88° at -22.5° azimuth (N=1,178) to 1.55±0.82° at 52.5° azimuth (N=455). This actually
436 corresponded to a weak negative correlation (r=-0.03, df=6295, p<0.01). This differs from the previous
437 measurements taken on the same scan, which found that the anatomical IO angle increases from about
438 1.26° to 1.90° from the lateral to central eye with a strong positive correlation (r=0.50, df=3,471,
439 p≪.0001). The 0.0001; Taylor et al., 2018). This is likely because the previous measurements did not account for
440 skewness which we find increases laterally, explaining partially why our anatomical IO angle
441 measurement exceeds theirs by about 0.16°.

13
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

442 Finally, Δφ). The IO h decreases vertically from a maximum of 2.07±1.11° at -52.5° elevation (N=274) to a
443 minimum of 1.49±0.74° at 7.5° elevation (N=1,528), with an insignificant correlation (r=0.03,
444 df=4,750, p=0.06; Figure 7B, bottom right).

445 Sensitivity, Acuity, and FOV Tradeoffs

446 Adjusted lens diameter also varied as a function of azimuth and elevation angle. Adjusted lens diameter
447 increases horizontally from a minimum of 19.9±6.7µm at -37.5° azimuth (N=476) to a maximum of
448 24.1±3.1µm at 37.5° azimuth (N=430), with a moderate positive correlation (r=0.28, df=3,542,
449 p≪.0001). The 0.0001; Figure 7B, bottom left, blue). This correlation is largely reduced for the absolute lens
450 diameter (r=0.09, df=3,542, p≪.0001). The 0.0001; Figure 7B, bottom left, blue dashed line).

451 Adjusted lens diameter also increases vertically from a minimum of 18.5±4.0µm at -67.5° elevation
452 (N=100) to a maximum of 23.7±3.9µm at 7.5° azimuth (N=559; Figure B, bottom left, blue). This
453 corresponds to a moderate positive correlation below 7.5° elevation (r=0.25, df=1,845, p≪.0001). The 0.0001),
454 which is completely reduced for the absolute lens diameter (r=-0.006, df=3,542, p=0.8; Figure 7B, top
455 right, blue dashed line).

456 These positional changes in adjusted lens diameter and anatomical IO angle components reveal
457 interesting tradeoffs between FOV, sensitivity, and the horizontal and vertical spatial acuities caused by
458 skewness across the eye. As we stated in the introduction, skewness always reduces sensitivity but can
459 lead to FOV and IO angle tradeoffs. In this case, adjusted lens diameter, and consequently sensitivity,
460 increases from the lateral to central visual field while vertical spatial acuity first increases slightly until
461 near the center and then decreases medially. Similarly, adjusted lens diameter, and consequently
462 sensitivity, increases from the ventral to dorsal visual field while vertical spatial acuity also increases.
463 This results in high sensitivity and vertical spatial acuity near the central visual field at the expense of
464 FOV, and high FOV laterally and ventrally at the expense of sensitivity and vertical spatial acuity.
465 Both of these effects are largely invisible based on absolute lens diameters or IO angles that do not
466 account for skewness. Hence, the regional tradeoffs and specializations are due to ommatidial
467 skewness. These same tradeoffs are found among many apposition eyes, though they usually coincide
468 with visible changes in the curvature of the eye. These results highlight the flexibility gained by
469 skewing ommatidia as well as their deceptive nature.

470 Skewness-Based Eye Regions

471 Skewness angle varied systematically across the bee eye (Figure 7C). In particular, there were 3
472 regions of generally different skewness angles (separated by the dashed white lines in Figure 7C): (1)
473 small angles centrally and medial-dorsally (19±10°) and (2) moderate angles dorso-laterally (35±13°),
474 and (3) ventrally (37±14°). This skewness resulted in 3 regions of generally different adjusted lens
475 diameters: (1) larger diameters centrally and medially (24±3µm); (2) moderate diameters dorso-
476 laterally (22±4µm); and (3) small diameters ventrally (20±5µm).

477 While Δφ). The IO h remained constant across these three regions (Δφ). The IO h=[1.5±0.8°, 1.5±0.7°, 1.5±1.1°], N=[3048,
478 1626, 2765]), Δφ). The IO v varied. (1) The lowest vertical IO components, Δφ). The IO v=0.81±0.52° (N=2,005), resulting
479 in the lowest total anatomical IO angle, Δφ). The IO =1.7°, were in the medial-dorsal region of lowest skewness.

14
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

480 (2) The largest vertical IO angles, Δφ). The IO v=1.3±0.7° (N=1,041), resulting in the largest total anatomical IO
481 angle, Δφ). The IO =1.98°, were in the dorso-lateral region of moderate skewness. (3) Moderate vertical IO
482 angles, Δφ). The IO v=0.97±0.56° (N=1,740), resulting in the second lowest total anatomical IO angle, Δφ). The IO =1.75°,
483 were in the ventral region of maximum skewness.

