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

This paper proposes a method that enhances retinal artery and vein segmentation in Doppler holograms by integrating cardiac signal-derived features into standard U-Net architectures, achieving performance comparable to complex models while leveraging the temporal dynamics of the data.

The Gist

Imagine your eye is a bustling city, and the tiny blood vessels inside it are the roads. Some roads are "arteries" (highways carrying fresh, oxygen-rich blood pumped by the heart), and others are "veins" (roads carrying used blood back). Doctors need to tell these two types of roads apart to spot diseases like diabetes or high blood pressure early.

Usually, doctors take a "snapshot" (a still photo) of these roads. But in this paper, the researchers are using a special camera called Doppler Holography. Instead of a still photo, this camera takes a high-speed video of the blood flowing.

Here is the problem they solved, explained simply:

The Problem: The "Still Photo" Confusion

When you look at a single still frame of this video (which the paper calls an M0 image), the arteries and veins look almost identical. They are both thin, red lines. Trying to tell them apart using just a still photo is like trying to tell a delivery truck from a garbage truck just by looking at a frozen photo of them parked on the street. You can't see which one is moving or what it's doing.

The researchers tried using fancy AI models (like U-Nets, which are like super-smart digital painters) to look at these still photos and guess which is which.

• The Result: The AI got confused. It could see the roads, but it couldn't reliably say, "That's an artery" or "That's a vein." It was like guessing the truck's job based on a blurry photo.

The Solution: Listening to the Heartbeat

The researchers realized they were ignoring the most important clue: Time.

In a video, arteries and veins behave differently because of the heartbeat:

• Arteries pulse hard and fast every time the heart beats (like a drum).
• Veins flow more steadily and gently (like a river).

The team created a new method that acts like a detective listening to the rhythm of the traffic.

1. Step 1: Find the Roads. First, they used a simple AI to just find where the roads are (ignoring whether they are arteries or veins).
2. Step 2: Listen to the Pulse. They analyzed the video to find the "heartbeat" signal. They looked for the exact moments when the heart pumps (systole) and when it rests (diastole).
3. Step 3: Create "Rhythm Maps." They created two special visual maps:
   • The Correlation Map: A heat map showing which parts of the eye are "dancing" in sync with the heartbeat (these are the arteries).
   • The "Diasys" Map: A difference map showing exactly how the blood flow changes between the heartbeat's peak and its low point.

The Magic Trick: Simple Tools, Smart Data

Here is the most surprising part of the paper.

Usually, in AI, people think you need a massive, super-complex robot brain (like a Transformer or a Transformer-based model) to solve hard problems. The researchers tried these complex robots on the "still photos," and they failed.

But when they fed the simple rhythm maps (the Correlation and Diasys maps) into a very simple, basic AI (a standard U-Net), the results were amazing.

• The Analogy: It's like trying to identify a singer.
  • Old Way: You give a complex AI a photo of a singer's face and ask, "Is this a rock star or a jazz singer?" It guesses wrong because they look similar.
  • New Way: You give a simple AI a recording of the singer's voice. Even a simple AI can instantly tell the difference because the rhythm and sound are unique.

The Results

By adding this "heartbeat" information:

• The simple AI models became just as good as the complex ones.
• They could distinguish arteries from veins with over 80% accuracy (up from about 60%).
• They didn't need fancy, expensive computer power; they just needed to look at the right data.

Why This Matters

This paper teaches us a valuable lesson: Sometimes, the data you have is more important than the complexity of the tool you use.

In the world of medical imaging, we often try to build bigger, smarter AI models. But this research shows that if we just take a moment to understand the temporal nature of the data (how things change over time), even a simple tool can perform miracles. It opens the door for better, cheaper, and faster ways to detect eye diseases by simply "listening" to the blood flow.

Technical Summary

1. Problem Statement

Context: Doppler holography is an advanced retinal imaging technique that captures high-temporal-resolution blood flow dynamics, allowing for the quantitative assessment of retinal hemodynamics. This is crucial for diagnosing conditions like diabetic retinopathy, glaucoma, and hypertension.
The Challenge: To extract quantitative hemodynamic data (e.g., flow velocity and direction), retinal arteries and veins must be accurately segmented. However, traditional segmentation methods face significant hurdles in this domain:

• Low Contrast & Noise: Power Doppler images (M0) have lower vessel contrast and resolution compared to standard fundus or OCT images.
• Spatial Ambiguity: Arteries and veins often have overlapping intensity distributions and variable calibers, making them hard to distinguish based solely on spatial features.
• Ineffectiveness of Spatial-Only Models: State-of-the-art deep learning models (U-Net variants, Transformers) designed for static fundus images rely heavily on spatial features. When applied directly to Doppler holograms, they fail to distinguish between arteries and veins because they ignore the rich temporal dynamics inherent in the data.

