Automated monitoring of early life‐stage development in Atlantic cod (Gadus morhua) embryos exposed to a reference toxicant

Early life stages of fish are widely used for regulatory toxicity testing, and marine fish display high sensitivity to pollutant exposure. Exposure to pollutants during embryogenesis causes acute effects on embryonic development and survival, but also sub‐lethal impacts manifested as maldeveloped larvae. Acquiring time‐ and exposure‐dependent responses to pollutant exposure and other stressors in small organisms is labor intensive and often subjective. This leads to studies obtaining small sample sizes, with measurements often made infrequently during development. Automated monitoring methods can maintain consistency between measurements and allow many more measurements to be made, improving the quantity and quality of such data. We exposed Atlantic cod embryos to 3,4‐dichloroaniline, a reference chemical widely used as a positive control agent in regulatory fish embryo toxicity testing. We monitored their growth through daily imaging with an automated flow‐through imaging system. Biologically relevant sublethal endpoints were estimated from these images with a neural network and traditional machine vision methods. We demonstrate the automated capture and analysis of tens of thousands of images, producing detailed morphometric data from hundreds of fish over a 10‐d study period, and assess the effectiveness of the automated system. The automated method presented allows measurements to be made frequently without sacrificing the sampled organisms, making detailed time series of development obtainable. We show dose‐dependent effects of the toxicant on development and capture nonlinear responses that would not be attainable under a conventional manual sampling regime.

Fish embryos develop rapidly within the egg following fertilization and are highly sensitive to pollutant exposure, which can cause acute mortality.Exposure can also cause sublethal effects that can be observed during embryogenesis and after hatch.A good example is exposure of marine fish embryos to petroleum, where short-term exposure during embryogenesis causes effects on larvae development which can reduce their long-term survival (Laurel et al. 2019).Due to their sensitivity, and to reduce the use of juvenile and adult fish, fish embryo are extensively used for toxicity testing of pollutants.
Common toxicity tests include the Fish Embryo Acute Toxicity Test (FET) (OECD 2013b), a short-term test aimed at characterizing acute toxicity of chemicals by measuring mortality and deriving an LC 50 , and the Fish Early-Life Stage Toxicity Test (OECD 2013a).The latter test goes beyond the FET in also specifying the collection of sublethal effects in animals up to 60 d posthatch.These sublethal effects typically occur after hatching and may include developmental retardation (smaller larvae), reduced yolk consumption, and deformed larvae.Assessing these developmental effects on early life-stage fish typically requires an expert user to microscopically examine individual fish embryos or larvae and make visual judgments and manual measurements to assess their development.Photomicrography can aid in accurate measurement, and software tools have been developed for this task (Teixid o et al. 2019; Kvaestad et al. 2022;Rasmussen et al. 2022).However, the preparation and imaging of fish embryos one at a time is still a time-consuming task, leading to low sample sizes and reduced statistical power in such studies.
A number of instruments have been developed for automated imaging of marine plankton (Davis et al. 2005;Graham and Smith 2010;Davies et al. 2017;Picheral et al. 2021) and even ichthyoplankton (Iwamoto et al. 2001;Cowen and Guigand 2008;Colas et al. 2018), and some of these systems also include integrated segmentation or classification of the sample images.These tools mainly focus on in situ imaging rather than laboratory studies.Since the majority of these tools were developed there have also been major advances in machine learning (He et al. 2017), allowing for improved image processing and more accurate and detailed segmentation of images.
To take advantages of these developments we previously introduced a new, low-cost flow-through imaging system suitable for laboratory or ship use (Williamson et al. 2022).Built from off-the-shelf parts and including open-source software, it uses machine learning and classical computer vision techniques for segmentation and measurement of fish eggs and larvae.This system was designed with toxicological and developmental studies in mind, and aims to allow a nonexpert user to obtain useful biological data with much larger sample sizes than was previously practical, quickly and with minimal manual intervention.
In this article, we present a case study using the system: an acute exposure experiment of 3,4-dichloroaniline (3, in Atlantic cod (Gadus morhua) embryos.3,4-DCA is widely used as a positive control agent in regulatory fish embryo toxicity testing and is specified as such in the FET, as it has a well-documented and predictable toxicity (Busquet et al. 2014).The study demonstrates both the potential use and current limitations of the system and points a way forward for automation in the monitoring of early fish life stages.To illustrate the use of our imaging system, we present results from the exposure experiment, in which we obtain growth time series for several biologically relevant endpoints through daily measurements of hundreds of fish from the same population, showing dose-dependent responses to the toxicant.

Materials
Reference toxicant and fish embryos 3,4-DCA (Sigma-Aldrich, purity >98%) was dissolved in filtered (1 μm, Sterivex) natural sea water to a stock solution of 3000 μg.This stock solution was diluted in seawater to the six concentrations ranging 8-747 μg L À1 .The concentrations were verified using extraction and analyzed with gas chromatography-mass spectrometry (see Hansen et al. 2021 for details-that study and this one were performed simultaneously using fish and chemicals from the same source and stock, respectively).Newly fertilized Atlantic cod (Gadus morhua) eggs of good quality (> 75% fertilization success) were purchased from Havbruksstasjonen in Tromsø, Norway.Eggs were shipped to Trondheim the morning after fertilization, and acclimated in the laboratory for 2 d from their temperature at arrival (4 C) to a temperature of 8.5 C.They were then exposed to the 3,4-DCA exposure solutions for 4 d starting 3 d postfertilization (days postfertilization, dpf ).

Imaging equipment
A flow-through imaging system was used to collect images of the cod embryos, described in detail in Williamson et al. (2022).The system comprises a length of transparent tubing connected at one end to an elevated glass funnel and at the other to a collection flask (borosilicate, 250 mL).The tubing runs through an imaging chamber, illuminated by a lightemitting diode (LED) array (Advanced Illumination, SL246) and observed by a digital camera running at 8 Hz (2448 Â 2048 pixels resolution, Allied Vision, Prosilica GC2450C, 2/3 00 sensor) with a telecentric lens (8.8 Â 6.6 mm field of view, 1Â magnification, Edmund Optics #55-350).A peristaltic pump (Watson-Marlow, model 205S) is connected to the collection flask, in which it draws a light vacuum.Seawater containing fish embryos is introduced to the glass funnel and pulled through the system by gravity and the vacuum in the collection flask.Samples are imaged as they pass through the imaging chamber, and these images are saved in the camera-native Bayer8 image format to a computer for analysis.A diagram representing the system and a photograph of the complete imaging system are shown in Fig. 1.

