This study demonstrated the longitudinal effects of excitotoxicity on retinal integrity upon unilateral NMDA injection, using OCT-based deep learning-assisted retinal layer segmentation and thickness monitoring estimation. The retinal thicknesses showed distinct layer-specific temporal patterns, while similar spatial patterns were observed between central and peripheral regions. Within the inner retina, we observed early IPL thinning followed by INL thinning, whereas RNFL initially thickened before normalizing and thinning, implicating the IPL thickness as an early imaging biomarker of excitotoxic retinal degeneration. Within the outer retina, early but slight thinning occurred in ELM, RPE, and BM, followed by normalization of ELM and RPE up to 4 weeks post-NMDA injection, while PRL showed no thinning but delayed thickening at day 28. The distinct temporal patterns of thickness changes across different retinal layers indicated the importance of determining the dynamics of neurodegenerative events for more targeted interventions at different stages of disease progression.

NMDA-induced retinal excitotoxicity manifested in terms of distinct, layer-specific structural changes during disease progression

Excitotoxicity from glutamate uptake impairment and NMDA overstimulation is believed to play a key role in various retinal pathologies [5,6,7]. Our study provides a comprehensive framework for quantitative analyses of layer-specific changes in retinal thickness across 28 days of NMDA-induced excitotoxicity in a rat model. Specifically, to investigate the longitudinal changes in retinal layer thickness, we created an end-to-end artificial intelligence (AI)-assisted automatic pipeline to correct retinal axial motions, segment the retinal layers, and calculate the mean retinal layer thicknesses in the central and peripheral retinas. Our results showed that the total retina of the NMDA-injured eye became significantly thinner compared to the contralateral eye from day 7 up to day 28 after intravitreal injection, with a majority of the thickness reduction attributed to the inner retina. Furthermore, individual layer thickness estimation provided a more sensitive and earlier imaging biomarker than the total retinal thickness, inner retinal thickness, and outer retinal thickness, with the first sign of inner retinal degeneration being observed in the IPL at day 3, followed by INL at day 7. This observation appeared consistent with recent histological studies demonstrating early shrinkage of RGC dendrites and presynaptic connections in the IPL before observable RGC and axonal damage in glaucoma and optic nerve injury models [4, 8, 37,38,39,40,41], demonstrating potential clinical relevance of early IPL alteration in glaucoma patients associated with worsening visual function [10, 42]. On the other hand, the ILM-RNFL in the NMDA-injected eye underwent different levels of increases in thickness at day 3, followed by pseudo-normalization at day 7 and significant thinning at day 28 as compared to the contralateral control eye. This temporal pattern of an increase in thickness preceding a decrease in RNFL suggests inflammation before cellular and axonal death. Prior studies have also reported thickening of the inner retina or RNFL due to inflammation [43,44,45], inflammatory and oxidative stress signaling with NMDA overstimulation [46,47,48], early inner retina thickening with NMDA overstimulation [49], as well as dendritic shrinkage visible by fluorescence imaging prior to ganglion cell complex thinning in other disease models [50]. Taken together, these findings call for caution in differentiating healthy tissues from pseudo-normalization when interpreting changes in retinal thickness, and the need to take longitudinal measurements when examining OCT scans in retinal diseases.

Glutamate uptake impairment is thought to be a major factor in neurological diseases where excess glutamate in the extracellular compartment leads to excessive activation of NMDA receptors and causes excitotoxic damage to neurons [51]. The NMDA receptor ligand-gated calcium channel contains four subunits (GluN2, GluN3, and two GluN1 subunits) [52]. Different subunits have been shown to play layer-specific roles in the retina. For example, GluN2 has been shown to potentially serve a neuromodulatory role in the IPL, whereas increased expression of GluN2B isoform has been implicated in the degeneration of the RGC layer in glaucoma [53]. RGCs exposed to elevated intraocular pressure increased their susceptibility to glutamate-induced death, and subjecting these cells to both elevated pressure and glutamate led to induction of apoptosis and BAX, suggesting glutamate and increased intraocular pressure together may play a part in the pathogenesis of glaucoma [54,55,56]. Several clinical studies have suggested RNFL thickness to be useful in the diagnosis and monitoring of glaucoma [57, 58], while others suggested GCL-IPL or IPL thickness alone were more strongly associated with the severity of disease [18, 19]. On the other hand, INL thickness was found to be relatively unaffected in patients with long-standing glaucoma [19]. However, these studies only measured retinal layer thickness at a single time point. It is essential to evaluate the longitudinal effects of excitotoxicity on the retinal cytoarchitecture and functionality in order to unveil and localize the pathological cascades across retinal layers, to guide early disease detection, and to monitor and optimize targeted neuroprotective treatment [59, 60].

