Choroidal Cynomolgus monkey reference database derived from a deep learning hybrid optical coherence tomography segmentation.

Animals and animal husbandry

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 on safety assessment so that the animals were observed sequentially. Therefore, only OCT imaging data from untreated cynomolgus monkeys (Macaca’s footnote) of both sexes in the current study. Thus, no additional animals were examined to obtain these data. Preliminary studies were reviewed and approved by the Institutional Animal Care and Use Committees (IACUC) at the respective institutions. Approval for 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 US National Research Council or Canadian Council on Animal Care guidelines.

To ensure animal safety and care, studies have been previously reviewed and approved by the Institutional Animal Care and Use Committees. The animals were specially bred for laboratory use and obtained from certified suppliers in two geographical regions: Mauritius and Asia. The room temperature was kept constant between 20 °C and 26 °C; Humidity ranged between 20 and 70%, with a light-dark cycle of 12:12h. Nutrition was provided by a standard diet of pellets fortified with fresh fruits and vegetables. Freely available, clean tap water was provided and purified by reverse osmosis and ultraviolet radiation. Psychological and ecological enrichment was introduced to the animals.

October image data

Only foveolar OCT imaging data from healthy cynomolgus monkeys of Mauritian or Asian origin were included. These monkeys are between 30 and 50 months old and weigh between 2.5 and 5.5 kg. OCT measurements were performed under sedation, as previously reported, with pupil dilation using the Spectralis HRA + OCT Heidelberg apparatus (Heidelberg Engineering, Heidelberg, Germany)16. The scanning protocol was the same for all animals and included a horizontal line (centered over the fovea) scan pattern of 20° × 20°, consisting of 25 B-scans spaced 221 μm (scan length 5.3 mm, 512 ×) 496 pixels, scanning depth 1.9 mm. ). Images obtained from the OCT device were exported as an original scan file in Bitmap Data Format (BMP). Only image data with a scan quality of at least 25, which is provided by the manufacturer’s software, is included.

image processing

The obtained images were analyzed by two automated processes (Fig. 1): (1) using a previously developed and validated deep learning (DL) procedure, the OCT images were divided into their corresponding parts16allowing the division of the choroid above the choroid to the junction of the choroid and sclera.

shape 1
shape 1

Illustration of automatic deep learning placental segmentation. (a) The original B-scan was exported from the device followed by (B) Automatic deep learning-based segmentation of the posterior eye into its principal compartments. Deep learning predictions are shown as overlays. The vitreous is highlighted in brown (arrow), retina in blue (two open arrowheads), and choroid in yellow (two white arrowheads), respectively. (c) Then specific measurements were performed on the segmented choroid (yellow, two white heads).

In short, the DL procedure used a modified U-Net architecture21, a type of convolutional neural network (CNN). CNN was trained and validated using a representative subset of the cynomolgus monkey . OCT dataset16. This subset—the ground truth (GT)—contains 1,100 B-scans obtained from 44 eyes from 44 individuals (each eye contributed 25 B-scans). GT annotation was performed by three experienced retina specialists. Then, 44 eyes in GT were randomly assigned to the training, validation, and test set containing 27, 9, and 8 eyes, respectively (675, 225, and 200 B-scans, respectively). Each human trainee annotated 225 and 75 different scans of the training and validation sets. 200 B-scans for the test set were annotated by each human grader (to investigate intergrader agreement for base truth nomenclature). The data in the training set were augmented by applying vertical inversion and adding a random rotation between -8° and 8° for each B-scan, which increased the size of the training set to 2025 B-scans. In the test set, the differences between the CNN predictions and the three human grade annotations were, on average, smaller than the human-grade differences. A detailed description of the ground truth annotation, CNN architecture, training, and evaluation is provided in Maloca et al.22.