484 Although regions 2 and 3 have nearly the same amount of skewness, 35° and 37°, their anatomical IO
485 angles, 1.98° and 1.75°, and thus spatial resolutions differ by about 12%. As we mentioned in the
486 introduction, this is because skewness confers a trade-off between FOV and spatial acuity. As a result,
487 the skewness in region 2 extends the FOV dorsally, while in region 3 it maintains a moderate spatial
488 resolution.

489 Also, the region of minimal skewness, largest adjusted lens diameters, and smallest anatomical IO
490 angles (region 1) coincides with two visual regions of ecological significance. First, during visual
491 fixation, bees incline their dorsal-ventral axis to 70° (Wehner and Flatt 1977), adjusting our elevation
492 parameter by -20°. Thus they direct the approximate center of this region (azimuth=0°, elevation=20°),
493 which is near the maximum adjusted lens diameter and minimum anatomical IO angle, towards visual
494 targets. Second, binocular overlap occurs in the dorsal part of this region (Taylor et al. 2018). As a
495 result, the region involved in visual fixation and binocular overlap has the lowest average skewness and
496 the highest sensitivity and acuity.

15
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

497 Discussion

498 The methods proposed here successfully automate the estimation of multiple visual parameters of
499 compound eyes. We tested the ommatidia detecting algorithm (ODA), which filters spatial frequencies
500 based on the hexagonal arrangement of the ommatidial lattice and applies a local maximum detector to
501 identify the centers of ommatidia in images of compound eyes. We applied this algorithm to calculate
502 the ommatidial lens diameter and count from different media (eye molds, microscope images, and µCT
503 scans), taxa (ants, flies, moths, and a bee), sizes (hundreds to tens of thousands of ommatidia), and eye
504 type (apposition, neural superposition, and optical superposition). In all cases, measurements provided
505 by the program matched well with manual measurements on the same data, previous measurements in
506 the literature, or both. Ommatidial counts were accurate, with an average proportion across all tested
507 species of 93–102% (mean of low means – mean of high means%) of previous or manual
508 measurements. Ommatidial diameters were even more accurate, with an average proportion across all
509 tested species of 100.2–103.4% of previous or manual measurements.

510 Integrating the ODA into a larger µCT pipeline proved successful on scans of two spherical moth eyes
511 (D. elpenor and M. sexta) and one oval bee eye (B. terrestris). In addition to ommatidia counts and lens
512 diameters, the pipeline provided estimates of anatomical IO angles, FOV, and skewness. This offered
513 the calculation of an adjusted lens diameter, approximating the aperture-diminishing effect of skewed
514 ommatidia. As expected for spherical eyes, skewness angles were insignificant in the two moth eyes,
515 which generally require spheracity for proper optical superposition. However, when applied to the oval
516 eye of the bumblebee, there were significant skewness angles, implying a reduction in optical
517 sensitivity. Again, the estimates provided were consistent with manual measurements on the same data,
518 previous measurements in the literature, or both.

519 On the bumblebee eye, we demonstrated additional advantages of high resolution 3D data, world-
520 referenced coordinates, and regional measurements across the visual field. These included the
521 measurement of regional changes in skewness, vertical spatial acuity, and sensitivity as well as regional
522 three-way tradeoffs between spatial acuity, FOV, and sensitivity caused by skewness. Though many of
523 these insights were consistent with previous studies and measurements on the same scan, some were
524 previously unknown.

525 One example reveals an unexpected advantage of using 3D data. Traditional methods of measuring IO
526 angle facilitate measurements oriented along the same axis but make it challenging to measure IO
527 angles oriented along a perpendicular axis. For example, when using histological sections, it is easier
528 and more efficient to measure multiple IO angles parallel to the sectioning plane than to measure
529 individual IO angles perpendicularly along many sections. With a 3D dataset, however, the two are
530 equally difficult. As a result, across the bee eye we were able to measure horizontal resolution
531 increasing vertically and vertical resolution increasing horizontally from the peripheral to the central
532 visual field. As far as we know, this has not been measured previously though it is consistent with
533 previous measurements showing the greatest resolution in the central visual field (Baumgärtner 1928).

534 In addition to the empirical benefits of our programs, they provide several technical advantages by
535 design. First, they are both open source and easy to install. The ODA can be downloaded as a Python

16
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

536 module (https://github.com/jpcurrea/fly_eye) and we are in the process of adding several command-
537 line tools for those who are not programming savvy. The µCT pipeline can also be downloaded as a
538 Python module (https://github.com/jpcurrea/compound_eye_tools) with command-line and graphical
539 user interfaces available to run on individual image stacks. Both have been designed in a modular
540 fashion so that collaborators can easily incorporate portions of the library into other, custom pipelines.
541 For instance, the pipeline can be easily customized to run on a computer cluster.

542 The two programs presented here have some limitations. First, the ODA requires images with sufficient
543 resolution in order to properly detect ommatidia. For single images, this requires that the pixel length
544 be at most half the smallest ommatidial diameter. For the µCT pipeline, this is ultimately affected by
545 the resolution of the scan: if individual crystalline cones cannot be resolved at each layer, they are
546 likely indiscriminable once fed into the ODA. Also, some species, preparations, and µCT scanning
547 procedures appear better than others at capturing the contrast between crystalline cones and the other
548 structures while avoiding structural damage to the specimen. For instance, crystalline cones in the M.
549 sexta scan were prefiltered simply by using the threshold function because the crystalline cones
550 contrasted sharply with the background. This was not the case for D. elpenor, which had additional
551 noise outside of the eye, and B. terrestris. The B. terrestris scan had a particular problem with uneven
552 exposure, resulting in different density measurements depending on the location in the scan and
553 ultimately the omission of data. This may be an unintentional consequence of the contrast enhancing
554 property of a synchrotron light source (Baird and Taylor 2017).