Fish exposure
Approximately, 400-500 cod eggs were transferred to individual borosilicate beakers (500 mL) containing filtered sea water (controls) or exposure solutions at 8.5 C. The eggs were exposed for 96 h, and the exposure solution renewed every 48 h.At the end of exposure (7 dpf), they were transferred to new beakers (500 mL) containing filtered sea water.During recovery in clean sea water, the eggs developed at 9.5 C. Dead were removed on a daily basis and the seawater refreshed at least every 48 h.

Imaging
Images of the control and exposed groups were collected daily from 7 to 16 dpf using the flow-through imaging system.In addition, from 3 to 16 dpf approximately 50 eggs or larvae were retrieved daily from an unexposed population kept in a large holding tank held at 8.5 C and imaged alongside the experimental groups (sample images shown in Fig. 2).These images were used in training the neural network (NN), but were not included in the experimental results.

Morphometrics
Areas of interest in the collected images were identified using a NN.Two instances of Mask-RCNN (He et al. 2017) were trained on images of either cod eggs or cod larvae, with The NNs were trained to identify and segment in eggs: the entire egg, the yolk sac, the fish embryo, the eyes; and in larvae: the entire fish, the yolk sac, the eyes.
Following image segmentation by the NN, classical machine vision techniques were used to measure the segmented areas and biologically relevant endpoints were estimated (Kvaestad et al. 2022;Williamson et al. 2022).The results of the automated analysis were then manually checked and inaccurate measurements thrown out.
As indicators of embryonic and larvae development and growth, morphometric endpoints were measured.Estimated endpoints were the standard length of larvae, the minor and major axes and areas of several regions of eggs and larvae.The regions estimated included egg embryo, larva body, yolk sac, and the eyes of both eggs and larvae.In larvae, body area was estimated both as structural body area (area without the yolk sac) and total body area.Measurements were taken at approximately 24-h intervals.
Endpoints most prevalent in the literature were selected to be shown in Assessment section; these were mean egg diameter, eye area, yolk area and fraction, embryo body area, and in  larvae total and structural body area and standard length.These measurements provide indications on toxic effects of pollutants.
Changes in egg diameter could indicate osmoregulatory issues during embryonic development.As the embryo grows, the body area (and eye diameter) increases at the expense of the yolk sac (which is reduced), thus, potential differences for these variables between exposed and control eggs/larvae are good indicators of impacts on growth (Hansen et al. 2021).
Analysis steps included checking that detected animals did not cross the image boundaries; associating all measured endpoints with a specific egg or larva (to avoid erroneous measurements of bubbles or detritus in the water); estimating endpoints as described above; aggregating the results of automated and manual checks; and performing statistical tests and producing summary plots of the results.
For the observed results plots, the mean measured values for each endpoint for each day and 95% confidence intervals (CI) were calculated.Generalized additive models (GAMs) were fitted to the endpoint measurements to quantify rates of development and test for effects of treatments (Wood 2011).As all endpoints are strictly positive and bound by zero, GAMs were fitted using a Gamma distribution with a log link function.Smooth functions were applied to dpf and treatment concentration levels, allowing for the modeling of nonlinear endpoint progression and dose-response.To reduce overfitting, the basis dimension in both smooths (i.e., day and treatment) were set to three.GAMs were chosen due to their flexibility in modeling nonlinear effects without a predefined form, which allows the same general form to be used across the 11 endpoints analyzed in this study (Wood 2011).Following model fitting, we identified the lowest observed effect concentration (LOEC), defined as the minimum treatment concentration showing a statistically significant difference from the control, and above which all other treatments are also statistically significant (OECD 2013a).Specifically, we implemented pairwise comparisons between the marginal means of the control and every treatment concentrations between 8 and 747 μg L À1 for each day of measurement to identify the LOEC.We adjusted for multiple comparisons using Bonferroni correction by dividing the significance alpha level by the number of unique hypotheses in the developmental stage.Specifically, the adjustment for egg endpoints seven, corresponding effect of day and effect of treatment across five endpoints; and the adjustment for larvae endpoints was eight, corresponding to effect of day and treatment across six endpoints.Statistical analyses were conducted in R version 4.3.1 (R Core Team 2023), GAMs were fitted using the mgcv package (Wood 2011), and estimation and significance testing of marginal means were conducted using the emmeans package (Lenth 2023).

Measurements of the control group
We first look at the performance of the automated imaging and analysis system, considering the control group alone.The development of Atlantic cod is stereotypical and well documented (Hall et al. 2004), so images and measurements of this group, expected to develop normally, ensure that our system produces results in line with those reported elsewhere.We also use these measurements to understand the precision of the system-how small a change in size it can reliably measure.
Figures 6 and 7 show measured endpoints for eggs and larvae, respectively, in the control group.Summary statistics from the GAM models can be found in Supplementary Tables S1 and S2.
Qualitative image quality Figures 4 and 5 show sample images taken of the control group during the experiment, giving two examples from each day of the development of the embryos.Figure 2 shows sample images of earlier egg stages obtained with the system.The images show clearly the development of embryo, including increased length of the larvae and development of optic vesicles to more pigmented eyes.Distinct developmental stages as described in Hall et al. (2004) can be identified.Some organs, such as the heart, can be identified, but the image resolution does not permit detailed assessment of organ development.General morphology is apparent, and sufficient to identify gross abnormalities such as spinal or craniofacial deformations.

Development of the control group-Eggs
In the control group, eye area measurements showed a significant linear increase over time (Fig. 6).
The trend overall was for a significant increase in embryo area, which is in line with the expected growth of the embryos.However, a small decrease was observed at 10 and 11 dpf.As noted in NN segmentation accuracy section, embryo measurements in the last days of egg development become more difficult due to self-occlusion.Measurements of self-occluding embryos are thrown out.As larger embryos are naturally more likely to self-occlude, the resulting measurements for days 10-11 tend to be of the smaller (and potentially underdeveloped) embryos.Thus, we expect the current methodology to show some bias in this period.
Egg mean diameter showed small fluctuations over the period measured.While the GAM showed a significant deviation from constant size, the effect size was extremely small and a biological explanation is not apparent.Note the smaller scale and differing unit of this plot-the difference in daily means is within $ 0.2%.We attribute the majority of the variation here to measurement error, possibly due to small movements in the imaging system during cleaning.
There was a significant decrease in yolk area over the period measured, consistent with the embryos consuming yolk in order to grow.This was also reflected in the yolk fraction of each egg.Since the size of the eggs did not change significantly, the yolk fraction trend closely follows that of yolk area.