Previous studies using ex vivo immunohistochemistry staining of rat retinal samples after NMDA injection have shown retinal layer thinning and apoptotic changes, especially in the GCL, and the severity of neurodegeneration increased with NMDA dosage [61,62,63]. However, prior histology studies only reported the inner retinal layer degeneration at either day 7 [61, 63] or 14 [62], lacking the ability to evaluate longitudinal effects within the same animals. A recent study introduced in vivo OCT on a chicken model of NMDA-induced retinal injury and reported significant retinal thickness reduction in the IPL, although at a relatively late time point at 14 days after injection [64], and the thickness was derived through manual selection of eight measurement points across the retina. Comparatively, the results of our current study using a rat model showed time-dependent NMDA-induced retinal thickness alterations, especially in the inner retinal layers, across 3 to 28 days after excitotoxic retinal injury. Specifically, our longitudinal thickness analysis revealed significant thinning of the IPL as early as at 3 days post-NMDA injection, with a significant RNFL layer thinning at a later time point (28 days post-NMDA injection) after initial retinal thickening. These observations align temporally with early phases of inflammatory signaling. Furthermore, our study identified NMDA-derived outer retinal layer alterations, indicating a potential neurotoxicity effect towards the outer retina. However, as the outer retinal changes reported are small in the order of 1–2 microns, the axial resolution of the instrument and the accuracy of the algorithm should be taken into account while interpreting the results. Overall, the results of this study provided novel insights about the dynamic and layer-specific patterns of neurotoxicity in the retina. The differential structural changes in retinal thicknesses of the NMDA-injured eyes implicated different pathological processes as well as compensatory mechanisms across retinal layers, which offered an important step to guide further studies to identify the underlying cellular mechanisms at each time point. Last but not least, our findings indicated that OCT with appropriate segmentation protocols could serve as a high-throughput, cost-effective, and non-invasive alternative to complement histological studies. For example, when using histology to assess early mouse retinal changes from 4 h to 7 days after NMDA-induced excitotoxicity, early TUNEL reactivity was found in the INL followed by increased TUNEL reactivity in the GCL and PRL that peaked at 24 h post-NMDA injection [65]. This early neuropathology was accompanied by distinct phases of inflammatory signaling ranging from 24 h to 7 days. The spatiotemporal changes in retinal thickness detected by our AI-assisted OCT imaging generally aligned with these pathological events, but in a non-invasive, in vivo, and longitudinal imaging setting within the same cohort of rats. These technological advancements can allow a close monitoring of the disease progression with or without pre- or post-conditioning in order to facilitate testing of causal pathophysiological mechanisms and neurotherapeutic effects with rigor.

Translational applications of AI-integrated pipeline for automated processing of retinal thickness changes in small animal studies

Recent innovations in AI methods have benefited clinical research substantially. For instance, deep learning-based medical image analysis has been used in ophthalmic big data containing OCT for computer-assisted diagnosis of retinal diseases [22, 66, 67]. In contrast, preclinical animal studies usually involve small sample sizes, hindering the effective adaptation of modern deep learning-based AI applications, which generally require large, labeled samples to train the models accurately. Reverse translation of clinically derived AI methods into preclinical small animal studies may offer a solution to this data availability challenge [30]. A robust automatic retinal layer segmentation pipeline for animal OCT data would significantly improve the processing throughput, measurement repeatability, and analysis accuracy. In this study, we have extended our previously developed and clinically validated deep learning-based automatic retinal layer segmentation framework [27, 28, 68] into adult rat OCT images. It is worth noting that, compared to the human retina, the rat retina lacks a fovea. To this end, the LF-UNet deep learning-based retinal layer segmentation framework that we developed and used in this study was trained on 2D B-scans extracted from the original 3D OCT volume, with a large proportion of 2D B-scans at the non-foveal locations, which ensured that the segmentation model had a good understanding of the general retinal structure across different B-scan locations.