2. Methodology

The authors propose a sequential segmentation pipeline that integrates cardiac pulse analysis into the standard deep learning workflow. Instead of treating the data as a static image, the method exploits the temporal signal to generate discriminative features.

A. Data Processing & Feature Extraction

1. Input Data: The pipeline starts with raw Doppler holograms processed into M0 images (Power Doppler images representing the integrated Doppler Power Spectrum Density over time) and the raw temporal video sequence.
2. Initial Vessel Segmentation: A standard U-Net performs binary segmentation on the M0 image to identify all blood vessels (Dice score ~0.82).
3. Artery Identification: The vessel mask is skeletonized. The mean temporal signal of each vessel segment is analyzed. Segments showing a positive derivative peak exceeding a threshold (indicating a strong systolic response) are classified as arteries.
4. Temporal Feature Generation:
   • Global Pulse Signal: An average cardiac pulse signal is derived from the identified arteries.
   • Correlation Map: A zero-lag cross-correlation is computed between the signal of every pixel in the video and the global arterial pulse signal. This highlights pixels that pulsate in sync with arteries.
   • Diasys Image: The difference between the average systolic frame (peak) and the average diastolic frame (valley) is calculated. This "Diasys" image emphasizes the dynamic flow changes.

B. Segmentation Architecture

• Input Strategy: The final artery/vein segmentation model receives a multi-channel input consisting of:
  1. The M0 image (spatial/topological info).
  2. The Correlation Map (temporal sync with arteries).
  3. The Diasys Image (temporal flow dynamics).
• Models Tested: The authors evaluated a wide range of architectures, including:
  • Standard: U-Net.
  • Enhanced Fusion: U-Net++, CE-Net, MSU-Net.
  • Iterative/Recursive: IterNet, W-Net, RRWNet.
  • Transformer/Hybrid: Swin-Unet, WNet (Hybrid CNN-Transformer), CSTA-Net.

3. Key Contributions

1. Temporal Preprocessing Pipeline: The primary contribution is a novel preprocessing strategy that converts temporal Doppler data into static feature maps (Correlation and Diasys) that encode cardiac pulse information.
2. Democratization of Model Complexity: The study demonstrates that simple architectures (like standard U-Net) can achieve performance comparable to complex, parameter-heavy models (Transformers, iterative networks) if provided with the correct temporal features.
3. Resolution of the Artery-Vein Gap: The method successfully eliminates the performance disparity where models typically segment arteries well but fail on veins, achieving balanced accuracy for both vessel types.
4. Public Dataset: The authors released a dataset of 145 samples from 47 patients with handcrafted artery/vein masks, addressing the lack of public data in this specific modality.

4. Experimental Results

The experiments were conducted on a dataset of 145 samples (121 training, 24 testing) using five-fold cross-validation.

• Performance Comparison (M0 Only vs. Temporal Cues):

  • Direct Segmentation (M0 only): All models performed poorly in distinguishing arteries from veins. The best model (WNet) achieved a Dice score of ~0.46, while the standard U-Net scored ~0.48. There was a significant gap between artery and vein segmentation accuracy.
  • Sequential Segmentation (with Temporal Cues): Performance improved drastically across the board.
    • U-Net: Jumped from ~0.48 to 0.82 Dice score.
    • WNet: Improved from ~0.46 to 0.84.
    • CE-Net & UNet++: Also reached Dice scores >0.84.
  • Uniformity: The performance gap between different architectures narrowed significantly. The standard U-Net became competitive with the most complex models, suggesting that the quality of input features matters more than architectural complexity in this domain.

• Ablation Study:

  • Adding the Correlation Map provided the most significant boost in discriminative power.
  • Combining Correlation + Diasys yielded the best results, reducing variability (standard deviation) particularly for vein segmentation.
  • The M0 image remained essential for spatial topology, but the temporal cues were the key to semantic classification.

• Efficiency: The simple U-Net with temporal cues achieved results similar to the WNet (a hybrid model) but with significantly fewer parameters (17k vs 69k) and faster inference times.

5. Significance and Conclusion

• Paradigm Shift: The paper challenges the assumption that complex deep learning architectures are required for difficult medical imaging tasks. It proves that time-resolved preprocessing is the critical factor for Doppler holography, unlocking the potential of simpler, more interpretable models.
• Clinical Impact: By enabling robust, automated artery/vein segmentation, this method facilitates the quantitative analysis of retinal hemodynamics. This is a prerequisite for using Doppler holography as a diagnostic tool for systemic diseases affecting microvasculature.
• Future Directions: The findings suggest that future deep learning frameworks for Doppler imaging should explicitly incorporate temporal dynamics (either through preprocessing or 3D/4D convolutions) rather than relying solely on spatial feature extraction.

In summary, the paper demonstrates that leveraging the cardiac pulse signal transforms a difficult semantic segmentation problem into a solvable one, allowing standard U-Nets to outperform complex models and providing a robust pathway for quantitative retinal hemodynamic analysis.