CASR-Net: An Image Processing-focused Deep Learning-based Coronary Artery Segmentation and Refinement Network for X-ray Coronary Angiogram

This paper introduces CASR-Net, a three-stage deep learning pipeline featuring a novel multichannel preprocessing strategy and a Self-ONN-based UNet architecture that achieves state-of-the-art coronary artery segmentation and refinement on X-ray angiograms, thereby enhancing the accuracy of coronary artery disease diagnosis.

The Gist

Imagine your heart is a bustling city, and the coronary arteries are the main highways delivering oxygen and fuel to every neighborhood. When these highways get clogged with "traffic jams" (plaque buildup), it's called Coronary Artery Disease (CAD). If the traffic stops completely, it's a heart attack.

Doctors use a special X-ray camera called an angiogram to take pictures of these highways. But here's the problem: these pictures are often blurry, dark, and full of "fog" (noise). It's like trying to find a specific, thin thread of silver wire tangled inside a messy ball of gray wool.

This paper introduces a new AI detective named CASR-Net (Coronary Artery Segmentation and Refinement Network). Its job is to look at these messy X-ray photos and perfectly trace the highways, even the tiny, broken, or clogged ones.

Here is how CASR-Net works, broken down into three simple steps:

1. The "Photo Enhancement" Step (Preprocessing)

Before the AI even looks at the picture, it cleans it up. Think of this like a photographer using two different filters to make a dark, foggy photo pop.

• Filter A (CLAHE): This is like a smart flashlight that brightens the dark corners without blowing out the bright spots. It makes the contrast between the road (artery) and the background (fog) much sharper.
• Filter B (Improved Ben Graham): This is like a "noise-canceling" headphone for images. It subtracts the background "static" so the road stands out clearly.
• The Magic Combo: Instead of using just one filter, CASR-Net uses both at the same time on different layers of the image. It's like wearing 3D glasses where one lens brightens the image and the other removes the fog. The result? A crystal-clear view of the highways.

2. The "Tracing" Step (Segmentation)

Now that the image is clear, the AI needs to draw a line around the arteries.

• The Old Way: Most AI models use a standard "brick-and-mortar" approach (called Convolutional Neural Networks). They are good at drawing thick, straight walls, but they often break when the road gets very thin, twisted, or clogged. It's like trying to trace a winding river with a square stamp; you miss the curves.
• The CASR-Net Way: This paper introduces a special tool called Self-ONN. Imagine a standard AI as a robot that follows a strict rulebook. Self-ONN is like a chameleon. As it looks at the image, it can change its own "shape" and "rules" on the fly to fit the curve of the artery.
  • If the artery is a straight highway, it acts like a straight ruler.
  • If the artery is a tiny, broken, clogged branch, it morphs to trace that tiny, jagged line perfectly.
  • This ensures that even the most fragile, broken parts of the road are connected, preventing the AI from saying, "Oh, the road stopped here," when it actually just got thin.

3. The "Cleanup Crew" Step (Refinement)

Even the best AI makes mistakes. Sometimes it draws a line where there isn't one (a "false positive"), or it misses a tiny piece of a road (a "false negative").

• The Cleanup: CASR-Net has a final "quality control" team.
  • Contour Refinement: This is like a gardener trimming the hedges. If the AI accidentally drew a tiny, isolated blob of green that isn't part of the main road, the gardener snips it off.
  • Patch Line Generation: If the AI drew a road but left a tiny gap in the middle (a broken bridge), this tool acts like a construction crew. It looks at the two ends of the gap and builds a tiny bridge to reconnect them, ensuring the highway is continuous.

Why Does This Matter?

In the past, doctors had to squint at these blurry X-rays and guess where the clogs were. If the AI missed a tiny, broken branch, a doctor might miss a dangerous blockage.

CASR-Net is like a super-powered, tireless assistant that:

1. Clears the fog so the roads are visible.
2. Morphs its shape to trace every twist and turn, no matter how thin or broken.
3. Fixes its own mistakes before showing the result to the doctor.

The Results

When the researchers tested this new AI, it was significantly better than all the previous "detectives." It successfully traced the highways with high accuracy, even in the most difficult, clogged cases.

In short: CASR-Net turns a messy, confusing X-ray into a clear, connected map of the heart's highways, helping doctors spot dangerous blockages earlier and save more lives.

Technical Summary

1. Problem Statement

Coronary Artery Disease (CAD) is a leading cause of global mortality. Early detection relies heavily on X-ray Coronary Angiography (XCA), which is the clinical gold standard for visualizing coronary arteries. However, automated segmentation of these vessels faces significant challenges:

• Image Quality Issues: XCA images often suffer from uneven illumination, low resolution, poor contrast, and low signal-to-noise ratios.
• Structural Complexity: Vessels are often narrow, tortuous, and overlapping. Stenotic (narrowed) regions often appear as discontinuities or breaks in the vessel structure.
• Artifacts: Catheters, the diaphragm, and the spine can obscure vessel boundaries.
• Limitations of Current Methods: Existing deep learning models often struggle to preserve the continuity of thin, stenotic branches and frequently produce false positives or fragmented masks, limiting their clinical utility for precise diagnosis and treatment planning.

