Beam Stack Search-Based Reconstruction Of Unhealthy Coronary Artery Wall Segmentations In CCTA-CPR Scans

Antonio Tejero-de-Pablos, Hiroaki Yamane, Yusuke Kurose, Junichi Iho, Youji Tokunaga, Makoto Horie, Keisuke Nishizawa, Yusaku Hayashi, Yasushi Koyama, Tatsuya Harada

TL;DR

Researchers have developed a "masking" technique that significantly improves the automated detection of arterial plaques in heart scans. By first identifying the artery wall and filtering out surrounding tissues (like fat and neighboring vessels), their deep learning model can segment difficult-to-see non-calcified plaques with much higher accuracy, potentially offering a more reliable tool for early heart disease diagnosis.

What Problem Does This Paper Address?

Coronary artery disease is a leading cause of death globally, often driven by the buildup of plaques. Doctors use Coronary Computed Tomography Angiography (CCTA) to find these plaques non-invasively. While "calcified" plaques (hardened by calcium) show up clearly on scans, "non-calcified" plaques are the real danger—they are unstable, prone to rupture, and notoriously difficult to detect.

The challenge lies in the "neighborhood." Non-calcified plaques have intensity values very similar to the fat and background tissues surrounding the artery. Current deep learning models often get "confused" by these ambiguous intensities at the edges of the artery wall, leading to missed detections or "rough" segmentations that aren't useful for detailed clinical analysis.

Methods

The authors propose a two-step pipeline to solve this "needle in a haystack" problem:

  1. Artery Wall Masking: Instead of looking at the whole scan, the system first identifies the inner and outer surfaces of the artery wall (using a method called HybridResUNet). It creates a binary mask that includes only the space between these two surfaces.
  2. Plaque Segmentation: This mask is applied to the input volume (specifically Curved Planar Reformation or CPR scans, which "unroll" the artery into a straight line). The model then performs a 4-class segmentation—background, artery wall, calcified plaque, and non-calcified plaque—only within that restricted area.

By using a weighted Dice loss function, the researchers forced the model to pay extra attention to non-calcified plaques, which are less common in the data but more clinically significant.

Datasets

The study utilized a non-public dataset consisting of:

  • 100 CCTA scans, resulting in 291 individual coronary arteries (LAD, LCX, and RCA).
  • 102,983 individual image slices.
  • Expert Ground Truth: Three expert radiologists provided the manual annotations used for training and validation.

Key Insights and Ideas

  • The Power of Constraints: The most important finding is that less is more. By intentionally throwing away data (the background voxels), the model's performance on the remaining data skyrocketed. This suggests that "noise" from surrounding fat is a primary bottleneck for current plaque-detection AI.
  • Robustness to "Messy" Masks: Even when the researchers intentionally added noise to the masks (simulating a less-than-perfect wall detection algorithm), the system still outperformed the "no-mask" approach. This indicates that even a "fuzzy" idea of where the artery wall is provides a massive boost to segmentation accuracy.
  • Clinical Discovery: Interestingly, the radiologists noted that the model was occasionally able to find plaques that were missed in the original manual annotations, suggesting the AI might eventually surpass human consistency in these complex 3D volumes.

Key Results

The researchers compared five configurations, ranging from a "perfect" ground-truth mask to no mask at all. The results (measured by Dice score, where 1.0 is a perfect match) were telling:

  • Non-Calcified Plaques: Accuracy jumped from 0.49 (no mask) to 0.73 (perfect mask). Even with an automated wall-segmentation mask, the score was 0.58, a significant improvement over the baseline.
  • Calcified Plaques: Accuracy improved from 0.73 (no mask) to 0.83 (perfect mask).
  • Qualitative Improvement: Visual results showed that without masking, plaques often "protruded" out of the artery wall into the surrounding fat. Masking kept the segmentations anatomically realistic.

Limitations and Future Work

The primary limitation is the dependency on the initial wall detection. If the mask is significantly off-center or misses a section of the artery, the plaque segmentation will inevitably fail in those areas. For instance, the authors noted that inaccuracies in the mask could still lead to false positives if fat tissue is accidentally included.

Future Work: The authors suggest moving toward an end-to-end training approach. Instead of two separate steps, they envision a single neural network that learns to detect the wall and the plaques simultaneously, allowing the two tasks to inform and improve each other in real-time.

Summary updated

Abstract

The estimation of the coronary artery wall boundaries in CCTA scans is a costly but essential task in the diagnosis of cardiac diseases. To automatize this task, deep learning-based image segmentation methods are commonly used. However, in the case of coronary artery wall, even state-of-the-art segmentation methods fail to produce an accurate boundary in the presence of plaques and bifurcations. Post-processing reconstruction methods have been proposed to further refine segmentation results, but when applying general-purpose reconstruction to artery wall segmentations, they fail to reproduce the wide variety of boundary shapes. In this paper, we propose a novel method for reconstructing coronary artery wall segmentations, the Tube Beam Stack Search (TBSS). By leveraging the voxel shape of adjacent slices in a CPR volume, our TBSS is capable of finding the most plausible path of the artery wall. Similarly to the original Beam Stack Search, TBSS navigates along the voxel probabilities output by the segmentation method, reconstructing the inner and outer artery walls. Finally, skeletonization is applied on the TBSS reconstructions to eliminate noise and produce more refined segmentations. Also, since our method does not require learning a model, the lack of annotated data is not a limitation. We evaluated our method on a dataset of coronary CT angiography with curved planar reconstruction (CCTA-CPR) of 92 arteries. Experimental results show that our method outperforms the state-of-the-art work in reconstruction.
BibTeX
@inproceedings{tejerodepablos2021beam,
  author = {Antonio Tejero-de-Pablos and Hiroaki Yamane and Yusuke Kurose and Junichi Iho and Youji Tokunaga and Makoto Horie and Keisuke Nishizawa and Yusaku Hayashi and Yasushi Koyama and Tatsuya Harada},
  title = {{Beam Stack Search-Based Reconstruction Of Unhealthy Coronary Artery Wall Segmentations In CCTA-CPR Scans}},
  booktitle = {ISBI},
  year = {2021},
  url = {https://openreview.net/forum?id=XA1kjaSTrdt2},
}

Loading related content...

Figure