555 Ultimately, anatomical measurements cannot replace optical techniques in measuring compound eye
556 optics (Stavenga 1979; Taylor et al. 2018). When light passes through the eye optics, the angle of
557 refraction is determined by both the incident angle and the index of refraction of the image space
558 (Stavenga 1979). In our approximations, however, we used only the incident angle. As a result, our
559 measurements of the aperture-diminishing effect of skewness represent lower bounds and our program
560 specifically provides measurements of anatomical IO angles as opposed to functional IO angles.
561 Ultimately, though skewness can be somewhat corrected for, no methods are as accurate as optical
562 techniques (Stavenga 1979).

563 Because of their great range of eye morphologies (Land and Nilsson 2012), and both inter- and
564 intraspecific variation in size (Casares and McGregor 2020; Currea et al. 2018; Gaspar et al. 2019;
565 Land 1997; Land and Nilsson 2012; Shingleton et al. 2007), compound eyes are ideal candidates for
566 studying environmental reaction norms and allometry. Very little research on allometry has dealt with
567 compound eyes, generally resorting to the allometry of organs that can be measured in one or two
568 dimensions (McDonald et al. 2018; Shingleton et al. 2007; Shingleton, Mirth, and Bates 2008). Our
569 programs automate the more tedious tasks of characterizing compound eyes and therefore should help
570 with this challenge. For instance, ommatidia counts and diameters derived from the ODA allow for
571 total cell count approximations and correspond to the independent effects of cellular proliferation and
572 growth during eye development. Further, our program facilitates measuring the allometry of visual
573 performance, allowing us to measure the environmental reaction norms of the anatomical determinants
574 of vision.

575 Characterizing compound eye morphology also applies to research in visual development, plasticity,

17
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

576 neurobiology, and neuroethology. Thanks in large part to progress in understanding eye development in
577 fruit flies (D. melanogaster; Callier and Nijhout 2013; Casares and McGregor 2020; Gaspar et al. 2019;
578 Ready et al. 1976), compound eyes are ideal for assessing principles of eye development and how they
579 may vary across different taxa (Casares and McGregor 2020; Friedrich 2003; Harzsch and Hafner
580 2006). Also, because eye optics are the first limit to incoming visual information (Land 1997; Stavenga
581 1979), they inform behavioral or electrophysiological data to infer intermediate neural processing
582 (Currea et al. 2018; Gonzalez-Bellido, Wardill, and Juusola 2011; Juusola et al. 2017; Land and
583 Nilsson 2012; Theobald et al. 2010; Wardill et al. 2017; Warrant 1999).

584 Conclusion

585 Arthropods are the most diverse eukaryotic phylum and occupy nearly every ecological niche on Earth.
586 They possess the most common eye type, the compound eye, which varies dramatically in size, shape,
587 and architecture across widely diverse visual environments. Because compound eyes are diverse,
588 ubiquitous, and have undergone substantial natural selection, studying their natural range of
589 morphologies and optical limitations is imperative to understanding the evolution and development of
590 vision. Our programs successfully measure optical parameters of a wide range of compound eye shapes
591 and sizes, minimizing labor and facilitating the study of the development, evolution, and ecology of
592 visual systems in non-model organisms.

593 Despite the programs we present here, more work is needed to understand the limitations of both the
594 ODA and the µCT pipeline. Because the ODA depends on passing the spatial frequencies
595 corresponding inversely to the ommatidial diameter, an eye with a wide range of diameters may not
596 work. The ODA should be tested on images of eyes with highly acute or sensitive zones, such as the
597 robberfly (Wardill et al. 2017) and ommatidia lattices with regions of transitioning arrangements, such
598 as the transition from hexagon to square in houseflies (Stavenga 1975) and male blowflies (Smith et al.
599 2015). Likewise, the µCT pipeline should be tested on non-spherical non-oval eyes. While our program
600 appropriately measured anatomical IO angles across the bumblebee’s oval eye and actually
601 corroborated its oval eye properties, it may not work when IO angles change dramatically like in the
602 robberfly (Wardill et al. 2017). Finally, the µCT pipeline should be tested on non-insect arthropods.

603
604 Funding: This research was supported by grants from the National Science Foundation: IOS-1750833
605 to JT, BCS-1525371 to JC, DEB-1557007 to AYK and JT and IOS-1920895 to AYK. The content is
606 solely the responsibility of the authors and does not necessarily represent the official views of the
607 National Institutes of Health or the National Science Foundation.

