Improving segmentation of retinal arteries and veins using cardiac signal in doppler holograms
📝 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/
💡 Analysis
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/
📄 Content
The retina provides a unique window into microvascular health, with blood flow and vessel morphology reflecting diseases such as diabetic retinopathy, glaucoma, hypertension, and Alzheimer’s. Recent advances in retinal imaging have improved visualization of the vascular network, yet most modalities remain limited to static or qualitative flow representations. Fluorescein and OCT angiography map vessel topology with high resolution, but rely on surrogate markers and cannot capture temporal dynamics throughout the cardiac cycle. Laser Doppler flowmetry, laser speckle contrast imaging, and ultrasound Doppler measure local flow but lack sufficient spatial resolution to resolve retinal layers or specific vessels [1]. Doppler holography, by contrast, enables simultaneous imaging of vessel morphology and hemodynamics at high frame rates [2], directly quantifying Doppler shifts induced by moving red blood cells and providing timeresolved measurements of flow velocity and direction over a wide field [1].
Extracting quantitative hemodynamic information requires semantic segmentation of retinal arteries and veins, a challenging task due to overlapping intensity distributions, variable vessel calibers and image noise. Traditional approaches rely on hand-crafted features such as vessel color, width, or proximity to the optic disc, while modern methods increasingly use deep learning architectures, particularly U-Net variants. Many are fine-tuned for fundus images, which exhibit a wide range of vessel calibers and resolutions. To handle these multiscale challenges, some methods adopt iterative strategies, specialized convolution layers, spatial attention, or graph-based networks.
Most of these spatially tuned strategies, however, lose effectiveness with Doppler holography. Power Doppler images [3] have lower vessel contrast and resolution, fewer small vessels, and overlapping retinal and choroidal vasculature. In this context, temporal information becomes the most informative dimension. By analyzing the pulse signal over time, arteries and veins can be reliably distinguished. Using this approach, we train several models and achieve highly satisfactory results, with simple U-Net-like architectures performing on par with more complex iterative, attention-based, or Transformer-based networks.
Doppler Holography exploits Doppler shifts induced by moving red blood cells to reveal the vascular network and quantify blood flow velocity and direction.
Using a Mach-Zehnder inline interferometer, a diffuse laser beam illuminates the retina. Its backscattered light interacts with a reference beam, creating interferogram patterns that are captured by a high-speed camera. The current device setup is depicted in [1], while [4] explains the experimental protocol.
The raw 512x320-pixel interferograms are acquired at 37.000 fps using the dedicated open-source acquisition and real-time rendering software Holovibes [2]. Offline rendering of the raw interferograms is then performed using the opensource Matlab program Holodoppler [5]. The interferograms are reconstructed at the retina plane via Fresnel propagation, then eye-motion artifacts are removed using singular value decomposition filtering [3]. High-pass filtering of the temporal frequency spectrum obtained from the short-time Fourier transform (512-frame windows) suppresses low-frequency components from static tissue and eye motion, retaining the frequency shifts associated with moving red blood cells [6].
The Doppler Power Spectrum Density (DPSD) is obtained by computing the squared magnitude of the filtered Fourier spectrum for each time window. The moment of order zero of this spectrum (denoted M0), obtained by integrating the DPSD over all frames, enables the visualization of blood flow dynamics across the cardiac cycle. It is the data used throughout our pipeline. The accumulation of each frame produces Power Doppler images (M0 image), providing a high-resolution map of the retinal vasculature, but losing its temporal fluctuations. An example of a M0 image is shown in figure 1.
Using a light scattering model, blood flow velocity in retinal arteries and veins is calculated by comparing local DPSD broadening in vessels with neighboring tissue [1], which requires semantic segmentation of arteries and veins.
The input of the segmentation pipeline are the rendered power Doppler videos and their averaged image, denoted M0. A private dataset of 145 samples over 47 different patients with handcrafted artery/vein masks was used to train diverse traditional and state of the art segmentation models, following different strategies. The dataset is publicly available.
The models used in this study mostly represent successive evolutions of the U-Net architecture for semantic segmentation, differing mainly in how they enhance basic convolutional units, integrate multi-scale context, refine predictions iteratively, or capture global dependencies.
The U-Net [7] serves as the foundation,
This content is AI-processed based on ArXiv data.