A Framework for Cross-Domain Generalization in Coronary Artery Calcium Scoring Across Gated and Non-Gated Computed Tomography

The paper presents CARD-ViT, a self-supervised Vision Transformer framework trained exclusively on ECG-gated CT data that successfully enables automated Coronary Artery Calcium scoring on non-gated scans, thereby facilitating scalable cardiovascular risk assessment using routine chest imaging without requiring additional scans or annotations.

The Gist

Imagine your heart's arteries are like old, rusty pipes. Over time, hard, rocky deposits (calcium) build up inside them. Doctors call this Coronary Artery Calcium (CAC). The more rock you have, the higher your risk of a heart attack.

To measure this, doctors usually need a special, high-speed camera called an ECG-gated CT scan. This camera is like a photographer who only snaps pictures when the heart is perfectly still (between beats), ensuring the image isn't blurry. This is great for accuracy, but it's expensive, rare, and only done in specialized heart centers.

However, millions of people get non-gated CT scans every year for other reasons, like checking their lungs for pneumonia or cancer. These scans are like taking a photo of a moving car; the heart might be a little blurry. Because of this blur, doctors usually ignore these scans for heart health, missing a huge opportunity to catch heart disease early.

The Problem:
We have a mountain of "blurry" heart scans (non-gated) that we aren't using, and a shortage of "sharp" heart scans (gated) to train AI to read them. Most AI models need to be trained on the exact type of data they will eventually see. If you train an AI on sharp photos, it usually fails on blurry ones.

The Solution: The "Smart Eye" (CARD-ViT)
The researchers in this paper built a new AI system called CARD-ViT. Think of this system as a super-smart student who only ever studied from perfect, high-definition textbooks (the sharp, gated scans).

Here is the magic trick:

1. Self-Taught Learning: Instead of being told "this is calcium, this is not," the AI was shown thousands of sharp heart scans and asked to find patterns on its own. It learned what a heart looks like and what calcium feels like without needing a teacher to point at every single spot.
2. The Leap: The researchers then asked: "Can this student, who only studied perfect textbooks, still read the blurry, messy notes (non-gated scans)?"
3. The Result: Surprisingly, yes! The AI learned the fundamental "shape" of calcium so well that it could recognize it even when the image was blurry. It didn't need to be retrained on the blurry scans.

How It Works (The Analogy):
Imagine you are learning to recognize a specific type of bird, the "Cardinal."

• Old Way: You study photos of Cardinals taken in perfect sunlight. Then, you try to spot them in foggy weather. You might get confused.
• This Paper's Way: You study the Cardinal so deeply in perfect sunlight that you understand its essence—its red color, its shape, its size. When you go into the fog, your brain fills in the gaps. You still know, "That's a Cardinal," even if you can't see the feathers clearly.

The Results:

• On Sharp Scans: The AI was a master, getting it right 91% of the time.
• On Blurry Scans: It performed just as well as other AI models that had been specifically trained on blurry scans (about 70% accuracy).
• The Big Win: This means we can now take the millions of routine chest scans people get for other reasons and instantly check their heart health without giving them any extra radiation or extra scans.

Why This Matters:
Currently, checking your heart calcium takes time, money, and a special appointment. This new framework is like adding a "heart health check" button to the software that already reads your lung scans. It turns a missed opportunity into a life-saving tool.

The Catch:
The AI isn't perfect yet. It's really good at saying "No calcium" or "A lot of calcium," but it sometimes struggles with the "medium" amounts. Think of it like a weather forecaster who is great at predicting "Sunny" or "Storm," but sometimes gets confused between "Cloudy" and "Overcast."

The Bottom Line:
The researchers proved that you don't need a million blurry photos to teach an AI to see through the fog. If you teach it well on clear photos, it can learn to see the truth even when the picture isn't perfect. This could revolutionize how we screen for heart disease, making it faster, cheaper, and available to almost everyone.

Technical Summary

1. Problem Statement

Coronary Artery Calcium (CAC) scoring is a critical, non-invasive predictor of cardiovascular risk, traditionally calculated using ECG-gated CT scans. However, gated scans are specialized, expensive, and less frequent. Conversely, non-gated CT scans (e.g., for pulmonary evaluation) are performed roughly 10 times more frequently, representing a massive "missed opportunity" for opportunistic cardiovascular screening.

Current challenges include:

• Domain Shift: Significant differences in image quality, motion artifacts, and acquisition protocols between gated and non-gated scans.
• Data Scarcity: High-quality, pixel-level annotated datasets for non-gated scans are scarce, making it difficult to train models specifically for them.
• Existing Limitations: Current domain adaptation methods often require access to both gated and non-gated data during training, which is impractical given the lack of labeled non-gated data.