608 Acknowledgements: We thank members of the Kawahara Lab at the Florida Museum of Natural
609 History, McGuire Center for Lepidoptera and Biodiversity, for raising and vouchering moth specimens.
610 The University of Florida CT facility is thanked for helping acquiring the M. sexta CT scan, and Kelly
611 Dexter for help with sample staining. We thank Deborah Glass and Ian Kitching from the Natural
612 History Museum of London for help with the CT procedure and sharing the D. elpenor CT scan, which
613 study was funded by NERC grant number NE/P003915/1 to IJK. Gavin Taylor and Emily Baird shared
614 their CT scans of B. terrestris. Michael Reisser and Arthur Zhao provided valuable discussions

18
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

615 regarding the uCT pipeline.

19
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

616 References

617 Baird, Emily, and Gavin Taylor. 2017. “X-Ray Micro Computed-Tomography.” Current Biology
618 27(8):R289–91. doi: 10.1016/j.cub.2017.01.066.
619 Baumgärtner, Herbert. 1928. “Der Formensinn Und Die Sehschärfe Der Bienen.” Zeitschrift Für
620 Vergleichende Physiologie 7(1):56–143.
621 Brodrick, Emelie A., Nicholas W. Roberts, Lauren Sumner‐Rooney, Christian M. Schlepütz, and
622 Martin J. How. 2020. “Light Adaptation Mechanisms in the Eye of the Fiddler Crab Afruca
623 Tangeri.” Journal of Comparative Neurology cne.24973. doi: 10.1002/cne.24973.
624 Callier, Viviane, and H. Frederik Nijhout. 2013. “Body Size Determination in Insects: A Review and
625 Synthesis of Size- and Brain-Dependent and Independent Mechanisms.” Biological Reviews
626 88(4):944–54. doi: 10.1111/brv.12033.
627 Casares, Fernando, and Alistair P. McGregor. 2020. “The Evolution and Development of Eye Size in
628 Flies.” WIREs Developmental Biology n/a(n/a):e380. doi: 10.1002/wdev.380.
629 Cronin, Thomas W., Sönke Johnsen, N. Justin Marshall, and Eric J. Warrant. 2014. Visual Ecology.
630 Princeton University Press.
631 Currea, John P., Joshua L. Smith, and Jamie C. Theobald. 2018. “Small Fruit Flies Sacrifice Temporal
632 Acuity to Maintain Contrast Sensitivity.” Vision Research 149:1–8. doi:
633 10.1016/j.visres.2018.05.007.
634 Diez-Hermano, Sergio, Jorge Valero, Cristina Rueda, Maria D. Ganfornina, and Diego Sanchez. 2015.
635 “An Automated Image Analysis Method to Measure Regularity in Biological Patterns: A Case
636 Study in a Drosophila Neurodegenerative Model.” Molecular Neurodegeneration 10(1):9. doi:
637 10.1186/s13024-015-0005-z.
638 Franceschini, N. 1972. “Pupil and Pseudopupil in the Compound Eye of Drosophila.” Pp. 75–82 in
639 Information Processing in the Visual Systems of Anthropods, edited by R. Wehner. Berlin,
640 Heidelberg: Springer Berlin Heidelberg.
641 Friedrich, Markus. 2003. “Evolution of Insect Eye Development: First Insights from Fruit Fly,
642 Grasshopper and Flour Beetle.” Integrative and Comparative Biology 43(4):508–21. doi:
643 10.1093/icb/43.4.508.
644 Gaspar, Pedro, Saad Arif, Lauren Sumner-Rooney, Maike Kittelmann, David L. Stern, Maria D. S.
645 Nunes, and Alistair P. McGregor. 2019. “Characterisation of the Genetic Architecture
646 Underlying Eye Size Variation within Drosophila Melanogaster and Drosophila Simulans.”
647 BioRxiv 555698. doi: 10.1101/555698.
648 Gonzalez-Bellido, Paloma T., Trevor J. Wardill, and Mikko Juusola. 2011. “Compound Eyes and
649 Retinal Information Processing in Miniature Dipteran Species Match Their Specific Ecological
650 Demands.” Proceedings of the National Academy of Sciences 108(10):4224–29. doi:
651 10.1073/pnas.1014438108.
652 Harzsch, Steffen, and Gary Hafner. 2006. “Evolution of Eye Development in Arthropods: Phylogenetic
653 Aspects.” Arthropod Structure & Development 35(4):319–40. doi: 10.1016/j.asd.2006.08.009.
654 Holthuis, Lipke B. 1991. “Marine Species Identification Portal : Green Rock Lobster - Jasus
655 Verreauxi.” Food and Agriculture Organization of the United Nations 13(125).
656 Huber, John, and John Noyes. 2013. “A New Genus and Species of Fairyfly, Tinkerbella Nana
657 (Hymenoptera, Mymaridae), with Comments on Its Sister Genus Kikiki, and Discussion on
658 Small Size Limits in Arthropods.” Journal of Hymenoptera Research 32:17–44. doi:
659 10.3897/jhr.32.4663.
660 Huber, John T., and John W. Beardsley. 2000. “A New Genus of Fairyfly, Kikiki, from the Hawaiian
661 Islands (Hymenoptera: Mymaridae).” Hawaiian Entomological Society.
662 Iyer, Janani, Qingyu Wang, Thanh Le, Lucilla Pizzo, Sebastian Grönke, Surendra S. Ambegaokar,
663 Yuzuru Imai, Ashutosh Srivastava, Beatriz Llamusí Troisí, and Graeme Mardon. 2016.