To address the challenges of the lack of standardized retinal layer segmentation labeling, the limited data with ground truth labels of layer segmentation, and the diverse levels of anatomical variations due to experimentally-induced retinal pathology, we have implemented a composition of two techniques, transfer learning [27] based domain adaptation [35] and pseudo-labeling [31], into our existing retinal layer segmentation pipeline, in combination with the data augmentation technique. Firstly, the transfer learning technique used a pretrained model trained from a large number of human retinal data with segmentation labels. Such an approach ensured that the segmentation model parameters were initialized with a good understanding of the general retinal anatomy and OCT image characteristics, significantly reducing the need of training data from rat retinal OCT with ground truth labels in the fine-tuning step. Secondly, during the fine-tuning step, data augmentation was used to introduce random affine transformation to the input data, increasing the range of structural variability even with small training data. The combination of these two steps has shown to allow effective training of AI models in few shots while achieving good performance with few samples [69]. Finally, the pseudo-labeling technique further improved the generalizability of the automatic segmentation model by gradually expanding the training data, so as to propagate and further fine-tune the model parameters with an increased semi-automatic training sample using (pseudo-)ground truth labels from both manual and automatic segmentations. Using this extended framework, we were able to automatically and robustly analyze the neurotoxicity-induced thickness changes across ten retinal layers simultaneously for spatiotemporal assessments.

Limitations and future directions

The current study focused on the in vivo examination of retinal layer thickness as a surrogate measurement of NMDA-derived excitotoxicity. The primary objective is to leverage a well-established animal model and use AI-based OCT image processing and analysis to facilitate the longitudinal assessment of the neurotoxic effects across layers over 4 weeks. While human and rodent eyes share many similarities, there are also challenges in reverse translation across species such as the size differences, intrinsic structural differences including the lack of fovea in rodents, as well as the differences in OCT devices used for collecting clinical and preclinical OCT data that should be further studied. In the current study, a wide margin mask was used to excuse the optic nerve region from deriving the layer segmentation and estimating the thickness analysis. Given the known anatomical differences of the optic discs between human and rodents, further studies with additional optic nerve labeling would be beneficial to segment the optic nerve head structures and analyze the effects of NMDA-induced excitotoxicity towards the optic disc.

A recent study reported the potential induction of retinal degeneration upon intravitreal normal saline injection in C57BL/6J mice [70]. While our previous study using Sprague–Dawley rats did not show apparent retinal thickness changes upon intravitreal normal saline injection [16], future studies can consider phosphate-buffered saline as the diluent of NMDA instead of normal saline to avoid any potential complications. The translatability and generalizability of the current findings in the NMDA-induced excitotoxic retinal injury model should also be validated in other animal models. When analyzing longitudinal data, different approaches could be used to address the research questions of interest. The current study focuses on assessing the excitotoxic effects of NMDA injection compared to the non-injected contralateral eye at each time point to account for the physiological and age-related changes that may occur in the rat retina. Therefore, we used pairwise group comparisons of the mean retinal layer thicknesses between the injected and control eyes at each time point, which intrinsically accounted for the potential longitudinal retinal layer variations in the contralateral non-injected eye. Future studies may consider repeated measures ANOVAs or other statistical models to examine the overall longitudinal thickness variations for both eyes with larger samples. Further histological studies can also be conducted to confirm the retinal layer boundaries and identify cell-type specific responses underlying the morphological changes detected in the current study. We can also combine non-invasive retinal OCT, brain magnetic resonance imaging, and visual functional assessments to determine the interactions between eye, brain, and behavior in health and disease.