Development of the control group-Larvae
Total body area showed a significant decrease over time, while structural body area showed a significant increase over the same period (Fig. 7).Since total body area includes the area of the yolk sac, while the structural area is non-yolk portions of the body only, this reflects yolk being consumed as the fish grows.Standard length increased significantly over time.As in the eggs, eye area showed a significant increase over time.Some unreliability is expected in eye area measurements in larvae-see Larvae pose and orientation section for a discussion.Yolk area and fraction both significantly decreased.

Measurement precision
Two factors contribute to the final precision of our measurements: the precision attained in the images captured; and the precision and accuracy of the NN and image processing software.
The image precision depends on the physical attributes of the camera and lens such as sensor resolution, lens distortion, optical properties of the sampling chamber and and whether sensor plane and sample are exactly parallel.The theoretical imaging resolution of the camera-lens system is approximately 0.0035 mm per pixel.This can reasonably be considered a limit on the resolving power of the system in a single image, though the population level precision resulting from multiple measurements can be higher.
To assess the accuracy and precision of the imaging system, spherical particle standards (0.230 mm diameter) were imaged prior to the start of the experiment, at 6 dpf.One hundred thirty-six images containing 430 spheres in total were collected.The spheres were automatically detected and measured from these images, with manual verification of the detections performed to ensure no debris was measured in error.This resulted in a mean estimated diameter of 0.2323 mm (standard deviation 0.0108 mm), a difference from the nominal mean of approximately 0.0023 mm ($ 1%).
The precision of the final morphometrics depends on the accuracy of automated segmentation by the NN and the accuracy of the post-processing steps and is likely to vary between endpoints.The error here is unknown since we do not obtain ground truth data for these measurements; however, we can estimate a limit on the precision achieved.
Lacking a reasonable mechanism for the observed fluctuation in mean egg diameter of the control group, we can assume the majority of the changes in the mean to be due to measurement error.We assume the mean measured diameter across all observations (1.3465 mm) approximates the true population mean.Then the largest deviation from this value among the daily means gives us a possible lower bound on the system's measurement resolution.This occurs at 11 dpf, with an observed mean of 1.3427 mm, or a difference of 0.0038 mm.
We conclude that the measurement resolution of the system is close to the optical resolution of the camera-lens system.

Mortality
This study focuses on demonstrating the use of the imaging system to observe sublethal effects, and mortality was not recorded.However, for comparison, we reproduce in Fig. 8 mortality figures from the study Hansen et al. (2021), which was conducted alongside this one using eggs from the same source and the same 3,4-DCA concentrations prepared from the same stock solutions.Note that embryos in that study were kept at 8.5 C rather than 9.5 C, exposure took place in smaller containers and exposure solutions were renewed every 24 h rather than every 48 h, so some caution must be used in applying these figures to our study.

Egg measurements
A summary plot of the main end-points measured in eggs is shown in Fig. 9, and a table of pairwise comparisons with the control group in Supplementary Table S3.S10.
Eggs in the highest treatment groups, 343 and 747 μg L À1 , showed a strong response to the toxicant across all endpoints.While mortality was not specifically recorded, it was high in these groups leading to fewer measurements and larger error in the model estimates.
The size of eyes increased over time in all treatment groups.Significant effects were seen in all groups from 220 μg L À1 and higher compared with the control, with growth appearing to slow in later days.
Embryo area overall followed a similar trend to the control group, with the latter 2 d appearing to become smaller-this is likely due to measurement bias as discussed above.A dose-dependent effect was seen, with higher treatment groups generally having smaller embryos; this was statistically significant across all groups.In later days, the 747 μg L À1 group appeared larger than other groups in the observed data points, though with a small number of surviving eggs.It may be that the more affected eggs in this groups, with smaller embryos, suffered higher mortality and only those developing normally survived to the end of the measurement period.
Mean egg diameter exhibited the lowest variability relative to magnitude (a range of just 0.7% of the mean), with the most notable difference appearing in eggs in the 343 and 747 μg L À1 treatment groups, which were significantly smaller than the control.The 747 μg L À1 group was more strongly affected, and egg size in this group appeared to decline in later days.
Yolk area declined with time across all treatment groups.Yolk area in the 747 μg L À1 group was much smaller than other groups, and the rate of yolk consumption over time was relatively constant, a significant departure from the control and other groups where the rate of yolk consumption increased in later days.Though the GAM revealed significant differences in all treatment groups except 220 and 343 μg L À1 , this difference is not readily apparent in the plots.
Significant differences in yolk fraction were found, but they were inconsistent across treatment groups and the effect size was small.In the 747 μg L À1 group the yolk portion of the egg appeared generally smaller than in other groups, though yolk utilization rate was similar.LOEC and no observed effect concentration (NOEC) were calculated for each endpoint, shown in Supplementary Table S5.In addition, the minimum concentration required to see a significant effect (Bonferroni-adjusted critical value α ≤ 0.05/7 in egg endpoints and α ≤ 0.05/8 in larvae endpoints) from within the range of concentrations tested was estimated for each endpoint from the GAM predictions.This is shown in Supplementary Table S6.The results suggest the toxicant has a varying affect across different endpoints.
While the LOECs for eye area, and egg mean diameter fell within the upper concentration ranges tested (220 and 343 μg L À1 , embryo area, yolk area and yolk fraction both had a LOEC of 747 μg L À1 , the highest concentration tested. In eggs, embryo area, and yolk area and yolk fraction were predicted by the model to see an effect from the minimum concentration tested, 8 μg L À1 , while eye area and egg mean diameter appeared to require a much higher dose to see an effect.The prediction for yolk area and fraction is somewhat contrary to the LOEC, a discrepancy likely explained by significant differences being found in some (but not all) treatment groups below the LOEC.

