This research was conducted in accordance with the ARVO Statement for the Use of Animals in Ophthalmic and Vision Research and was approved by the Animal Care and Use Committee of the National Eye Institute. C57BL/6J mice were reared under 50-104 lux cyclic lighting (12 h:12 h) with food and water available ad libitum. Seven 2-month-old male C57BL/6J mice were intraperitoneally injected with 25 mg/kg of NaIO₃. The size of animal used in this study is typical for investigating SI-induced retinal degenerations (Arifin and Zahiruddin, 2017; Moriguchi et al., 2018; Kim et al., 2022).

OCT images were collected from both eyes of the animals before (baseline) and 1, 3, 6, 13, and 20 days post-SI injection. Prior to imaging, the animals were anesthetized with xylazine 6 mg/kg and ketamine 100 mg/kg. Drops of tropicamide (1%) and phenylephrine (2.5%) were then topically applied to the eyes for pupil dilation. OCT images of the mice were obtained with Bioptigen's Envisu R2210 system (axial resolution 1.7 μm/pixel), using a protocol for 4 B-scans of the retina (1.4mm) at 45-degree intervals with each B-scan sampled 40 times. Averaged Bscan images (4 per eye) were fed to AI models for retinal layer segmentation. A total of 112 image sets were analyzed at each time points.

eAC were calculated based on published methods (Vermeer et al., 2013; Bissig et al., 2025). Briefly, OCT b-scans were transformed so that grayscale pixel values were linearly proportional to the power of the signal measured by the device. In each A-scan, attenuation is described by pixel size, the brightness of a pixel, and the relative brightness of remaining pixels more distant from the OCT device. In our commercial system, there was very little variance in retinal distance from the OCT sensor so the signal decay factor was ignored in these calculations [using the variable names of Vermeer et al. (2013), we stipulate I(z) = U(z)-N(z)]. Acknowledging this slight change from Vermeer et al. and the existence of other strategies to calculate the attenuation coefficient (Chang and Bowden, 2019), we refer to our results as estimated attenuation coefficients (eAC) with units of m⁻¹ and expressed on a logarithmic scale.

The vitreous portion of the image was identified as everything above the first segmentation line (boundary of inner limiting membrane with vitreous). Particles in vitreous were determined using the “Analyze Particles” function in ImageJ. A size range of 5 to 500 was selected to exclude small artifacts and the large optic nerve head in images.

The program for building of machine learning models were described previously (Berkowitz et al., 2022). Briefly, the model was based on a U-net convolutional neural network trained using the “dice loss” function and the Adam optimizer (learning rate = 0.001). A shortest-path algorithm was applied to improve the model's performance. Using “U-net” architecture, five models were built from annotated OCT images with training parameters varied in the number of convolution layers, kernel, starting neurons, and pool layers. Five attentional models were built with the same training parameters but from annotated OCT images pre-processed using a contrast-enhancing algorithm (Girard et al., 2011). Furthermore, two additional models were made with “deeplabv3plus” architecture. A custom-made MATLAB program displayed OCT images superimposed with the segmentation lines generated by each model, along with a segmentation line based on the median values of selected models. The accuracy of each segmentation line was verified manually by either selecting best fit from model-predicted pools or making manual corrections using Pisces2D (v2.1.8342, Voxeleron). As OCT images were captured with the optic nerve head centered, segmentations on either side of the optic nerve head were independently evaluated.

For each B-scan, two-200-pixel (A-scan) region (350 to 630 um away from center of ON head) were selected for measurement (Gao et al., 2021). Unless specified otherwise, averaged values from 8 ROIs were used to represent each eye and data are presented as mean ± SEM. Paired t-tests were used to check for significance of change from the baseline values. Statistics were performed with Prism (V10.2.3, GraphPad software).

Figure 1A shows examples of retinal OCT images captured from the same eye at indicated time points before and after SI treatment. We generated 10 “U-net” models (normal retina models) based on annotated OCT images previously captured from normal mouse retinas. These models segment OCT images with nine lines to produce seven retinal layers and a choroid layer (Figure 1B). The normal retina models can successfully segment OCT images captured at baseline, PI1, and PI3, with very few (< 0.2%) manual corrections needed. To compare each model, we calculated usage rate, ratio of image user selected as the best fit vs. total number of images analyzed. Usage rates for each normal retina model are summarized in Figure 2A (with details listed in Supplementary Table 1 and DICE score listed in Supplementary Table 2). We also calculated the median point for every A-scan from values 10 models predicted and called “Model_median”, which had the highest usage. On the other hand, each of 10 individual models had its own success rate indicating all models are useful in predicting segmentation lines.

However, normal retina models performed poorly for images captured at PI6, PI13, and PI20 when retinal degeneration becomes imminent. We generated 11 models using both “U-net” and “deeplabv3plus” architectures (degenerative retina models). Training dataset includes annotated OCT images previously captured from rd10 mice and images from normal retina with the photoreceptor layer artificially reduced. These models contain 8 segmentation lines to produce 6 retinal layers and a choroid layer (Figure 1C). Usage rates for these models in analyzing images captured at PI6, PI13, and PI20 are summarized in Figure 2B (with details listed in Supplementary Table 3 and DICE score listed in Supplementary Table 4). The median point for every A-scan from values 11 models predicted and called “DR_median”, which also had the highest success rate, whereas all individual “degenerative retina” models were useful in predicting segmentation lines. All OCT images were able to be analyzed with existing models and no manual correction was needed.

