Automatic analysis of the retinal avascular area in the rat oxygen-induced retinopathy model

Abstract

Purpose: The aim of this study was to create an algorithm to automate, accelerate, and standardize the process of avascular area segmentation in images from a rat oxygen-induced retinopathy (OIR) model.

Methods: Within 6 h of birth, full-term pups born to Sprague Dawley rat dams that had undergone partial bilateral uterine artery ligation at embryonic day 19.5 were placed into a controlled oxygen environment (Oxycycler, BioSpherix, Parish, NY) at 50% oxygen for 48 h, followed by cycling between 10% and 50% oxygen every 24 h until day 15. The pups were then moved into room air until day 18.5. Ten lectin-stained retinal flat mounts were imaged in montage fashion at 10x magnification. Three masked human reviewers measured two parameters, total retinal area and peripheral avascular area, for each image using the ImageJ freehand selection tool. The outputs of each read were measured as number of pixels. The gold standard value for each image was the mean of the three human reads. Interrater agreement for the measurement of total retinal area, avascular area, and percent avascular area was calculated using type A intraclass correlation coefficients (ICCs) with a two-way random effects model. Automated avascular area identification (A3ID) is a method written in ImageJ Macro that is intended for use in the Fiji (Fiji is Just ImageJ) image processing platform. The input for A3ID is a rat retinal image, and the output is the avascular area (in pixels). A3ID utilizes a random forest classifier with a connected-components algorithm and post-processing filters for size and shape. A separate algorithm calculates the total retinal area. We compared the output of both algorithms to gold standard measurements by calculating ICCs, performing linear regression, and determining the Dice coefficients for both algorithms. We also constructed a Bland-Altman plot for A3ID output.

Results: The ICC for percent peripheral avascular/total area between human readers was 0.995 (CI: 0.974-0.999), with p<0.001. The ICC between A3ID and the gold standard was calculated for three image parameters-avascular area: 0.974 (CI: 0.899-0.993), with p<0.001; total retinal area: 0.465 (CI: 0.0-0.851), with p=0.001; and the percent peripheral avascular/total area: 0.94 (CI: 0.326-0.989), with p<0.001. In the linear regression analysis, the slope for prediction of the gold standard percent peripheral avascular/total area from A3ID was 0.98, with R2=0.975. A3ID and the total retinal area algorithm achieve an average Dice coefficient of 0.891 and 0.952, respectively. The Bland-Altman analysis revealed a trend for computer underestimation of the peripheral avascular area in images with low peripheral avascular area and overestimation of peripheral avascular area in images with large peripheral avascular areas.

Conclusions: A3ID reliably predicts peripheral avascular area based on rat OIR retinal images. When the peripheral avascular area is particularly high or low, hand segmentation of images may be superior.

Figure 1

Comparison between rat and mouse retinas in oxygen-induced retinopathy models; a=ciliary body, b=avascular retina, c=vascular retina, and d=background. There is a central avascular zone in the mouse OIR model and a peripheral avascular zone in the rat OIR model.

Figure 2

Total peripheral avascular area algorithm identification image processing steps. The input for A3ID is a single image of a lectin-stained retinal flat mount from the OIR rat model (A). We used four small selections from the gold standard data set (B) to train the Weka Segmentation Tool to segment images into vascular, avascular retina, and background categories (C). We then used the Shape Filter plugin to distinguish the avascular area from inter-vascular spaces (D) (the rainbow coloring is a product of the IJBlob Shape Filter algorithm and merely indicates the sequence of analysis.) After two post-processing quality steps, the final output is the retinal avascular area (E).

Figure 3

Training the Weka Segmentation Tool. The Trainable Weka Segmentation Tool in FIJI allows users to apply various machine-learning segmentation algorithms to an image. We created a library of four sub-images with manually selected portions of the avascular retina (green), the vascular retina (red), and background (yellow). We then used these to train the Weka classifier.

Figure 4

Dice coefficients for A3ID and total retinal area measurement Dice coefficients (also called F1 measures) are reported for A3ID and the total retinal area algorithm. Note the difference in scale for the two plots (A). Precision and recall values supporting these figures are provided in the table below (B).

Figure 5

Examples of segmentations of the peripheral avascular area. The left column contains the unmodified lectin-stained retinal flat mounts. The middle column contains manual segmentations by one grader, and the right column shows the output of A3ID.

Figure 6

Examples of segmentations of the total retinal area The left column contains the unmodified lectin-stained retinal flat mounts. The middle column contains manual segmentations by one grader, and the right column shows the output of our total retinal area algorithm.

Figure 7

Bland Altman plot of A3ID versus the gold standard. The difference between manual and A3ID avascular area measurements for each of the 10 retinal flat mount images is plotted against the average avascular area of all reads (manual and A3ID) for each image. The positive slope of the line of best fit for this data indicates that A3ID underestimates avascular area in retinal images with low avascular area and overestimates the avascular area in retinal images with large avascular areas.