DigitalHolography/HoloDopplerSegISBI

Retinal Vessel Segmentation from Holographic Doppler Imaging

This dataset provides paired images and segmentation masks for retinal vessel segmentation from holographic Doppler imaging.
It is designed for research in computational imaging, retinal blood flow analysis, and deep learning-based vessel segmentation.

---

Dataset Summary

Holographic Doppler imaging enables quantitative blood flow mapping in the retina by analyzing temporal fluctuations in optical interferograms.
This dataset includes power Doppler images (M0) and their corresponding artery and vein segmentation masks, representing the spatial distribution of retinal blood flow observed through laser Doppler holography. Temporal informations derived from the arterial signal, namely the correlation map and the diasys image, are also provided. See the article submitted to ISBI2026 (linked soon) for more details

---

Data Description

Attribute	Description
Imaging modality	Holographic Laser Doppler imaging of the retina
Input type	Power Doppler intensity images (M0), Correlation map, Diasys image
Target type	Artery and Vein masks : binary segmentation masks (vessels vs. background)
Classes	0: Background, 1: Vessel
File format	PNG (8-bit grayscale)
Typical image size	1023 × 1023 pixels
Sampling	Retinal field of view from high-speed interferometric acquisitions
Preprocessing	Temporal SVD filtering, Fresnel reconstruction, power Doppler accumulation

---

Dataset Structure

HoloDopplerSegISBI/
├── dataset.py
├── README.md
├── train/
│ ├── measure/
│ │ ├── M0.png
│ │ ├── correlation.png
│ │ ├── diasys.png
│ │ ├── maskArtery.png
│ │ └── maskVein.png
│ └── ...
└── test/
├── measure/
└── ...\

Measures have the following form : id_eye__number, with:

• id corresponding to a unique patient.
• eye equals to 'R' or 'L', respectively for right eye or left eye
• number indicating which measure of the same eye it is

---

Dataset split

The data were split into 121 training and 24 test samples, with no patient overlap between splits.
The number of measures, detailed by eyes, are summarized in the metadata/train_patients.csv and metadata/test_patients.csv
Due to the recent nature of this dataset, the number of measures vary per patients, ranging from 1 measure up to 14, causing potential bias. Still, using the given train/test split, our experiments showed that using all available data yielded the best performance.