DSA-SRGS: Super-Resolution Gaussian Splatting for Dynamic Sparse-View DSA Reconstruction

This paper proposes DSA-SRGS, the first super-resolution Gaussian splatting framework for dynamic sparse-view DSA reconstruction, which integrates a Multi-Fidelity Texture Learning Module with confidence-aware supervision and Radiative Sub-Pixel Densification to recover fine-grained vascular details while avoiding blurring and hallucination artifacts.

The Gist

Imagine you are trying to draw a detailed map of a city's underground subway system, but you only have a few blurry, low-resolution photos taken from a distance. If you try to zoom in on these photos to see the tiny tunnels and switches, the image just gets pixelated and fuzzy. This is exactly the problem doctors face with DSA (Digital Subtraction Angiography), a medical imaging technique used to see blood vessels in the brain.

Currently, doctors often have to take many X-ray images from different angles to get a clear 3D picture. This exposes the patient to a lot of radiation. To reduce radiation, they take fewer images (sparse views), but the result is a blurry, low-quality 3D model that misses tiny, crucial blood vessels.

This paper introduces DSA-SRGS, a new "magic lens" that fixes this problem. Here is how it works, using simple analogies:

1. The Problem: The "Blurry Zoom" Trap

Think of traditional methods like trying to make a high-definition poster out of a tiny, pixelated sticker. If you just stretch the sticker (upsampling), it becomes a giant, blurry mess. You can't see the fine lines of the subway tracks (blood vessels), which is dangerous for surgery.

2. The Solution: A Team of Two Experts

The authors created a system that combines two experts working together to rebuild the 3D map:

• Expert A: The "Super-Resizer" (The SR Model)
  Imagine an artist who is amazing at guessing what a blurry photo should look like if it were clear. They take your low-res X-ray and "hallucinate" (predict) the missing details, like adding sharp edges to a sketch.

  • The Risk: Sometimes, this artist gets too creative and draws things that aren't actually there (like a fake tunnel). This is called a "hallucination artifact."

• Expert B: The "3D Builder" (Gaussian Splatting)
  This is the construction crew that builds the 3D model. They use thousands of tiny, glowing "paint splats" (Gaussian kernels) to form the shape of the blood vessels.

3. How They Work Together (The Secret Sauce)

DSA-SRGS doesn't just let Expert A draw whatever they want. It uses two clever tricks to ensure the final map is both sharp and accurate:

Trick #1: The "Trust-but-Verify" Manager (Multi-Fidelity Texture Learning)

This is the most important part. The system acts like a strict editor.

• It looks at the "Super-Resizer's" drawing.
• It asks: "Is this detail confident? Does it match the original blurry photo?"
• If the detail is in a safe, high-confidence area: The editor says, "Great! Keep that sharp detail."
• If the detail looks suspicious or risky: The editor says, "No, that might be fake. Stick to the original blurry shape to be safe."
• Result: You get a sharp image that doesn't invent fake blood vessels. It balances "what we think it looks like" with "what we actually saw."

Trick #2: The "Micro-Expansion" Strategy (Radiative Sub-Pixel Densification)

Imagine you are painting a tree. If you only have big, thick brushes, you can't paint the tiny leaves.

• The system looks at the 3D model and finds areas where the "paint" is too thick or blurry (like the tiny branches of a blood vessel).
• It says, "We need more detail here!" and automatically splits one big "paint splat" into many tiny, high-resolution splats.
• It does this specifically where the image has complex textures, allowing the model to grow tiny, intricate branches that were previously invisible.

The Result

When the doctors use this new system:

1. Less Radiation: They can take fewer X-ray shots because the AI can fill in the gaps intelligently.
2. Sharper Vision: The final 3D model shows tiny, fragile blood vessels that were previously just blurry blobs.
3. No Fake News: The system is careful not to invent fake anatomy, ensuring the doctor sees the real patient's brain.

In a nutshell: DSA-SRGS is like having a team of artists who can take a few blurry photos and reconstruct a crystal-clear, 3D hologram of a patient's blood vessels, while strictly checking their work to make sure they don't draw anything that isn't actually there. This helps doctors perform safer, more precise surgeries.

Technical Summary

1. Problem Statement

Digital Subtraction Angiography (DSA) is a critical imaging technique for diagnosing and treating cerebrovascular diseases. However, current reconstruction methods face two major bottlenecks:

• Sparse-View Limitations: Acquiring dense projections requires high radiation doses and is prone to motion artifacts. Sparse-view acquisition is safer but leads to incomplete data.
• Resolution Constraints: Existing 3D/4D Gaussian Splatting methods applied to DSA are constrained by the resolution of the input low-resolution (LR) projections. Naive upsampling of these inputs results in severe blurring, aliasing artifacts, and loss of fine-grained vascular details (e.g., small branches).
• Domain Gap in Super-Resolution (SR): While general computer vision SR models exist, applying them directly to medical imaging often introduces "hallucination artifacts" (fake textures) that compromise the physical fidelity and clinical reliability of vascular structures.