Larvae measurements
A summary plot of the main endpoints measured in larvae is shown in Fig. 10, and a table of pairwise comparisons with the control group in Supplementary Table S4.
No eggs from the highest treatment group (747 μg L À1 ) survived to hatching, and mortality was also high in the 343 μg L À1 group, leading to fewer measurements.A few eggs had hatched into larvae by 11 dpf, but due to the very low number of measurements and correspondingly high relative variance they were not included in our analysis.
All treatment groups showed similar trends to the control, with yolk sacs size and fraction decreasing over time and other body parts growing.
Animals in the 343 μg L À1 group were overall significantly smaller than in the control and other groups, both by standard length and total and structural body area.Animals in the 220 μg L À1 group were also smaller in total body area and standard length, but no significant difference from the control group was found in structural body area.Treatment groups 108 μg L À1 and above had smaller eyes than the control group, with those exposed to higher concentrations more affected.All treatment groups showed smaller yolk sac size than the control in absolute terms, and the two highest treatment groups showed significant differences in yolk fraction.
As in eggs, LOEC, NOEC, and minimum effect concentrations were calculated (Supplementary Tables S5 and S6).LOEC for total body area and standard length fell in the second highest treatment groups, 220 μg L À1 , while that for structural body area was in the highest group, 343 μg L À1 .The LOEC for  S9.
eye area was 108 μg L À1 , somewhat lower than in eggs.Somewhat surprisingly, the yolk area LOEC fell in the lowest treatment group, though the yolk fraction LOEC was again in a higher group at 220 μg L À1 .
Minimum effect concentration predictions were broadly in line with these LOECs in most cases, predicting concentrations from 115 to 245 μg L À1 .A notable exception was yolk area, where the predicted minimum effect concentration was above 343 μg L À1 despite a LOEC of 8 μg L À1 .This may reflect the relatively lower significance (p between 0.05 and 0.01) of the difference from the control across all treatments in this endpoint.

Related work
3,4-DCA toxicity Hansen et al. (2021) present similar morphometrics in Atlantic cod embryos exposed to 7-747 μg L À1 3,4-DCA from 3 to 7 dpf, although only for larvae on a single day: 15 dpf.The study found concentration-dependent effects on eye size in all treatment groups except 8 μg L À1 , and significantly smaller fish in higher treatment groups.They also reported significantly higher mortality in the 747 μg L À1 group at the end of the exposure period, and significantly higher mortality than the control in all exposure groups at 14 dpf.The animals in that study were kept at a lower temperature than those described here (8.5 C vs. 9.5 C), and so can be expected to have developed more slowly-at 15 dpf our fish have an additional 8 degree-days of growth over those in the Hansen et al. study.
No other 3,4-DCA exposure experiments are reported in Atlantic cod embryos, but acute exposure in early life stages of other fish species has been examined.
Early life stages of fathead minnows (Pimephales promelas) exposed to 5.1-157 μg L À1 3,4-DCA from hatching (at 4-5 dpf) (Call et al. 1987) were found to have slightly increased mortality rates at 5 d postexposure (days postexposure, dpe) and lower survival at 28 dpe than controls, and significantly smaller larvae (by standard length and weight) in one of two tests performed.
Sublethal effects comparable to the endpoints in our study were not measured, but oedema and anatomical abnormalities were reported in embryos exposed to concentrations from 500 to 5000 μg L À1 .
Several studies have looked at the effects of 3,4-DCA in zebrafish (Danio rerio) embryos.This species is the model used in the FET (OECD 2013b), which specifies 3,4-DCA as a positive control in this test with the expectation that a concentration of 4000 μg L À1 should result in at least 30% mortality after 96 h.Nagel et al. (1991) exposed several groups of newly spawned zebrafish ova to 3,4-DCA at concentrations from 2 to 200 μg L À1 .They found significant mortality at 2 weeks and beyond at 200 μg L À1 and at 4 weeks and beyond at 100 μg L À1 .Larva body length appears to be reduced at 100 and 200 μg L À1 , but there is insufficient data to draw a strong conclusion.Schiwy et al. (2020) exposed zebrafish for up to 5 dpf to 3,4-DCA concentrations of 500, 2000, or 4000 μg L À1 .They found no significant mortality at concentrations of 500 or 2000 μg L À1 , but significantly decreased survival at 48 h at 4000 μg L À1 .Exposure solutions were renewed every 24 h, and they found that the exposure concentration decreased markedly during these intervals compared to a control solution with no fish, which they attribute to uptake and biotransformation of the toxicant by the fish.
This toxicant uptake is also supported by Hertl and Nagel (1993), who in a study with nominal exposure concentrations from 0.1 to 0.5 μmol L À1 (16.2-81.0μg L À1 ) 3,4-DCA with a 48-h exposure time reported high accumulation of the chemical in 4-d-old zebrafish larvae, but much reduced uptake in eggs and older fish (17 d and 6 months).This suggests that young larvae could be particularly affected by the toxicant, while exposure in eggs may have reduced effects.Scheil et al. (2009) looked at zebrafish larvae exposed to nominal 5-500 μg L À1 concentrations of 3,4-DCA from fertilization onwards.They found significantly increased mortality from 6 dpf at 1000 μg L À1 , 7 dpf at 500 μg L À1 , and a lower (but significant) increase in mortality from 5 dpf at 5 μg L À1 .No significant differences in mortality were seen during the 11 d experiment at 10, 100, or 250 μg L À1 concentrations.Significant and dose-dependent increases in the occurrence of edema in surviving fish were seen at 1000, 1500, and 2000 μg L À1 compared to controls, and reduced locomotor activity at 500 and 1000 μg L À1 .
Our study showed decreased size in most endpoints in eggs at 747 μg L À1 and some effects at 343 μg L À1 .While not specifically recorded, mortality was high in both these groups and no eggs in the 747 μg L treatment group survived to the end of hatching at 12 dpf (8 dpe).In larvae we saw smaller fish in the 343 μg L À1 and reduced eye size in all treatment groups except 8 μg L À1 .The larvae results are consistent with those of Hansen et al. (2021), and increased mortality at higher exposures was seen in all studies.The decreased growth rate but normal yolk sac utilization seen in our 343 μg L À1 group might be explained by the findings of Hertl and Nagel (1993) and Schiwy et al. (2020), with energy from the yolk being used to process the toxicant.
The concentrations at which increased mortality was reported varied widely between studies.It is likely that sensitivity to 3,4-DCA varies between species: Call et al. (1987) reported effects in fathead minnow at lower concentrations than seen in our study, while Schiwy et al. (2020) saw no increase in mortality in zebrafish at more than twice our highest concentration.Ibrahim et al. (2020), who used Javanese medaka in their study, comment that their reported 96 h LC 50 is up to 15 times that found in zebrafish.Hansen et al. (2021) provide a summary of LC 50 values for different species from the literature and similarly show a large variation between the lethal concentration in different fishes.
Differences between studies in onset of exposure, exposure duration and exposure solution renewal regime make direct comparisons difficult.Call et al. (1987), Nagel et al. (1991), Schiwy et al. (2020), Hansen et al. (2021), and this study verified 3,4-DCA concentrations through analytical chemistry, but in other experiments the nominal concentrations reported may vary from the actual concentration, also affecting comparison.Furthermore, the rates at which fish embryos develop and even the developmental stage at which morphological changes occur varies widely between species, as do the temperatures at which they are typically held in the laboratory.Zebrafish incubated at 28.5 C are reported to hatch 2-3 dpf (Kimmel et al. 1995) and even in the studies presented here temperatures varied, while the Atlantic cod in our study were kept at 9.5 C and began hatching far later, at 10-11 dpf.An exposure period of 96 h, which in our study accounts for only part of the embryonic stage of cod, would clearly not be equivalent to 96 h in zebrafish where it might encompass all egg stages and the early larva.