2. Methodology: The CASR-Net Pipeline

The authors propose CASR-Net, a three-stage pipeline designed to enhance image quality, segment vessels accurately, and refine the output.

Stage 1: Multichannel Preprocessing

To address poor contrast and illumination, a novel multichannel strategy is employed:

• CLAHE (Contrast Limited Adaptive Histogram Equalization): Applied to one channel to enhance local contrast and vessel visibility while limiting noise amplification.
• Improved Ben Graham Method: Applied to a second channel. This is a modified version of a technique originally used for retinal images. It involves subtracting the local average color to accentuate vessels. The authors improved this method by integrating Canny edge detection and morphological operations (erosion/dilation) to eliminate circular patch artifacts and shadow effects common in angiograms.
• Concatenation: These two processed channels are concatenated to form a multichannel input, providing complementary contrast features to the network.

Stage 2: Segmentation Network (Core Innovation)

The segmentation backbone is a modified U-Net architecture with specific enhancements:

• Encoder: Uses DenseNet121 (pre-trained on ImageNet) to extract rich hierarchical features.
• Decoder: Replaces standard Conv2D layers and ReLU activations with Self-Organized Operational Neural Network (Self-ONN) blocks and Tanh activations.
  • Why Self-ONN? Unlike standard CNNs that use fixed linear operators, Self-ONNs employ generative neurons that can autonomously optimize their nodal operators (modeled as Q-th order Taylor approximations) during training. This allows the network to better adapt to the non-linear, complex geometry of coronary arteries, specifically preserving continuity in narrow and stenotic branches.
• Loss Function: The model is trained using Dice Loss, which was found to outperform Binary Cross-Entropy (BCE) and hybrid losses in this specific context.

Stage 3: Post-Processing Refinement

Three refinement strategies were investigated to correct segmentation errors:

1. Deep Learning-based Refinement: A secondary UNet++ (ResNet50 encoder) trained to remove false positives (though results showed it sometimes eroded true vessels).
2. Contour Refinement: A heuristic method that identifies small contours below a size threshold, inverts them, and removes them via bitwise operations to eliminate noise.
3. Patch Line Generation: A technique to restore continuity in fragmented vessels. It uses morphological thinning to create skeletons, detects endpoints, and connects disjoint branches with "patch lines" if the pixel continuity along the connection path meets specific criteria.

3. Key Contributions

• Novel Architecture: Introduction of a Self-ONN-based decoder within a U-Net framework, which significantly improves the segmentation of thin, stenotic vessels compared to standard Conv2D decoders.
• Multichannel Preprocessing: A hybrid approach combining CLAHE and an improved Ben Graham method, which outperforms single-channel enhancement techniques.
• Comprehensive Refinement Pipeline: A systematic evaluation of post-processing techniques, demonstrating that contour refinement and patch line generation effectively reduce false positives and restore vessel continuity.
• Robust Validation: The model was trained and validated on a combined dataset of healthy (DCA1) and stenotic (Danilov et al.) arteries, ensuring the model learns from clinically challenging cases.

4. Experimental Results

The model was evaluated using 5-fold cross-validation on a combined dataset of 348 images (256x256 pixels).

• Performance Metrics:
  • Dice Similarity Coefficient (DSC): 76.10%
  • Intersection over Union (IoU): 61.43%
  • clDice (Topology-aware metric): 79.36%
  • Accuracy: 97.52%
• Comparative Analysis: CASR-Net outperformed several state-of-the-art models, including DenseNet121_UNet, DenseNet201_UNet++, ResNet50_MAnet, and MedSAM.
  • It achieved a higher DSC and IoU than the next best performing models (e.g., DenseNet121_Unet++_scSE).
  • It demonstrated superior performance in preserving vessel continuity (high clDice score).
• Ablation Studies:
  • Preprocessing: The multichannel approach yielded a DSC of 75.97%, significantly better than original images (75.12%) or single-method enhancements.
  • Architecture: Replacing standard layers with Self-ONN blocks improved DSC from 75.37% (UNet+DenseNet) to 75.74%.
  • Hyperparameters: A Self-ONN q-order of 3 provided the optimal balance between accuracy and complexity.
  • Refinement: Contour refinement was the most effective post-processing step, reducing False Negative Rate (FNR) while maintaining high DSC.

5. Significance and Clinical Implications

• Clinical Utility: By accurately segmenting stenotic and narrow branches, CASR-Net provides clinicians with more reliable tools for assessing the severity and location of blockages, which is critical for planning interventions like stenting or bypass surgery.
• Robustness: The ability to handle low-contrast images and preserve vessel topology makes the system suitable for real-world clinical environments where image quality varies.
• Future Directions: The authors acknowledge limitations such as the lack of external dataset validation and the use of heuristic refinement. Future work aims to incorporate temporal modeling (for video sequences), larger multi-center datasets, and learnable adaptive refinement modules to further enhance generalizability.

In conclusion, CASR-Net represents a significant advancement in automated coronary artery segmentation by integrating advanced image processing, a novel Self-ONN architecture, and targeted refinement strategies to overcome the specific challenges of X-ray angiography.