Example segmentation lines are superimposed on OCT images in Figure 1A. Images from baseline, PI1, and PI3 are segmented with “normal retina” model, which uses nine segment lines to separate an image into eight layers: the nerve fiber layer (NFL), inner plexiform layer (IPL), inner nuclear layer (INL), outer plexiform layer (OPL), outer nuclear layer (ONL), inner and outer segment layer (IS + OS), retinal pigment epithelium (RPE) and choroid layer. The magnified view of a baseline OCT image outlined by a yellow box is shown in Figure 1B. It better-illustrates the placement of the nine lines. Starting at PI6, retinal degeneration became more extensive, illustrated by thinning of photoreceptor layers, merging of photoreceptor and RPE layers, and reduced prominence of the external limiting membrane (ELM). OCT images obtained from these injured retinas could not be adequately processed with “normal retina” model. We developed “degenerative retina” model to analyze these images, with eight segmentation lines to separate seven layers. As the IS + OS and RPE layers merged, we used the outer retina (OR) to define this merged layer. A magnified image of the region outlined by a yellow box for PI20 OCT is shown in Figure 1C.

Using the “normal retina” model, the RPE layer can be separated from OCT images captured from baseline, PI1, and PI3 eyes. As shown in Figure 3A, there is a significant increase in RPE layer thickness after SI injection for both PI1 and PI3, consistent with swelling of RPE cells in response to SI. We used estimated Attenuation Coefficient (eAC) to probe SI induced reflectance changes on the RPE layer (Vermeer et al., 2013; Moriguchi et al., 2018; Chang and Bowden, 2019; Sakai et al., 2024). Examples of eAC images calculated from the original OCT images are shown in Figure 3E. The effects of SI treatment on RPE layer reflectivity are summarized in Figure 3B. There is a significant reduction of reflectance in RPE cells after SI application.

SI is known to induce photoreceptor degeneration. Figure 3C plots ONL thickness during the course of the study, which showed gradual reduction after SI application. The time course of photoreceptor degeneration can fit with an exponential decay with a half-decay time of ~3 days. Application of SI also changed reflectivity of the ONL layer: there, eAC values are visibly elevated on PI1, and this elevated level of reflectivity is maintained thereafter (Figure 3). Photoreceptor degeneration reduces thickness for both ONL and OR. Consequently, there is a high correlation between ONL thickness and OR thickness following application of SI, as shown in Figure 4A.

We hypothesized that elevated reflectance for the ONL layer could serve as an indication of photoreceptor stress. To test this hypothesis, we calculated the correlation of the eAC values of ONL observed on PI1 and the extent of photoreceptor degeneration at PI20. Results are shown in Figure 4B. A linear fit of the data indicates good correlation (R² = 0.50) between the eAC on P-I1 and amount of ONL reduction on PI-20, suggesting that ONL reflectance could serve as an early marker for photoreceptor degeneration.

It has been suggested that the Dip ratio (OCT intensity of IS band divided by the intensity of the hyporeflective zone between IS and OS, illustrated in Supplementary Figure 1) could serve as an early biomarker for retinal degeneration (Zeng et al., 2024). Figure 4C presents the Dip ratio calculated from OCT images captured at baseline, PI-1 and PI-3. Compared with baseline levels, a significant reduction in Dip ratio was observed at the PI-3 time point, providing additional support that Dip ratios could serve as a biomarker for retinal degeneration. On the other hand, there is no significant different in the Dip ratios between PI-1 and baseline, suggesting this biomarker lags ONL reflectivity in predicting SI-induced degeneration.

In addition to alterations in photoreceptor and RPE layers, application of SI also elicited responses in the inner retinal layers. Changes in inner retinal layer thicknesses are shown in Figure 5A, and the thicknesses of each layer normalized to baseline value are shown in Figure 5B. There is an increase of inner retinal layer thicknesses after SI injection, suggesting swelling of all inner retinal layers. Maximum swelling was observed at PI3, with OPL and INL exhibiting large changes as shown on the normalized plot (Figure 5B). This swelling slowly regresses with time, and at PI20 most inner retinal layers (except NFL) showed a significant reduction of layer thickness than baseline. Table 1 summarizes retinal layer thicknesses at baseline and post-injection changes.

During OCT image acquisition, we noticed the appearance of many hyper-reflective spots floating in the vitreous cavity near the optic nerve head after mice received SI, especially on PI1 (Figure 1A). To quantify these vitreous particles, we isolated vitreous images by subtracting the retina portion from OCT images. Examples are shown in Figure 6A. After SI treatment, many small particles appear in vitreous around the vicinity of the optic nerve head (indicated by red arrows). The number of these small particles was quantified using ImageJ and results are shown in Figure 6B. There is a large increase in vitreous particles after SI injection, and the number progressively decreases with time. By PI13, the vitreous particle number is not significantly different from the baseline.

The application of SI also induces changes in the choroid layer. Figure 7A show examples of segmented choroid layer from OCT images. The mean thickness of choroid at baseline is 88.6 ± 4.2 um (mean ± SD), similar to the published data measured by magnetic resonance imaging (MRI) (Duong, 2014). The effects of SI on choroid thickness are shown in Figure 7B. Choroid thickness showed some reduction at PI1. However, due to the large variance of the value, the difference between PI1 and baseline is not statistically significant. Choroid thickness showed further decrease with time, and all those values are significantly different from the baseline.