20
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

664 “Quantitative Assessment of Eye Phenotypes for Functional Genetic Studies Using Drosophila
665 Melanogaster.” G3: Genes, Genomes, Genetics 6(5):1427–1437.
666 Johnsen, S. 2006. “Crepuscular and Nocturnal Illumination and Its Effects on Color Perception by the
667 Nocturnal Hawkmoth Deilephila Elpenor.” Journal of Experimental Biology 209(5):789–800.
668 doi: 10.1242/jeb.02053.
669 Juusola, Mikko, An Dau, Zhuoyi Song, Narendra Solanki, Diana Rien, David Jaciuch, Sidhartha Anil
670 Dongre, Florence Blanchard, Gonzalo G. de Polavieja, Roger C. Hardie, and Jouni Takalo.
671 2017. “Microsaccadic Sampling of Moving Image Information Provides Drosophila Hyperacute
672 Vision.” ELife 6. doi: 10.7554/eLife.26117.
673 Kensler, Craig B. 1967. “The Distribution of Spiny Lobsters in New Zealand Waters (Crustacea:
674 Decapoda: Palinuridae).” New Zealand Journal of Marine and Freshwater Research 1(4):412–
675 20. doi: 10.1080/00288330.1967.9515216.
676 Land, Michael F. 1997. “Visual Acuity in Insects.” Annual Review of Entomology 42(1):147–177.
677 Land, Michael F., and Dan-Eric Nilsson. 2012. Animal Eyes. OUP Oxford.
678 McDonald, Jeanne M. C., Shampa M. Ghosh, Samuel J. L. Gascoigne, and Alexander W. Shingleton.
679 2018. “Plasticity Through Canalization: The Contrasting Effect of Temperature on Trait Size
680 and Growth in Drosophila.” Frontiers in Cell and Developmental Biology 6:156. doi:
681 10.3389/fcell.2018.00156.
682 Mohrbeck, Inga, Pedro Martínez Arbizu, and Thomas Glatzel. 2010. “Tantulocarida (Crustacea) from
683 the Southern Ocean Deep Sea, and the Description of Three New Species of Tantulacus Huys,
684 Andersen & Kristensen, 1992.” Systematic Parasitology 77(2):131–51. doi: 10.1007/s11230-
685 010-9260-0.
686 Palavalli-Nettimi, Ravindra, and Ajay Narendra. 2018. “Miniaturisation Decreases Visual Navigational
687 Competence in Ants.” Journal of Experimental Biology 221(7):jeb177238.
688 Palavalli-Nettimi, Ravindra, Yuri Ogawa, Laura A. Ryan, Nathan S. Hart, and Ajay Narendra. 2019.
689 “Miniaturisation Reduces Contrast Sensitivity and Spatial Resolving Power in Ants.” Journal
690 of Experimental Biology 222(12). doi: 10.1242/jeb.203018.
691 Ready, Donald F., Thomas E. Hanson, and Seymour Benzer. 1976. “Development of the Drosophila
692 Retina, a Neurocrystalline Lattice.” Developmental Biology 53(2):217–40. doi: 10.1016/0012-
693 1606(76)90225-6.
694 Schneider, Caroline A., Wayne S. Rasband, and Kevin W. Eliceiri. 2012. “NIH Image to ImageJ: 25
695 Years of Image Analysis.” Nature Methods 9(7):671–75.
696 Schramm, Bartosz W., Agnieszka Gudowska, Filip Kapustka, Anna Maria Labecka, Marcin
697 Czarnoleski, and Jan Kozłowski. 2015. “Automated Measurement of Ommatidia in the
698 Compound Eyes of Beetles.” BioTechniques 59(2):99–101. doi: 10.2144/000114316.
699 Shingleton, Alexander W., W. Anthony Frankino, Thomas Flatt, H. Frederik Nijhout, and Douglas. J.
700 Emlen. 2007. “Size and Shape: The Developmental Regulation of Static Allometry in Insects.”
701 BioEssays 29(6):536–48. doi: 10.1002/bies.20584.
702 Shingleton, Alexander W., Christen K. Mirth, and Peter W. Bates. 2008. “Developmental Model of
703 Static Allometry in Holometabolous Insects.” Proceedings of the Royal Society of London B:
704 Biological Sciences 275(1645):1875–85. doi: 10.1098/rspb.2008.0227.
705 Smith, Dylan B., Galina Bernhardt, Nigel E. Raine, Richard L. Abel, Dan Sykes, Farah Ahmed, Inti
706 Pedroso, and Richard J. Gill. 2016. “Exploring Miniature Insect Brains Using Micro-CT
707 Scanning Techniques.” Scientific Reports 6(1). doi: 10.1038/srep21768.
708 Smith, J. L., N. A. Palermo, J. C. Theobald, and J. D. Wells. 2015. “Body Size, Rather Than Male Eye
709 Allometry, Explains Chrysomya Megacephala (Diptera: Calliphoridae) Activity in Low Light.”
710 Journal of Insect Science 15(1):133. doi: 10.1093/jisesa/iev114.
711 Snyder, Allan W. 1979. “Physics of Vision in Compound Eyes.” Pp. 225–313 in Comparative
712 Physiology and Evolution of Vision in Invertebrates, Handbook of Sensory Physiology, edited

21
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

713 by H. Autrum. Springer Berlin Heidelberg.

