A generative AI model for accurate reconstruction of retinal blood flow from static OCT scans

Abstract

Purpose : A newly developed optical coherence tomography (OCT) modality, called OCT Angiography (OCT-A) allows noninvasive visualization of underlying blood flow and vasculature. It currently drives advances in the understanding of prevalent chronic retinal diseases, such as age-related macular degeneration (AMD). However, this technology is not available in all clinics and even when it is available, relative to standard (static) OCT imaging, the acquisition process of OCT-A is time consuming and prone to signal quality issues, making OCT-A currently ill-suited to the clinic.

Methods : In this work, we develop an AI-based generative model for reconstructing OCT-A representations using standard OCT scans. We also introduce local contrast regularization for the loss function used during training, which enhances the overall performance of our model. Using 500 standard OCT scans acquired from independent patients, we compared our model's ability to infer 2D OCT-A scans for the superficial and choriocapillaris (CC) layers. The layer reconstruction performance was measured by three commonly used metrics: Mean Absolute Error (MAE), Structural Similarity Index Measure (SSIM), and Flow Deficit Error (FDE).

Results : We demonstrate that our proposed method is able to generate superficial and CC OCT-A en face slabs at a higher quality compared to competing methods. Our proposed model outperformed state-of-the-art (SoTA) AI models significantly in SSIM, MAE, and FDE scores, with improvements of 0.25, 0.05, and 0.08, respectively. Moreover, previous studies showed that Flow Deficit (FD) is correlated with age. We found that the CC FD values calculated from the reconstructed images are indeed significantly correlated with age (R2=0.39; p-value<0.05) corroborating our model’s soundness.

Conclusions :
Our framework surpasses current SoTA methods, generating superior superficial and CC OCT-A en face slabs from standard OCT scans for CC FD assessment. This approach promotes clinical research as it allows the regeneration of choriocapillaris OCT-A en face slabs from historical (standard OCT) datasets for which OCT-A was not captured. Moreover, it eliminates the complex and expensive OCT-A acquisition process, which in turn, will enhance diagnostic efficiency and accessibility in clinical settings, representing a notable advancement for the ophthalmology field.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.