Robust Detection of Retinal Neovascularization in Widefield Optical Coherence Tomography

Retinal neovascularization (RNV) is a vision threatening development in diabetic retinopathy (DR). Vision loss associated with RNV is preventable with timely intervention, making RNV clinical screening and monitoring a priority. Optical coherence tomography (OCT) angiography (OCTA) provides high-resolution imaging and high-sensitivity detection of RNV lesions. With recent commercial devices introducing widefield OCTA imaging to the clinic, the technology stands to improve early detection of RNV pathology. However, to meet clinical requirements these imaging capabilities must be combined with effective RNV detection and quantification, but existing algorithms for OCTA images are optimized for conventional, i.e. narrow, fields of view. Here, we present a novel approach for RNV diagnosis and staging on widefield OCT/OCTA. Unlike conventional methods dependent on multi-layer retinal segmentation, our model reframes RNV identification as a direct binary localization task. Our fully automated approach was trained and validated on 589 widefield scans (17x17-mm to 26x21-mm) collected from multiple devices at multiple clinics. Our method achieved a device-dependent area under curve (AUC) ranging from 0.96 to 0.99 for RNV diagnosis, and mean intersection over union (IOU) ranging from 0.76 to 0.88 for segmentation. We also demonstrate our method’s ability to monitor lesion growth longitudinally. Our results indicate that deep learning-based analysis for widefield OCTA images could offer a valuable means for improving RNV screening and management.

💡 Research Summary

The paper addresses the critical need for early detection and longitudinal monitoring of retinal neovascularization (RNV), a hallmark of proliferative diabetic retinopathy that can lead to severe vision loss. While optical coherence tomography angiography (OCTA) provides high‑resolution, non‑invasive visualization of retinal vasculature, most existing automated RNV detection algorithms have been developed for conventional, narrow‑field (≈6 × 6 mm) scans. Recent commercial OCTA devices now offer widefield imaging (up to 26 × 21 mm), which captures peripheral retinal regions where RNV frequently develops, but the larger field of view introduces substantial anatomical and optical variability that challenges traditional segmentation pipelines.

To bridge this gap, the authors propose a fully automated, two‑stage deep learning framework that operates directly on widefield OCT/OCTA volumes without relying on fixed‑height slab segmentation. The first stage segments the vitreoretinal interface (VRI), the boundary between the retina and the vitreous. This is achieved with a U‑Net‑based convolutional neural network that takes as input three consecutive B‑scans, each containing paired structural OCT and OCTA data, concatenated into a six‑channel tensor. Multi‑scale convolutions (1 × 1, 3 × 3, 5 × 5) capture both fine and coarse features, while a composite loss (binary cross‑entropy + Dice, weighted 0.5 each) balances pixel‑wise accuracy and overlap quality, handling the class imbalance inherent in boundary segmentation.

Using the predicted VRI, the algorithm defines a dynamic “vitreous slab” that extends from the VRI upward without a preset height, thereby accommodating the variable vertical extent of neovascular lesions. Five en face images are then generated from this slab: (1) vitreous OCT, (2) vitreous OCTA, (3) ganglion cell complex (GCC) OCT, (4) GCC OCTA, and (5) a “subtracted OCTA” image obtained by scaling and subtracting the GCC OCTA signal from the vitreous OCTA. This multi‑modal input provides complementary structural and flow cues while suppressing artifacts from superficial retinal vessels that can leak into the vitreous slab due to segmentation errors.

The second stage employs another U‑Net‑style network that ingests the five‑channel en face stack to simultaneously perform binary RNV classification and pixel‑wise lesion segmentation. The model was trained and validated on a heterogeneous dataset comprising 589–634 widefield scans collected from three commercial platforms (Solix, DREAM OCT, S1‑OCTA) across two institutions (Casey Eye Institute, Oregon Health & Science University, and Aichi Medical University, Japan). All scans were pre‑processed with column‑wise z‑score normalization and a fixed intensity threshold (‑0.5) to reduce device‑specific intensity bias and suppress low‑signal noise.

Performance was evaluated using receiver operating characteristic (ROC) area under the curve (AUC) for diagnosis and mean intersection‑over‑union (IoU) for segmentation. Device‑specific AUCs ranged from 0.96 to 0.99, indicating excellent discriminative ability, while mean IoU values fell between 0.76 and 0.88, reflecting robust lesion delineation despite the widefield context. Importantly, the same model could be applied to longitudinal follow‑up scans, enabling quantitative tracking of lesion growth or regression over time—a capability directly relevant to treatment monitoring.

Key contributions of the work include: (1) reframing RNV detection as a direct binary localization problem that avoids the pitfalls of fixed‑height slab definitions; (2) leveraging VRI‑guided dynamic slab generation together with a rich five‑channel en face representation to handle peripheral anatomical variations and optical aberrations; (3) demonstrating cross‑device generalizability on a large, multi‑center dataset; and (4) providing a framework for automated, quantitative longitudinal assessment of RNV.

The authors acknowledge several limitations. The preprocessing pipeline relies on a hard intensity threshold, which may be suboptimal for extremely low‑quality images or non‑standard acquisition protocols. The current approach operates on 2‑D en face projections, leaving 3‑D volumetric quantification of neovascular volume as future work. Additionally, ground‑truth annotations required expert manual delineation, which can be time‑consuming and may introduce inter‑observer variability despite consensus review.

In conclusion, this study delivers a practical, high‑performing solution for automated RNV detection and segmentation in widefield OCTA, paving the way for non‑invasive, large‑scale screening and precise monitoring of proliferative diabetic retinopathy in clinical practice.