714 Snyder, Allan W., Simon B. Laughlin, and Doekele G. Stavenga. 1977. “Information Capacity of
715 Eyes.” Vision Research 17(10):1163–75. doi: 10.1016/0042-6989(77)90151-1.
716 Snyder, Allan W., Doekele G. Stavenga, and Simon B. Laughlin. 1977. “Spatial Information Capacity
717 of Compound Eyes.” Journal of Comparative Physiology 116(2):183–207.
718 Spaethe, J. 2003. “Interindividual Variation of Eye Optics and Single Object Resolution in
719 Bumblebees.” Journal of Experimental Biology 206(19):3447–53. doi: 10.1242/jeb.00570.
720 Stavenga, Doekele G. 1975. “The Neural Superposition Eye and Its Optical Demands.” Journal of
721 Comparative Physiology ? A 102(4):297–304. doi: 10.1007/BF01464342.
722 Stavenga, Doekele G. 1979. “Pseudopupils of Compound Eyes.” Pp. 357–439 in Comparative
723 Physiology and Evolution of Vision in Invertebrates. Vol. 7 / 6 / 6 A, edited by H. Autrum.
724 Berlin, Heidelberg: Springer Berlin Heidelberg.
725 Stöckl, A. L., D. O’Carroll, and E. J. Warrant. 2017. “Higher-Order Neural Processing Tunes Motion
726 Neurons to Visual Ecology in Three Species of Hawkmoths.” Proceedings of the Royal Society
727 B: Biological Sciences 284(1857):20170880. doi: 10.1098/rspb.2017.0880.
728 Streinzer, Martin, and Johannes Spaethe. 2014. “Functional Morphology of the Visual System and
729 Mating Strategies in Bumblebees (Hymenoptera, Apidae, Bombus ): Bombus Compound Eye
730 Morphology.” Zoological Journal of the Linnean Society 170(4):735–47. doi:
731 10.1111/zoj.12117.
732 Taylor, Gavin J., Stephen A. Hall, Johan A. Gren, and Emily Baird. 2020. “Exploring the Visual World
733 of Fossilized and Modern Fungus Gnat Eyes (Diptera: Keroplatidae) with X-Ray
734 Microtomography.” Journal of The Royal Society Interface 17(163):20190750. doi:
735 10.1098/rsif.2019.0750.
736 Taylor, Gavin J., Willi Ribi, Martin Bech, Andrew J. Bodey, Christoph Rau, Axel Steuwer, Eric J.
737 Warrant, and Emily Baird. 2016. “The Dual Function of Orchid Bee Ocelli as Revealed by X-
738 Ray Microtomography.” Current Biology 26(10):1319–24. doi: 10.1016/j.cub.2016.03.038.
739 Taylor, Gavin J., Pierre Tichit, Marie D. Schmidt, Andrew J. Bodey, Christoph Rau, and Emily Baird.
740 2018. “Bumblebee Visual Allometry Results in Locally Improved Resolution and Globally
741 Improved Sensitivity.” BioRxiv. doi: 10.1101/380527.
742 Theobald, Jamie. 2017. “Optic Flow Induces Spatial Filtering in Fruit Flies.” Current Biology
743 27(6):R212–R213.
744 Theobald, Jamie C., Eric J. Warrant, and David C. O’Carroll. 2010. “Wide-Field Motion Tuning in
745 Nocturnal Hawkmoths.” Proceedings of the Royal Society of London B: Biological Sciences
746 277(1683):853–60. doi: 10.1098/rspb.2009.1677.
747 Vanhoutte, Kurt J. A., Kristel F. L. Michielsen, and Doekele G. Stavenga. 2003. “Analyzing the
748 Reflections from Single Ommatidia in the Butterfly Compound Eye with Voronoi Diagrams.”
749 Journal of Neuroscience Methods 131(1):195–203. doi: 10.1016/j.jneumeth.2003.08.011.
750 Walker, Simon M., Daniel A. Schwyn, Rajmund Mokso, Martina Wicklein, Tonya Müller, Michael
751 Doube, Marco Stampanoni, Holger G. Krapp, and Graham K. Taylor. 2014. “In Vivo Time-
752 Resolved Microtomography Reveals the Mechanics of the Blowfly Flight Motor.” PLOS
753 Biology 12(3):e1001823. doi: 10.1371/journal.pbio.1001823.
754 Wardill, Trevor J., Samuel T. Fabian, Ann C. Pettigrew, Doekele G. Stavenga, Karin Nordström, and
755 Paloma T. Gonzalez-Bellido. 2017. “A Novel Interception Strategy in a Miniature Robber Fly
756 with Extreme Visual Acuity.” Current Biology 27(6):854–59. doi: 10.1016/j.cub.2017.01.050.
757 Warrant, Eric J. 1999. “Seeing Better at Night: Life Style, Eye Design and the Optimum Strategy of
758 Spatial and Temporal Summation.” Vision Research 39(9):1611–30. doi: 10.1016/S0042-
759 6989(98)00262-4.
760 Wehner, R., and I. Flatt. 1977. “Visual Fixation in Freely Flying Bees.” Zeitschrift Für Naturforschung
761 C 32(5–6):469–472.

