Cynomolgus Monkey Choroid Reference Database Derived from Hybrid Deep Learning Optical Coherence Tomography Segmentation
Animals and breeding
A retrospective analysis of OCT data from studies conducted as part of routine pharmaceutical product development support was performed19.20. The purpose of these studies was to obtain OCT data for safety assessment so that animals were observed sequentially. Therefore, only OCT imaging data from untreated cynomolgus monkeys (macaca fascicularis) of both sexes were collected in the present study. Thus, no additional animals were examined to obtain these data. The primary studies were reviewed and approved by the Institutional Animal Care and Use Committees (IACUCs) of the respective institutions. Approval of the studies was granted by one of the following IACUCs: Charles River Laboratories Montreal, ULC Institutional Animal Care and Use Committee (CR-MTL IACUC), IACUC Charles River Laboratories Reno (OLAW Assurance No. D16-00594) and Institutional Animal Care and Use Committee (Covance Laboratories Inc., Madison, WI) (OLAW Assurance #D16-00137 (A3218-01) In this study, animals were handled and used strictly according to the guidelines of the US National Research Council or the Canadian Council on Animal Care.
To ensure animal safety and welfare, studies were reviewed and approved in advance by institutional animal care and use committees. Animals were bred specifically for laboratory use and sourced from certified suppliers in two geographic regions: Mauritius and Asia. The room temperature was kept constant between 20°C and 26°C; humidity was 20-70%, with a 12:12 light-dark cycle. Feed was provided via a standard diet of pellets enriched with fresh fruits and vegetables. Clean, freely available tap water was provided and purified by reverse osmosis and UV irradiation. The animals were offered attractive psychological and environmental enrichment.
OCT image data
Only OCT foveolar imaging data from healthy cynomolgus monkeys of Mauritian or Asian origin were included. These monkeys were between 30 and 50 months old and weighed between 2.5 and 5.5 kg. OCT measurements were performed under anesthesia, as previously reported, with the pupil dilated using the Spectralis HRA + OCT Heidelberg device (Heidelberg Engineering, Heidelberg, Germany)16. The scanning protocol was the same for all animals and included a horizontal line scan pattern (centered on the fovea) with a size of 20° × 20°, consisting of 25 B-scans spaced 221 μm (scan length 5.3 mm, 512 × 496 pixels, scan depth 1.9 mm). The obtained images were exported from the OCT device as an original B-scan file in bitmap image data (BMP) format. Only image data with a scan quality of at least 25, provided by the manufacturer’s software, was included.
The obtained images were analyzed via two automatic processes (Fig. 1): (1) Using a previously developed and validated deep learning (DL) procedure, the OCT images were segmented into their corresponding compartments16allowing the choroid to be segmented just above the choriocapillaris to the choroid-sclera junction.
In summary, the DL procedure used a modified U-Net architecture21, a type of convolutional neural network (CNN). CNN training and validation was performed using a representative subset of the cynomolgus monkey OCT dataset16. This subset – the ground truth (GT) – contains 1100 B-scans obtained from 44 eyes of 44 individuals (each eye contributed 25 B-scans). GT annotation was performed by three experienced retina specialists. Subsequently, the 44 GT eyes were randomly assigned to a training, validation, and test set containing 27, 9, and 8 eyes, respectively (675, 225, and 200 B-scans, respectively). Each human proofreader annotated 225 and 75 different B-scans for the training and validation sets. The 200 B-scans in the test set were annotated by each human rater (to study inter-rater agreement of ground truth labels). The training set data was augmented by applying vertical mirroring and adding random rotation between −8° and 8° degrees to each B-scan, thus increasing the size of the training set at 2025 B-scans. On the test set, the differences between the CNN predictions and the annotations of the three human raters were, on average, smaller than the differences between the human raters. A detailed description of ground truth annotation, CNN architecture, training and evaluation is provided in Maloca et al.22.
(2) The second step of image processing was performed using a deterministic and classical structure-based computer vision algorithm to detect the deepest location in the fovea so that the whole approach can be described as hybrid image processing. This algorithm has been implemented in C# (v7.0, .NET Framework v4.6). Since the extracted inner limiting membrane (ILM) line as the boundary between vitreous and retinal compartment segmentation was rather noisy, the extracted ILM was smoothed using a moving average with a sampling window two-dimensional to determine the deepest point of the fovea. . Thus, it was possible to automatically identify and define the deepest point of the fovea from the smoothed ILM, which was denoted by the nulla16. The nulla was therefore defined as the deepest position in a series of OCT B-scans of a particular macular OCT volume scan. This is particularly important because the nulla is the thinnest part of the fovea, where receptors can most directly interact with light and which is commonly considered the place of sharpest vision. In the case of several deepest points (usually adjacent to each other), the coordinates of their center of mass were used as the deepest point.
Therefore, using the nulla as a reference point, an imaginary line was projected orthogonally to the underlying retinal pigment epithelium to measure the axial diameter of the choroid. Successive measurements of the choroid were made at distances of 500 µm laterally, up to a maximum distance of 2000 µm from the nulla23.24. This allowed the measurement of nine choroidal diameters (marked as thicknesses) in the axial direction, as well as eight of the intermediate choroidal areas, giving a total of 17 parameters for the quantification of choroidal properties, as shown in Fig. 2.
Given the importance of the nulla as the putative site of highest receptor density (central cone bunch), further measurements of the choroid were performed to determine if higher receptor density was also associated to a higher choroidal thickness.1.25. Thus, the choroidal thickness and the intermediate choroidal areas were measured laterally at an interval of 100 μm from the mentioned nulla. Thus, four additional values were added: an additional nasal thickness (TUn) and a temporal thickness (TUt) at a distance of 100 µm nasal and temporal to nulla, respectively, as well as an additional nasal choroid area (AUn) and a temporal choroid. area (AUT). Including the choroidal thickness at the level of the nulla itself, subanalysis of the nulla provided a total of 5 parameters. Due to incomplete records, accurate data on the age and weight of the monkeys was missing. This made it impossible to include these parameters in the analyses.
For each of the thickness and area coefficients measured, summary statistics (mean, standard deviation, minimum, and maximum) were calculated for subsets of data. Summary statistics were calculated for left and right eyes separately, and box plots were used to visualize the distribution of data and differences between subgroups (e.g. Mauritian versus Asian origin) . With respect to the nulla, for the choroidal thickness (T5) and areas of its adjacent choroidal surfaces (A4 and A5), mean, minimum, maximum, and coefficient of variation (CV) mean values were further calculated for all eyes. The CV was calculated as a relative measure of the dispersion (defined as the ratio of the standard deviation to the mean). The Pearson correlation coefficients were calculated among the thickness and surface coefficients. All calculations were performed in Python v3.8.5. The boxplots were generated using the Seaborn v0.11.1 Python library. The impact of the categorical variables of sex (male, female) and origin (Mauritius, Asia) on each of the measured thickness coefficients was studied by a two-way analysis of variance (ANOVA) using a calculation of the sum of the type II squares. Adding the sex:origin interaction term to the ANOVA analyzes did not change the significance levels of their results. Thus, the interaction terms were removed. Since some monkeys contributed both left and right eyes, these eyes were not independent of each other and were analyzed separately. The 374 eyes contained 16 eyes of unknown origin, which were excluded from the ANOVA analyses. The ANOVA was performed using the Python library statsmodels v0.12.1. The significances of the differences between the group means were calculated using the F statistic, which is part of the ANOVA implementation of statsmodels. Bonferroni’s correction of significance levels was applied to adjust for the multiple testing problem by dividing the significance levels by nine, the number of statistical tests per eye.