Other automated systems
Other systems for automated imaging and analysis of small marine organisms have been developed.Numerous instruments exist for in situ marine particle imaging, such as the Video Plankton Recorder (Davis et al. 2005), In Situ Ichthyoplankton Imaging System (Cowen and Guigand 2008), LISST-Holo (Graham and Smith 2010), SilCam (Davies et al. 2017), and Underwater Vision Profiler (Picheral et al. 2021).While the target particles are often of a similar size to those in this study, our system is aimed primarily at laboratory studies rather than use in the field.In situ instruments cannot reliably image the same population of animals repeatedly, and typically measure only the overall size of particles.Much effort has also gone into machine learning systems for plankton classification (see Irisson et al. 2022 for a recent review).Fish eggs and larvae are sometimes considered a class of plankton; however, in this paper, we focus not on classification but on detailed measurements of body parts over time.We therefore limit ourselves in this section to other laboratory or shipboard systems, aimed at either collecting images of millimeter-scale organisms or analyzing such images.
The system described here builds on the AutoMOMI method of Kvaestad et al. (2022), which used a similar NN to measure endpoints in microscopy images of cod larvae.Significant developments over that system are the ability to analyze eggs as well as larvae, and the use of automated imaging as well as analysis.
FishInspector (Teixid o et al. 2019) and FishSizer (Rasmussen et al. 2022) are software tools developed to assist in the measurement of fish from existing images.FishSizer can be used with several different fish species, but is limited to segmenting fish from the background and measuring myotome height and standard length.While FishInspector can measure many more endpoints, it is designed for use with zebrafish only.Both systems use hand-tuned image processing algorithms rather than a machine learning approach and require additional manual intervention in extracting measurements.Neither system can obtain measurements from eggs.
ZooScan is an integrated imaging and analysis system aimed at identifying and counting zooplankton (Grosjean et al. 2004).Water samples containing plankton are imaged in a scanning cell, and the plankton are then detected and classified by taxa through machine learning.Its aim of particle classification is distinct from our goal of feature measurement; however, a similar imaging method could feasibly be used with our analysis system rather than the flow-through method we describe.
REFLICS (Iwamoto et al. 2001) and ZooCAM (Colas et al. 2018) are imaging systems conceptually similar to our own, also using flow-through imaging and transmitted illumination.Both are intended for shipboard use, and target particles in the millimeter scale such as fish eggs.REFLICS has limited image processing, performing image segmentation and feature extraction to isolate images of fish eggs.ZooCAM additionally uses a Random Forest algorithm to identify zooplankton and classify fish eggs by species and developmental stage.
Our system presents a novel combination of features: integrated image collection and analysis, detailed measurements of multiple endpoints, and targeting translucent millimeterscale particles in a laboratory setting.Where in situ instruments capture only a snapshot of an organism's life history, we aim to monitor development over a longer period.The systems described above tend to perform either imaging or image processing, or where they perform both they lack the breadth and flexibility of measurements we can obtain.We believe our system uniquely fulfills the requirements of carrying out the type of developmental studies described in this article and offers a valuable new tool to researchers performing such work.

Subjectivity of the method
The method currently requires decisions to be made by the operator at two steps.
First, and most critically, the annotation of the training data requires an operator to accurately identify and mark the body parts under consideration.In some cases, it is not possible to unambiguously identify a body part-this is most common in eggs where part of the yolk sac is obscured, and in larvae where the eyes overlap or the yolk sac is very small.In such instances, it is better not to make an annotation at all, which does not affect NN training, than to risk an incorrect annotation that might mislead the training process.
Second, during manual validation of the NN segmentation some judgment must be used as to whether a segmentation is "good enough."Obviously incorrect segmentations might include misidentified body parts or gross departures from the body part, but in cases where a segmentation is mostly correct with only small discrepancies it is necessary to decide whether the measurement falls within an acceptable error range.
The first of these decision points is analogous to the traditional method of manual measurements of microphotographs, where an expert operator is required to make similar judgments.Documented training plans, not working when fatigued, having the same operator make all measurements or repeating measurements between different operators can all help mitigate sampling error, or at least maintain a consistent and systematic error that can be accounted for during analysis.Since a NN once trained applies its training consistently, repeatably and without fatigue to any number of samples, we consider that this method is less subjective than the traditional method, assuming good training data.
The second case requires less operator expertise than the first, since most errors are immediately obvious at a glance.As discussed in NN segmentation accuracy section, a majority of bad measurements were due to dead animals, or a small number of stereotypical limitations of the system.Ideally these problems should be identified automatically, and possible approaches to this are discussed in Further automation section.Measurements that are only slightly wrong might require more operator judgment, but will also influence the overall results of a study less.While the system is not quite at this stage, the goal is that manual checking should not be required, since reliable, robust measurements can be expected without needing human verification.

