Abstract
Automated detection of coronary artery plaques in Coronary Computed Tomography Angiography (CCTA) scans is a critical but challenging task in cardiac disease diagnosis. Non-calcified plaques, in particular, are difficult to segment due to their visual similarity to surrounding fat and background tissues. We present a novel two-step pipeline that utilizes artery wall masking to constrain the search space for plaque detection. By isolating the artery wall before segmentation, our deep learning model significantly improves accuracy for high-risk non-calcified plaques, achieving a Dice score improvement from 0.49 to 0.58 with automated masking, and up to 0.73 with ground-truth masks.
Introduction
Coronary artery disease remains a leading cause of global mortality. While CCTA allows for non-invasive plaque detection, the process is often hindered by the "neighborhood effect." Calcified plaques are easily identified due to their high intensity, but non-calcified plaques—which are often unstable and prone to rupture—share similar intensity values with the surrounding fat. Current deep learning models frequently struggle with these ambiguous boundaries, leading to inaccurate segmentations that are unsuitable for clinical use. This research addresses the problem by introducing a "masking" technique that filters out noise from surrounding tissues.
Methodology
The proposed method employs a two-stage pipeline to refine the segmentation process:
- Artery Wall Masking: The system first identifies the inner and outer surfaces of the artery wall using a HybridResUNet architecture. This step generates a binary mask that isolates the space between these surfaces, effectively removing the "distraction" of surrounding fat and neighboring vessels.
- Plaque Segmentation: The generated mask is applied to Curved Planar Reformation (CPR) scans. The model then performs a 4-class segmentation (background, artery wall, calcified plaque, and non-calcified plaque) exclusively within the masked region.
To address the class imbalance—where non-calcified plaques are less frequent but more clinically significant—the researchers utilized a weighted Dice loss function. This forces the model to prioritize the accurate detection of these difficult-to-see features.
Results and Key Insights
The study was validated using a dataset of 100 CCTA scans (291 individual arteries) and over 100,000 image slices, with manual annotations from three expert radiologists serving as the ground truth.
The quantitative results demonstrate that constraining the model's focus leads to substantial performance gains. As shown in the table below, the accuracy for non-calcified plaques jumped from a baseline of 0.49 (no mask) to 0.58 with an automated mask, and reached 0.73 when using a perfect ground-truth mask.
Key Findings:
- The Power of Constraints: Intentionally discarding background data (noise) is a primary driver of accuracy. Even when the masks were intentionally noisy or imperfect, the system consistently outperformed the "no-mask" baseline.
- Anatomical Realism: Qualitative analysis showed that masking prevents plaques from "protruding" into surrounding fat, a common error in unconstrained models.
- Surpassing Human Consistency: Radiologists noted that the model occasionally identified plaques missed in the original manual annotations, suggesting a potential for higher consistency than human experts in complex 3D volumes.
Conclusion and Future Work
By leveraging the anatomical structure of the artery wall as a constraint, this masking technique provides a robust framework for improving plaque detection. While the current system relies on a two-step process, future work will explore end-to-end training. A unified neural network that simultaneously learns to detect the wall and segment plaques could allow these tasks to inform one another, further reducing false positives and improving clinical reliability.
Summary updated