22
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

762 White, R. H. 2003. “The Retina of Manduca Sexta: Rhodopsin Expression, the Mosaic of Green-, Blue-
763 and UV-Sensitive Photoreceptors, and Regional Specialization.” Journal of Experimental
764 Biology 206(19):3337–48. doi: 10.1242/jeb.00571.
765 Woodman, Peter N., Amy M. Todd, and Brian E. Staveley. 2011. “Eyer: Automated Counting of
766 Ommatidia Using Image Processing Techniques.” Drosophila Information Service 94:142.
767 Zhang, Zhi-Qiang. 2013. “Animal Biodiversity: An Update of Classification and Diversity in 2013. In :
768 Zhang, Z.-Q. (Ed.) Animal Biodiversity: An Outline of Higher-Level Classification and Survey
769 of Taxonomic Richness (Addenda 2013).” Zootaxa 3703(1):5–11. doi:
770 10.11646/zootaxa.3703.1.3.

23
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

771 Figures

772 Figure 1: Diagrams of apposition and superposition eyes demonstrating the geometric tradeoffs
773 between lens diameter (D), interommatidial angle (Δφ). The IO ), and field of view (FOV) for spherical (A. and
774 B.) and non-spherical eyes (C. and D.). A. In spherical apposition eyes, D directly determines
775 sensitivity while Δφ). The IO inversely determines acuity. B. In superposition eyes, migrating pigment
776 (indicated by the arrows) allows the ommatidia to share light, increasing the eye’s sensitivity. As a
777 result, these eyes must adhere to a spherical design. C-D. In non-spherical eyes, the intersection of
778 ommatidial axes differs from the center of curvature, with ommatidial axes askew from the surface of
779 the eye. Consequently, FOV and Δφ). The IO are not externally measurable and the effect of D on sensitivity is
780 reduced by greater angles of skewness. C. When the distance to the intersection is greater than the
781 radius of curvature, FOV and Δφ). The IO decrease, increasing average spatial acuity by directing more
782 ommatidia over a smaller total angle. D. Inversely, when the distance to the intersection is less than the
783 radius of curvature, FOV and Δφ). The IO increase, decreasing average spatial acuity by directing fewer
784 ommatidia over a smaller total angle. In both cases, optical sensitivity is lost because skewness reduces
785 the effective aperture of the ommatidia.

24
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

786
787 Figure 2: The ommatidia detecting algorithm (ODA) extracts periodic signals in a 2D image using the
788 FFT. A. In the frequency domain of a 2D FFT, called the reciprocal space, gratings are represented by
789 an x- and y-frequency. The polar coordinates represent visual properties of the corresponding grating.
790 The radial distance is a grating's spatial frequency, with high frequencies farther from the origin. The
791 polar angle is the grating's orientation , which is perpendicular to the grating’s wavefront. Notice that
792 the reciprocal space has local maxima (in red) approximately equal to the input grating parameters
793 (polar angle=45° and radial distance=.047±.005). B. In a hexagonal lattice, there are three major axes
794 (here in blue, green, and red). Each axis corresponds to a 2D spatial frequency (and it’s negative),
795 visible in the image’s reciprocal space. The periodic nature of the axes results in harmonic frequencies.
796 A low-pass filter returns a version of our original image primarily representing these three axes. The
797 center of each ommatidium is found at the local maxima of the filtered image.

25
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

798
799 Figure 3: The µCT pipeline starts with an image stack, which may be pre-filtered and cleaned of
800 unrelated structures. Then we filter the relevant density values and fit a surface to the coordinates,
801 allowing us to split the points into two sets or cross sections, inside and outside the fitted surface. For
802 each cross section, we generate spherically projected, rasterized images that are processed by the FFT
803 method for locating ommatidia, yielding approximate centers for the ommatidia. With these centers, we
804 can find the coordinate clusters corresponding to independent ommatidia and automatically measure
805 eye parameters.

