Improving segmentation of retinal arteries and veins using cardiac signal in doppler holograms

TL;DR

The paper addresses semantic artery–vein segmentation in time-resolved retinal Doppler holography to enable quantitative hemodynamics. It introduces a sequential pipeline that leverages cardiac-pulse–driven temporal cues—specifically arterial signal extraction, a global cardiac-pulse correlation map denoted $C$, and the $Diasys$ image defined as $D_{sys}-D_{dia}$—alongside standard segmentation networks using $M0$ as input. The results show that temporal preprocessing substantially improves segmentation, with eight models exceeding a Dice score of $0.8$ and simple U-Nets achieving competitive performance when augmented with temporal features. This highlights the importance of time-resolved analysis in Doppler holography and motivates development of DL architectures that explicitly exploit temporal dynamics. The dataset used is publicly available at HuggingFace, facilitating reproducibility.

Abstract

Doppler holography is an emerging retinal imaging technique that captures the dynamic behavior of blood flow with high temporal resolution, enabling quantitative assessment of retinal hemodynamics. This requires accurate segmentation of retinal arteries and veins, but traditional segmentation methods focus solely on spatial information and overlook the temporal richness of holographic data. In this work, we propose a simple yet effective approach for artery-vein segmentation in temporal Doppler holograms using standard segmentation architectures. By incorporating features derived from a dedicated pulse analysis pipeline, our method allows conventional U-Nets to exploit temporal dynamics and achieve performance comparable to more complex attention- or iteration-based models. These findings demonstrate that time-resolved preprocessing can unlock the full potential of deep learning for Doppler holography, opening new perspectives for quantitative exploration of retinal hemodynamics. The dataset is publicly available at https://huggingface.co/datasets/DigitalHolography/

Figures (3)

• Figure 1: Different processing of the M0 video. Its average M0 image (b) looses temporal fluctuations, but offers better visualization of the vasculature, including choroidal vessels, not connected to the centered optic disc. Correlation map (c) and Diasys image (d) both convey retinal hemodynamic information, but they respectively better reveal arteries and veins, as shown by red (artery) and blue (vein) selections.
• Figure 2: Sequential segmentation pipeline. 1. Segmentation of retinal blood vessels, using a classic retinal segmentation model 2. By analyzing the local signal of each labeled vessel from the binary vessel mask, arteries with the most dominant signal are identified. 3. Using the arterial signal, temporal cues are extracted: zero lag cross-correlation with the Power Doppler video gives a correlation map with the arterial signal for each pixel, and the peaks and valleys are used to extract the diastolic and systolic frames. 4. The temporal cues are concatenated with the Power doppler (M0) image, and given as input to a classic artery / vein segmentation model.
• Figure 3: Segmentation results of WNet (top) and U-Net (bottom) trained with different inputs. With only M0, the U-Net shows more misclassifications than WNet (b). Both improve greatly when using the diasys and correlation maps, yielding nearly identical outputs (c). Using only the correlation map misses intersections and some details (d), while using only diasys loses small vessels (e). Combining both without M0 increases false positives (f). Using all temporal cues benefits more to the U-Net. A detailed ablation study will follow in future work.