Ease of setup and use
Setting up the imaging system takes around half an hour, but once setup requires little maintenance-flushing with clean water before and after use is usually sufficient.The imaging chamber requires topping up with water every few days as air creeps into its outer layer.If left in place for an extended period the imaging chamber, camera lens and LED array should be checked for dust and cleaned if necessary.
The imaging system is straightforward to use, and a nonexpert operator should be able to carry out experiments with less than an hour of training.Some care must be taken when introducing samples to the glass funnel, especially eggs, as there is a possibility for the funnel exit to become blocked by the samples.This can be mitigated by introducing samples slowly and with sufficient seawater to avoid clumping.Image capture requires an operator to input a single, short command on the computer to begin imaging, and another to end.
Image annotation for NN training requires judgment as described above.Software has been developed to assist in this task, the use of which can be quickly learned and does not require any special knowledge.Training the NN is not entirely straightforward.We consider that most users should not need to carry out these steps, since once a network is trained it can be used on all similar data sets.
The analysis software used is still under active development, and this is where a nonexpert user is likely to require most guidance.Efforts have been made to make it simple to run standard analyses, and output is to the portable and widely-used Comma-Separated Values format, but any runtime errors that occur may be time-consuming or beyond the capabilities of users unfamiliar with programming in Python to solve.

Imaging issues
In general, the flow-through imaging system provides highquality images of sufficient detail for accurate segmentation and measurement.However, some issues were noted during collection of this dataset, and have subsequently been addressed.For a summary of images captured by the system, and the portions of these images that were rejected in either automated or manual checks, see Supplementary Tables S7 and S8.
One of the largest sources of unusable measurements was images where samples crossed the image boundary, leading to a partial image of an egg or larva (see also NN segmentation accuracy section).Images were initially captured at a constant 8 Hz without regard to the image contents.Since collecting this dataset, image capture has been improved such that the camera monitors the sample flow at 15 Hz, but writes an image to disk only when a sample is entirely inside its field of view.This not only eliminates badly framed images, but also allows a higher effective framerate (because slow disk-write operations are reduced).Since this issue is caught by automated checks, it is not expected to have had an effect on our results for this dataset.
A less easily solved problem is that of the same sample being imaged more than once.The flow rate of the system is such that particles moving with the flow of water pass through the imaging chamber quickly and are generally imaged only once.However, fish larvae are able to freely swim within the tubing of the imaging system, sometimes swimming against the direction of flow.In some cases, they are briefly able to maintain their position or even swim upstream, and when this occurs inside the imaging chamber they can appear in multiple images.This may bias the population statistics of the study: well developed larvae are better able to resist the current than poorly developed or malformed ones.The incidence with which this occurs is not accurately known as it is difficult to reidentify a cod larva with certainty.However, manual inspection of images suggests that it occurs primarily towards the end of a sampling run where water pressure and thus flow rate is lower, and occurs infrequently (two or three fish per sampling run), typically leading to four or five images of the same animal.Where possible these duplicate images were thrown out in manual checks, and the problem can be mitigated by ensuring that water in the funnel is kept topped up even at the end of a sampling run, maintaining more consistent water pressure and flow rate.
A final imaging issue is that some measurements were disrupted by scratches on the surfaces of the imaging chamber, which can obscure parts of the samples.The plastic cuvette used has since been replaced with a glass imaging chamber that is more resistant to scratches and has eliminated this problem.Where the issue occurred, it should have been caught in manual checks.

Reliability of volume estimates
A limitation of the current analysis is that while it is able to identify fish embryos within an egg and successfully segment them from their surroundings, the accuracy of volume estimation as described in Williamson et al. (2022) is unreliable.The primary reason for this is that the eggs are imaged at a single angle at random orientation.In very early stages, when the cell mass of the fish still forms a disc or ellipsoid, the embryo has a regular shape where volume estimation may be possible.Later, however, as the fish elongates and wraps around the inside of the egg, this becomes more difficult as the fish can self-occlude or simply be at a viewing angle that does not give enough information to estimate one or more of its dimensions.
Volume estimation in larvae is typically more straightforward, but larvae in a curved pose or with irregular or occluded yolk sacs still present difficulties.This is not a problem that can be readily solved with more training data, but requires a new approach.One possibility is to image samples from multiple angles, and reconstruct a three-dimensional model that can be used to estimate morphometrics.Others possible methods might take inspiration from similar work in plankton, such as distance maps for volume estimation (Moberg and Sosik 2012), or the use of statistical modeling to estimate reconstruction from a single angle as in Levis et al. (2018) and Ronen et al. (2021).

NN segmentation accuracy
Supplemental Tables S7 and S8 show how many measurements were made by the NN for each date and treatment group, how many were kept after automated checks, and how many after manual checks.
Automated checks typically throw out measurements either because they are identified as crossing the edge of the image or because endpoints are not associated with an egg or larva (usually indicating an erroneous identification).Manual checks by a human are able to identify many more bad measurements: air bubbles misidentified as eggs, overlapping eggs or larvae wrongly segmented as a single animal, mis-segmentations of body parts, or simply inaccurate measurements.
Overall accuracy in the eggs dataset was good, with over 70% of NN measurements kept.About half of the measurements excluded were detected by automated checks, and the rest caught in manual checks.By far the most common measurements thrown out during manual checking were of the embryo body, particularly later in development.This occurred for two main reasons.First, large embryos wrap around the inside of the egg and self-occlude, making measurements unreliable.Second, the NN segmentation method used does not allow for "holes" in the area segmented.This means that when an embryo wraps around the inside of the egg with clear space in the center, this space is included in the embryo's body area.This limitation of the method is responsible for the majority of the egg measurements thrown out, and will be addressed in a future version of the software.
The other group with many bad measurements was the 747 μg L À1 treatment group, especially later in development.This is because many eggs imaged were dead, and dead animals were not considered in the analysis.
Accuracy in the larvae dataset was considerably lower than in eggs.Overall, around 45% of measurements were kept, and somewhat over half of the exclusions occurred during automated checking.Body and eye measurements had an especially high failure rate-as discussed in Larvae pose and orientation section, measuring these two endpoints from a single camera angle in freely swimming animals poses a particular difficulty.Similar to the egg dataset, the highest exposure group in larvae (343 μg L À1 ) included many dead animals, and these measurements were thrown out in manual checks.