(2) The second step of image processing was carried out using a conventional and structure-based deterministic computer vision algorithm to detect the deepest location within the fovea so that the whole approach can be described as hybrid image processing. This algorithm was implemented in C# (v7.0, .NET Framework v4.6). Since the extracted intrinsic specific membrane (ILM) line as a boundary between vitreous and retinal segmentation was somewhat noisy, the extracted ILM was smoothed using a moving average with a 2D sampling window to identify the deepest point within the fovea. Thus, it was possible to automatically identify and define the deepest point of the click from the smooth ILM, which was denoted as nulla16. The null was therefore defined as the deepest locus within a series of OCT B scans examining a given macular OCT volume. This is particularly important because the null corresponds to the thinnest part of the fovea, where the receptors can interact directly with light which is generally thought to be the most acute place of vision. In the case of multiple deep points (usually adjacent to each other), the coordinates of their center of mass were used as the deepest point.

Therefore, from the vacuole as a reference point, an imaginary line was projected perpendicular to the primary retinal pigment epithelium to measure the axial diameter of the choroid. Successive choroidal measurements were performed at distances of 500 µm to the side, up to a maximum distance of 2000 µm from the void.23,24. This allowed the measurement of nine placental diameters (marked as thickness) in the axial direction, as well as eight overlapping placental areas, yielding a total of 17 parameters for estimating placental characteristics, as shown in Figure 2.

Figure 2
Figure 2

Identify choroidal anatomical landmarks in the left eye with respect to the deepest site of the fossa. (a) In the B-sectional examination of healthy macaques, the deepest locus at the bottom was automatically identified and marked as null (red dot). Under the box, successive measurements of choroidal thickness were performed at intervals of 500 μm up to 2000 μm for the side (marked by T1–9 thicknesses, purple diameters). (B) Between the diameters of the choroidal thickness, eight choroidal surface areas (A1–A8, shaded in light blue) were identified and measured. For the central bundle of cones (highlighted in light green), choroidal subfield analysis was similarly performed at 100 μm distances to determine additional choroidal parameters of nasal and temporal thickness (a, white lines) and choroidal morphology. nasal surface areas and time (B). The same procedures were performed for all eyes. Bars = 500 µm.

Due to the importance of the null value as the putative site of the highest receptor density (central cone bundle), further measurements of the choroid were performed to determine whether higher receptor density was also associated with higher choroidal thickness.1,25. Thus, the choroidal thickness and overlapping areas were measured laterally at an interval of 100 μm to the stated null value. Thus, four more values: additional nasal thickness (TUn) and temporal thickness (TUt) at a distance of 100 μm nasal and temporal were added to nulla, respectively, as well as additional nasal choroid area (AUn) and temporal choroid. area (AUt). Including choroidal thickness at the same null, sub-analysis of nulla provided a total of 5 parameters. Due to incomplete records, accurate data on the age and weight of the monkeys were missing. This made it impossible to include these parameters in the analyses.

statistical analysis

For both the parameters of thickness and area measured, summary statistics—mean, standard deviation, minimum, and maximum—were computed for subsets of the data. Summary statistics for the right and left eyes were calculated separately, and Boxplots were used to visualize the distribution of data and differences between subgroups (eg, Mauritius versus Asian ancestry). Regarding null, for placental thickness (T5) and areas of adjacent placental surfaces (A4 and A5), mean values, minimum, maximum, and coefficient of variation (CV) were additionally calculated for all eyes. CV was calculated as a relative measure of dispersion (defined as the ratio of the standard deviation to the mean). Pearson correlation coefficients between thickness and area coefficients were calculated. All calculations were made in Python v3.8.5. Boxplots was created using the Python Seaborn v0.11.1 library. The effect of categorical variables of gender (male, female) and origin (Mauritius, Asia) on each of the measured thickness parameters was investigated by two-way analysis of variance (ANOVA) using a type II sum of squares. Addition of the gender interaction term: origin to the ANOVA analyzes did not change the significance levels of their results. Hence, the reaction conditions were dropped. As some monkeys contributed to both the 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. ANOVA was performed using the statmodels library Python v0.12.1. Semantics of differences between group averages were calculated using the F-statistic, which is part of the application of ANOVA in statsmodels. Bonferroni correction for significance levels was applied to adjust for the multiple test problem by dividing significance levels by nine, the number of statistical tests for each eye.

Leave a Comment

Your email address will not be published. Required fields are marked *