The core challenge is to reconstruct high-fidelity, high-resolution 4D vascular models from sparse, low-resolution dynamic inputs without introducing structural hallucinations.

2. Methodology: DSA-SRGS

The authors propose DSA-SRGS, the first framework integrating super-resolution learning with 4D dynamic Gaussian Splatting in an end-to-end optimization process. The architecture consists of three key components:

A. Radiative Sub-Pixel Densification (RSD)

Standard Gaussian Splatting relies on the input resolution for detail. To overcome this, the authors introduce an adaptive densification strategy:

• Mechanism: Instead of static densification, the method accumulates gradients from high-resolution sub-pixel sampling during the rendering process.
• Adaptive Splitting: Regions with high gradient magnitudes (indicating texture-rich areas like vessel edges) trigger the splitting of Gaussian kernels.
• Residual Guidance: A residual-guided mechanism actively adds small-scale Gaussian kernels in regions where the difference between the rendered image and the ground truth is high, specifically targeting small vascular branches.
• Dynamic Attenuation: A Dynamic Neural Attenuation Field (DNAF) models the time-dependent concentration of the contrast agent, ensuring the 4D evolution is captured accurately.

B. Multi-Fidelity Texture Learning

To provide high-frequency texture priors without relying solely on potentially hallucinatory SR models, the framework employs a Multi-Fidelity Learning strategy:

• SR Prior Integration: A fine-tuned, DSA-specific Super-Resolution model generates high-resolution pseudo-labels ($I_{SR}$).
• Confidence-Aware Strategy: To prevent the model from learning "illusory" textures, a confidence map ($C$) is generated. This map adaptively weights the supervision signal:
  • High-Confidence Regions: Trusts the SR-generated high-frequency details.
  • Low-Confidence Regions: Reverts to the original upsampled low-resolution observations ($I_{\uparrow}^{LR}$) to maintain structural authenticity.
• Loss Function: The total loss combines a high-fidelity loss (using the confidence-weighted teaching reference) and a low-fidelity loss (strict adherence to original LR inputs) to balance detail recovery with structural truth.

C. Optimization Pipeline

The system unifies the SR model and the 4D Gaussian reconstruction. It uses a teaching reference image constructed via the confidence-aware fusion to supervise the rendering of high-resolution projections, while simultaneously optimizing the Gaussian kernels to match the original low-resolution inputs.

3. Key Contributions

1. First Framework: Introduction of DSA-SRGS, the first super-resolution Gaussian splatting framework specifically designed for dynamic sparse-view DSA reconstruction.
2. Novel Modules:
   • Multi-Fidelity Texture Learning Module: Integrates high-quality SR priors while suppressing hallucination artifacts via a confidence-aware fusion mechanism.
   • Radiative Sub-Pixel Densification: An adaptive strategy that leverages high-resolution gradient accumulation to refine Gaussian kernels in texture-rich regions, enabling the recovery of micro-vascular structures.
3. Clinical Validation: Extensive experiments on two clinical datasets demonstrating superior performance over state-of-the-art (SOTA) methods.

4. Experimental Results

The method was evaluated on two clinical datasets (DSA-15 and DSA-28) covering multi-center samples.

• Quantitative Performance:
  • DSA-SRGS achieved the best results in both PSNR and SSIM across all view settings (30 and 40 views).
  • On DSA-15 (30 views), it achieved 34.323 dB PSNR and 0.8563 SSIM, outperforming the previous best method (4DRGS) by a significant margin.
  • It demonstrated stronger robustness in sparse sampling scenarios compared to competitors.
• Qualitative Performance:
  • Visual comparisons showed that while other methods (R2-Gaussian, TOGS, 4DRGS) suffered from missing structures, rough contours, or lost fine branches, DSA-SRGS successfully restored vascular margins and intricate branching structures with high visual fidelity.
• Ablation Studies:
  • Removing the Confidence-Aware strategy led to SSIM degradation due to structural hallucinations.
  • Using the Multi-Fidelity module was crucial for balancing high-frequency detail recovery with observational constraints.
  • The specific CATANet architecture, fine-tuned on DSA data, proved to be the optimal SR backbone, outperforming general models like SwinIR and Real-ESRGAN.

5. Significance

• Clinical Impact: By enabling high-resolution reconstruction from sparse, low-dose inputs, DSA-SRGS facilitates precision diagnosis and treatment of cerebrovascular diseases, potentially reducing patient radiation exposure while improving the visibility of critical micro-vascular details.
• Technical Advancement: The paper bridges the gap between general computer vision super-resolution and medical imaging by solving the "hallucination" problem through confidence-aware multi-fidelity learning.
• Future Potential: The framework sets a new standard for 4D medical reconstruction, with future work planned to explore self-supervised SR and further optimization of reconstruction speed for real-time clinical application.