Larvae pose and orientation
In manual measurement of fish larvae, the fish are immobilized and manually posed such that they lie straight and present a side-on view to the camera (see example microscopy images in Supplementary Fig. S1).This allows for the most accurate measurement, and ensures consistency between samples.Larvae flowing through the imaging chamber in our automated system are able to swim freely, resulting in a range of poses and orientations.The narrow tubing encourages fish to swim either with or against the flow, putting their long axis parallel to the image plane, but there is enough room for them to bend or even turn around within the tubing.As larvae swim within the tubing they flex their tail, appearing foreshortened from the camera's viewpoint and resulting in underestimation of the length of the fish.The system relies upon finding the center line of the animals to estimate standard length, and in highly curved larvae this line is also shortened, again leading to an underestimate.See Fig. 11 for images illustrating the problem.
Fish can also be oriented in any direction around the long axis, presenting their side, belly or top to the camera.We can broadly consider imaged larvae to fall into one of two orientations-lateral (side-on) and dorsoventral (top-down) (see Fig. 12).This variation in orientation presents a difficulty in recovering consistent measurements of the larvae.Standard body length is not affected, as animals appear the same length in any orientation of the long axis.However, apparent body area is less in the lateral view, as the fish are narrower observed from above than from the side.The yolk sac is generally obscured or only partially visible in dorsoventral images, leading to missing or erroneous measurements.
Eyes are also particularly affected, presenting a smaller area from above than from the side.In the lateral view, there is another problem-since the bodies of the fish are at this stage largely transparent, but the eyes opaque, the eye facing away from the camera is sometimes visible through the body of the animal, partially overlapping with the closer eye.There is low contrast between the two eyes and it is generally not possible to separate them in segmentation, leading to a single, "figure 8"-shaped eye being measured.Such false eyes result in overestimation of diameter and area.
These problems of pose and orientation are difficult to overcome in software, and the measurement errors they cause are currently caught in manual checks.However, they might also be addressed by improvements to the hardware of the system.Imaging from multiple angles (see Further automation section) could eliminate problems arising from orientation, and improved imaging contrast through more even lighting and an optically clearer imaging chamber could allow separation of the eyes in cases where they overlap.

Advantages of the method
The clearest advantage of this system is its ability to image and measure more endpoints in many more samples, much more quickly, than is practically feasible with manual methods.Other toxicity studies reviewed in this article used as few as 2 replicates of 10 animals (Schiwy et al. 2020), and 800 embryos were measured in the largest study (Nagel et al. 1991), with the work divided over 8 laboratories.In most cases, few endpoints were measured, and typically only mortality was reported at multiple timepoints.
Our method allows for thousands of animals to be repeatedly imaged and analyzed.To image all eight treatment groups took one person roughly 2 h of laboratory time each day; around 15 min per imaging run.This permits far higher sample sizes than are possible with manual methods, making the system more sensitive to small effects and leading to smaller statistical uncertainty.Increasing the number of animals measured per imaging run necessarily also increases the time taken for imaging since a larger sample volume is required, but does not affect the amount of manual intervention required and no additional preparation is needed, so our system also scales well to larger experiments.
Manual measurements are usually lethal for the animals being measured, as they must be removed from the water and immobilized.Our method does not appear to damage the test subjects and they remain in water at all times.This means it is possible to follow the development of a population over time without diminishing the size of that population.In contrast with other studies, we were able to measure the same population daily throughout development, obtaining a time-series of their growth rather than a single set of measurements at the end of the study period.This is especially valuable given the nonlinear response seen in several endpoints, since these would not be possible to detect with a single measurement.While we imaged consistently once per day, this could naturally be extended to more frequent measurements, for instance sampling more often during the critical and rapid stages of early development.Apart from frequency and quantity of images obtained, we also show the advantages of automated analysis.Measuring multiple endpoints does not take additional human labor since all that is required is to run a script.This may take several hours to produce results, but does not need intervention while running.Time-consuming manual checks of the measurements are currently still required since there is high error rate in this first iteration of the system.This can be improved as discussed in Limitations of the current method and Further automation sections, and even now checking using our software tools is significantly faster than making the measurements manually-typically a few seconds per image.
Analyzing additional endpoints at a later time is also potentially faster than with manual measurements.All images are saved for future reanalysis.Some measurements will require retraining the NN with additional annotations, for instance identifying other body parts such as the heart.Once trained though, the entire dataset, and future datasets, can be analyzed without further significant effort.We consider this flexibility an advantage over other automated systems using hand-written algorithms such as FishInspector Teixid o et al. (2019).Additional measurements derived from the NN segmentations can be obtained through traditional computer vision methods, which necessitates writing code but no retraining.Examples might include obtaining the eye to forehead distance or myotome height in larvae.

Measurement accuracy
We have explored the precision of the system's measurements in Measurements of the control group section, and show that we obtain results broadly consistent with those from the literature in Related work section.However, a quantitative assessment of the system's accuracy, as compared to manual measurements, has not been performed.Kvaestad et al. (2022) looked at the performance of a similar automated analysis system with an expert human using microscopy images, and found they achieved comparable results.We anticipate that our system will have a lower accuracy in individual measurements when compared with microscopy images due to the difficulties presented by randomly posed and oriented animals, but will fall within an acceptable error range and provide many more measurements, possibly improving accuracy at a population level.While we feel that the data the system provides are already useful, demonstrating the system's performance empirically is an important next step if it is to be used in toxicological or developmental studies.

Software
A large source of error in measurements was measuring the wrong thing: air bubbles erroneously measured as eggs, dead animals being included in measurements, eggs being measured as larvae or larvae as eggs.One solution to this problem is an additional classification step before the NN segmentation.This could be approached through classical machine vision, but since we already have an annotated training set available it seems natural to train a NN to classify objects in the images.Classes might include live egg, live larva, dead egg, dead larva, air bubble and dust particle.The segmentation step would then only run on areas of an image identified as being live animals, reducing processing time and improving measurement accuracy.
Another shortcoming to the method is that the segmentation used does not allow for holes within a segmented area.This was a problem in measuring embryos inside eggs, where the embryo is often wrapped around the edge of the egg (from the camera's point of view) with a gap in the center.In such cases, the gap is erroneously considered to be a part of the embryo.Allowing holes within segmentations would improve the accuracy of these measurements and lead to fewer being thrown out.This is a limitation of Mask-RCNN, which allows only convex polygons to be used as segmentation masks.It is possible to represent "donut" shapes such as the embryos as two polygons, but this seems inconvenient and unintuitive for the user performing annotation.To overcome this, a preprocessing step is needed to convert an annotation with a hole into two masks, one representing the hole, and to use an additional class when training the NN to identify such holes.