26
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

806
807 Figure 4: Checkpoints along the µCT pipeline for the bumblebee scan. A. Import Stack: The
808 bumblebee µCT pipeline started with an image stack that we pre-filtered and manually cleaned of
809 unrelated structures. B. Get Cross-Sections: A surface was fitted to the coordinates identifying a
810 cross-section of points within 50% of the residuals. We generated a spherically projected, 2D
811 histogram of the crystalline cone coordinates (in grayscale). C. Find Ommatidial Clusters: The inset
812 zooms into stages of applying the ODA to label individual crystalline cones. The ODA locates the
813 cluster centers (red dots in top), allowing us to partition the coordinates based on their distance to those
814 centers (boundaries indicated by the red lines in middle). Then we apply the spectral clustering
815 algorithm to all of the 3D coordinates using the ommatidia count (bottom), finding the clusters
816 approximating the individual crystalline cones (each colored and labelled separately in bottom). D.
817 Measure Visual Parameters: Ommatidial diameter corresponds to the average distance between
818 clusters and their adjacent neighbors. Ommatidial count corresponds to the number of clusters. For
819
27
820
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

821 accounting for skewness in the ommatidial axes. The insets zoom into regions around the 25th, 50th,
822 and 75th percentile ommatidia along the y-axis (left) and x-axis (right). Note: in C and D the different
823 colors signify different crystalline cone clusters. In D, the black lines follow the ommatidial axes and
824 the black dots indicate the cluster centroids.

28
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

825
826 Figure 5: The ODA successfully approximated ommatidial counts and diameters when compared to
827 manual measurements. A. When applied to 6 ant eye molds of 5 species ranging in overall size and
828 lattice regularity, the automated counts were 97% and the diameters were 99% of those measured by
829 hand. For each image, the program missed relatively few ommatidia and the lens diameter
830 measurements were successful even when they varied substantially within an image (as in #5 from the
831 left). Species from left to right: Notoncus ectatommoides, Notoncus ectatommoides, Rhytidoponera
832 inornata, Camponotus aeneopilosus, Myrmecia nigrocincta, and Myrmecia tarsata. B. When applied to
833 39 microscope images of fruit fly eyes from the same species (D. melanogaster), the automated counts
834 were 104% and the diameters were 105% of those measured by hand, with correlations of .93 and .82.
835 C. Two example fruit fly eyes, one small and one large, with the automated ommatidia centers
836 superimposed as white dots. Again, there were relatively low rates of false positives and negatives.
837 Scale bars are 50µm.

29
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

838 Figure 6: A. The MicroCT program allows us to map the visual fields of our three specimens. Using
839 their spherical projection, we map the azimuth (along the lateral-medial axis) and elevation (along the
840 ventral-dorsal axis) angles of ommatidia and represent lens diameter with color. The colorbar to the
841 right indicates lens diameter, and shows diameter histograms for each species in white. Insets zoom in
842 on a 20°x20° region in the center of each eye, showing ommatidial lattices in more detail. Note that the
843 spherical projection of the bee eye underestimates its visual field and is missing a portion dorsally, as
844 explained in the text. B. Optical parameters for the three species are presented as a table. Values are
845 mean ± s.d except for angular measurements, indicated with an asterisk, which show median ± IQR. C.
846 Adapted from (Stavenga 1979). Bee eyes are approximately oval, and, unlike spherical eyes, have
847 different intersection points for horizontal (Ih) and vertical (Iv) ommatidial pairs corresponding to
848 different IO angle components horizontally (Δφ). The IO h) and vertically (Δφ). The IO v). For oval eyes, the horizontal
849 component of the IO angle of horizontal pairs (orientation=0°) is 2Δφ). The IO h while the vertical component is
850 ~0° and the horizontal component of diagonal pairs (orientation=±60°) is Δφ). The IO h while the vertical
851 component is Δφ). The IO v. Because the ommatidia form an approximately regular hexagonal lattice, Δφ). The IO v /Δφ). The IO h =
852 1/ √ 3 for spherical eyes, and √ 3 for oval eyes.

30
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

853

31
 bioRxiv preprint doi: https://doi.org/10.1101/2020.12.11.422154; this version posted December 12, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

854 Figure 7: A, B. Grayscale heatmaps indicate kernel density estimation of the entire dataset and red lines
855 indicate statistics about surrounding clusters: the longer, half opacity lines indicate IQR, tick marks
856 indicate the 99% CI of the median using bootstrapping, and points (often overlapping with the tick
857 marks) indicate the median. Solid blue lines, ticks, and points indicate the median and IQR of adjusted
858 lens diameters in the same way as red lines, and correspond to inserted axes on the right. Dashed blue
859 lines indicate the median absolute lens diameters also corresponding to the axes on the right. A. The
860 vertical and horizontal subtended angles of ommatidial pairs with respect to their orientation fit well
861 into the oval eye lattice from Figure 6C. The pair orientations form a trimodal distribution with means
862 close to ±60° (diagonal pairs) and 0° (horizontal pairs). Medians, IQRs, and 99% C.I.s are plotted for
863 each orientation ±15°. B. Using the lattice described in Figure 6C, we convert the horizontal and
864 vertical subtended angles into the horizontal and vertical IO angle components (Δφ). The IO h and Δφ). The IO v) and plot
865 them with respect to their azimuth and elevation positions on the projected visual field. C. Skewness
866 varies systematically across the eye influencing adjusted lens diameters and the horizontal and vertical
867 IO angle components (Δφ). The IO h and Δφ). The IO v). The position of each IO angle component is the average position
868 of the two ommatidia per IO pair. Dashed white lines in the skewness map indicate three regions of
869 generally different skewness: low skewness centrally and medially (top-right region; 19°) and moderate
870 to high skewness ventrally (bottom region; 35°) and dorsal-laterally (bottom region; 37°).

32