2. Methodology

The authors propose an end-to-end framework that achieves cross-domain generalization by training exclusively on high-quality gated data, without any non-gated supervision. The pipeline consists of three main stages:

A. Feature Extraction: CARD-ViT

• Architecture: A Vision Transformer (ViT) backbone named CARD-ViT.
• Pretraining Strategy: The model is pre-trained using DINO (Self-Distillation with NO labels), a self-supervised learning approach. This allows the model to learn robust anatomical representations from large amounts of unlabeled CT slices.
• Input Adaptation: The standard RGB-based patch embedding layer of ViT is modified to accept single-channel grayscale CT inputs by averaging pretrained weights across channels.
• Token Processing: Input slices ($512 \times 512$) are divided into $16 \times 16$ patches, creating 1,024 tokens. The model utilizes a [CLS] token and four register tokens (auxiliary learnable embeddings) to refine spatial attention patterns.
• Feature Fusion: The model extracts three feature types from the final layer: patch embeddings, attention maps, and register token features. These are concatenated to form a unified 784-channel representation.

B. Segmentation and Scoring

• Decoder: A custom decoder processes the 784-channel features. It reduces dimensionality and upsamples through four stages ($32 \to 512$ resolution) using bilinear interpolation and convolution.
• Skip Connections: Selective skip connections from the original input CT slices are integrated at $256 \times 256$ and $512 \times 512$ resolutions to restore fine anatomical details lost during decoding.
• Output: The network outputs a single-channel binary mask highlighting all calcified regions (artery-agnostic), rather than separate masks for specific arteries (LM, LAD, CX, RCA). This design improves robustness and mitigates class imbalance.
• Scoring Algorithm: A lesion-specific adaptation of the Agatston scoring method is applied. Connected-component analysis identifies individual lesions, combining physical area ($mm^2$) and maximum Hounsfield Unit (HU) intensity to calculate a weighted score.

C. Clinical Deployment

The framework is deployed via MONAI Label integrated with the OHIF Viewer, allowing radiologists to review segmentation overlays, interactively refine results, and validate predictions in real-time.

3. Key Contributions

1. Cross-Domain Generalization without Non-Gated Data: The primary contribution is demonstrating that a model trained exclusively on gated CT data can generalize effectively to non-gated CT scans. This bypasses the need for scarce, labeled non-gated datasets.
2. CARD-ViT Architecture: Introduction of a self-supervised ViT backbone specifically adapted for medical CT imaging, utilizing DINO pretraining and register tokens to capture complex spatial relationships and calcification patterns.
3. Lesion-Specific Agatston Scoring: Implementation of a scoring pipeline that moves beyond slice-wise classification to pixel-level segmentation and lesion-specific scoring, aligning with clinical standards and improving risk stratification.
4. Clinical Integration: Successful deployment of the pipeline in a clinical environment (OHIF/MONAI Label), facilitating real-world validation and interaction.

4. Results

Datasets

• Training: Heartlens dataset (University of Kentucky), containing 2,651 gated cases.
• Internal Testing: Heartlens test set (469 gated cases).
• External Testing: Stanford Coronary Calcium dataset, comprising:
  • 443 gated cases (with pixel-level annotations).
  • 205 non-gated cases (with patient-level scores, no pixel-level labels).

Performance Metrics

• Gated CT Performance:
  • Accuracy: 0.910 on both Heartlens and Stanford gated test sets.
  • Cohen's $\kappa$: 0.871 (Heartlens) and 0.874 (Stanford), indicating strong agreement with ground truth.
  • Segmentation: Dice coefficient of 0.865 (Heartlens) and 0.839 (Stanford).
• Non-Gated CT Generalization (Cross-Domain):
  • CARD-ViT (Gated-trained): Achieved 0.707 accuracy and $\kappa$ of 0.528 on the Stanford non-gated dataset.
  • Comparison: This performance matches the AI-CAC model (0.707 accuracy, $\kappa$ 0.542), which was trained directly on non-gated data.
  • Implication: The model achieves comparable results to a domain-specific model without ever seeing non-gated training data.
• Ablation Study: The Base size DINOv2 backbone yielded the best performance (Dice 0.865). Larger models (Large, Giant) suffered from overfitting and reduced generalization on medical CT data.

Qualitative Findings

• The model successfully distinguished true coronary calcifications from confounders like aortic root calcification, mitral annular calcification, and pacemaker leads.
• Errors in non-gated scans were concentrated in intermediate risk categories (11–400), often due to motion artifacts affecting peak HU or lesion area near category thresholds.

5. Significance and Conclusion

This paper establishes the feasibility of opportunistic cardiovascular screening using routine non-gated chest CTs. By leveraging self-supervised learning on gated data, the framework overcomes the bottleneck of data scarcity for non-gated annotations.

• Clinical Impact: The framework enables scalable risk stratification without requiring additional imaging or radiation exposure.
• Deployment Strategy: Given the lower performance on non-gated scans compared to gated scans, the authors suggest the system is best used as a triage tool with clinician over-read, rather than fully automated decision-making for intermediate risk cases.
• Future Work: The authors plan to explore 3D extensions, domain adaptation techniques to further improve non-gated performance, and multi-center validation to address scanner-specific variations.

In summary, CARD-ViT demonstrates that robust, transferable anatomical features learned from high-quality gated data can effectively bridge the domain gap to non-gated scans, offering a practical solution for widespread cardiovascular risk assessment.