Hardware
A useful extension to the current imaging system would be to simultaneously image samples from multiple angles.This would both provide a better chance of producing an image from which reliable measurements can be made, and open the possibility of combining the images for more accurate measurement, especially volumetric measurements.With sufficient coverage, a 3D reconstruction of the samples would be possible.This is achievable with multiple cameras, but may come with its own problems in providing lighting to samples and synchronizing image capture.A simpler approach is to use angled mirrors to allow a single camera multiple views of the sample.
Another interesting possible extension is a system where samples flow constantly between two holding tanks, passing through an imaging chamber between the tanks.Rather than performing discrete imaging runs, the population would be continuously imaged, with a new image being captured whenever a sample passes through the imaging chamber.This would reduce the human intervention required in imaging, and provide a finer-grained view of development with a resolution in minutes rather than days.Providing a continuous flow between tanks without harming the test animals and without the possibility for clogging is, however, a significant technical challenge.

Other organisms
While this study has considered Atlantic cod, the system described can be used for any translucent millimeter-scale organisms, and with different lenses and imaging chambers the same concept can be extended to other particle sizes.Some examples of organisms other than cod imaged with the system are shown in Fig. 13.Tests of a similar automated analysis system have been performed with several other fish species (Kvaestad et al. 2022), and such animals can even use the same image processing and the same NN classes, with some additional training.Organisms other than fish require new classes to be defined and trained, but are also possible.

In situ use
A version of the system, modified with a glass imaging chamber, has also been used on-board a ship using samples from a rosette water sampler deployed in the Norwegian Sea.In this case, the system was used to capture images, but not for analysis.Sample images are shown in Fig. 13.Shipboard operation was generally similar to in a land-based laboratory, with two exceptions.
Wave motion affected the flow of samples through the imaging system, leading to inconsistent flow rate and in rough conditions even a reversal of the flow.This was partly mitigated by increasing the rate of the peristaltic pump, but could not be entirely overcome.This effect lead to some samples being imaged more than once as they remained or reappeared in the imaging chamber, while others were missed entirely as they moved through too quickly.
Samples taken from the ocean also contained a larger variety of particles than those from the laboratory, including those too large to pass through the imaging system.To avoid clogging, water samples were passed through a 2000-μm sieve before introduction to the imaging system to remove larger particles.In some samples even this was insufficient, and particles of interest were picked out of the sample under a microscope for imaging.Clearly, this removes one of the primary advantages of the imaging system-that minimal manual sample preparation is required.Higher dilution of the samples in seawater can also mitigate clogging, but at the expense of longer sampling runs.
Combining the ability to measure a variety of organisms as outlined above, and object classification such as used in the ZOOSCAN system (Grosjean et al. 2004), samples of natural seawater containing a mixture of target organisms could in future feasibly be analyzed in detail.
the training data comprising a subset of images selected at random from those recorded during this experiment and a similar toxicity study (B.H. Hansen unpubl.),divided into a training set (approximately 90% of images) and a validation set (approximately 10% of images).Networks were monitored with TensorBoard during training and training was ended when overfitting began to occur.The larvae network was trained for 2115 epochs of 900 iterations on a dataset of 345 training images and 34 validation images.The egg network was trained for 1151 epochs of 900 iterations on a dataset of 464 training images and 52 validation images.The network was provided with hand-drawn annotations for each image identifying the outlines of regions of interest.Examples of training images with annotations are shown in Fig. 3.

Fig. 2 .
Fig. 2. Sample images of eggs between 3 and 6 dpf (left to right), showing development during this time.The eggs in these images come from the population not used in the exposure experiment.Scale bar 0.5 mm.

Fig. 1 .
Fig. 1.(a) Diagrammatic representation of the imaging and analysis pipeline (reproduced from Williamson et al. 2022).(b) Photograph of the imaging equipment setup for use.

Fig. 3 .
Fig. 3. Example training images with annotations for (a) eggs (8 dpf) and (b) larvae (14 dpf).Overlays show the areas presented to the NN as belonging to a given body part: cyan the egg, blue the eyes, purple the embryo inside an egg, green the yolk sac, red the body of a larva.

Fig. 4 .
Fig. 4. Cropped sample images showing eggs from the control group on each day of the experiment, from 7 to 11 dpf.Scale bar 0.5 mm.

Fig. 5 .
Fig. 5. Cropped sample images showing larvae from the control group on each day of the experiment, from 12 to 16 dpf.Scale bar 1 mm.

Fig. 6 .
Fig. 6.Measured endpoints of eggs in the control group.Black error bars show 95% CI of the measurements, red lines show GAM predictions and red bands 95% CI of the model.Number of observations at each day for each endpoint are given here, and in Supplementary TableS9. S9 .

Fig. 7 .
Fig. 7. Measured endpoints of larvae in the control group.Black error bars show 95% CI of the measurements, red lines show GAM predictions and red bands 95% CI of the model.Number of observations at each day for each endpoint are given here and in Supplementary TableS10.

Fig. 8 .
Fig. 8. (a) Cumulative dead fraction as a function of time (dpf) for cod embryos exposed to six concentrations of 3,4-DCA between 3 and 7 dpf (indicated in gray area).Light blue area indicates hatching days.(b) Estimated LC 50 values (and confidence intervals) plotted as a function of time.Gray area indicates the exposure period.(c) Cumulative dead fraction at the end of the exposure (7 dpf).(d) Cumulative dead fraction at the end of the experiment (14 dpf).Data are given as average AE standard deviation (N = 6 for control, N = 3 for all 3,4-DCA exposures).Significantly higher dead fraction in 3,4-DCA treatments than controls are given as **p < 0.01, ***p < 0.001 and ****p < 0.0001.Figure and caption reproduced with permission from Hansen et al. (2021)-see the text for discussion of the differences between that study and ours.

Fig. 9 .
Fig. 9. Evolution of five different endpoints over time in eggs.Plots in (a) show mean measured values, (b) GAM predictions.Translucent bands show the 95% CI.Numbers of observations for each data point are given in Supplementary TableS9.

Fig. 10 .
Fig. 10.Evolution of five different endpoints over time in larvae.Plots in (a) show mean measured values, (b) GAM predictions.Translucent bands show the 95% confidence interval.Numbers of observations for each data point are given in Supplementary TableS10.

Fig. 11 .
Fig. 11.Examples of (a) tail foreshortening, and (b) a curved body shape in larvae.In both cases the standard length of the animals is underestimated.The automatically estimated center line of the animals (marked in red) is affected by foreshortening in (a), and by the high curvature of the fish in (b).Scale bar 1 mm.

Fig. 12 .
Fig. 12. Cropped images of larvae showing: (a) top, (b) side-on views, and (c) an example of an image with partially overlapping eyes.Scale bar 1 mm.