Introduction
Coronary Artery Disease (CAD) remains a primary cause of global mortality. While deep learning has advanced the analysis of Coronary CT Angiography (CCTA), clinical adoption is hindered by the "black-box" nature of most models and their tendency toward overconfidence. Presented at AAAI-2025, the Trusted Lesion-assessment Network (TLA-Net) addresses these issues by integrating causal reasoning and a hierarchical uncertainty model that aligns with clinical workflows.
The Challenge of Trusted CAD Diagnosis
Automated CAD diagnosis faces two significant hurdles:
- Lack of Interpretability: Models often rely on spurious correlations, such as image noise or scanner-specific artifacts, rather than biological markers.
- Unreliable Confidence: Standard models frequently provide high confidence scores even for incorrect predictions, making it difficult for clinicians to identify when a manual second opinion is necessary.
The TLA-Net Framework
TLA-Net is designed to be a "trusted" architecture by mimicking the diagnostic logic of radiologists. The framework processes multi-view imaging signals through two primary modules: Causality-informed Evidence Collection (CEC) and Clinically-aligned Uncertainty Integration (CUI).
Causality-informed Evidence Collection (CEC)
The CEC module moves beyond simple pattern matching by employing a causal graph to model the diagnostic process. By implementing "causal interventions" ($P(Y|do(X))$), the network uses an intervention medium to aggregate multi-view representations. This process effectively "blocks" the influence of confounders—such as poor image quality or blurred boundaries—that typically disrupt a model's logic. This ensures the model focuses on features that are invariant across different scanners and image qualities.
Clinically-aligned Uncertainty Integration (CUI)
The CUI module replicates the "global-to-local" workflow used by physicians, who typically start with a global overview of the artery before examining specific segments.
As shown in the detailed architecture, the module uses Dirichlet distributions to quantify uncertainty for both global and local views. These uncertainties are then merged using Dempster’s combination rule, ensuring that the final confidence score is mathematically grounded in the consistency of evidence across different perspectives.
Experimental Evaluation
The researchers evaluated TLA-Net using a robust dataset of 2,618 lesions from two clinical centers (CC1 and CC2), utilizing different CT scanner technologies. The study employed two data splits:
- DS1 (Proportional Split): To test general performance.
- DS2 (Center-based Split): To evaluate domain adaptability across different hospital environments.
Key Results
TLA-Net demonstrated state-of-the-art (SOTA) performance across all major metrics:
- Accuracy: Achieved 0.953 on DS1 and 0.944 on DS2, showing remarkable stability.
- F1-Score: Reached 0.962, indicating a strong balance between precision and recall.
- Calibration: The model achieved a significantly lower Expected Calibration Error (ECE) of 0.159 compared to the 0.217 of standard Softmax models. This improvement means the model's confidence scores are more representative of its actual accuracy.
Conclusion
TLA-Net represents a significant step toward trustworthy AI in medical imaging. By prioritizing causality over simple correlation and aligning uncertainty estimation with clinical practice, it provides a more reliable and interpretable tool for CAD diagnosis. Future work will explore incorporating additional clinical modalities, such as patient history, to further enhance diagnostic accuracy and disease monitoring.
Summary updated