Coronary Wall Segmentation in CCTA Scans Via a Hybrid Net with Contours Regularization

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

TL;DR

This paper introduces a novel deep learning approach for segmenting coronary artery walls in CCTA scans, addressing the common issue of disconnected and inaccurate boundaries. By employing a hybrid 3D-2D network and a new contour-regularized loss function, the method achieves more accurate, closed, and smooth boundary detections, crucial for diagnosing coronary artery disease.

What Problem Does This Paper Address?

Coronary Artery Disease (CAD) is a major global health concern, and accurate diagnosis relies heavily on precise visualization of coronary arteries. Coronary CT angiography (CCTA) provides detailed 3D images, but segmenting the artery walls, particularly the inner and outer boundaries, is a challenging task for automated systems. Existing deep learning methods often struggle with the inherent noise, partial volume effects, and the need for continuous, closed boundaries, leading to fragmented or imprecise segmentations. This necessitates manual correction by cardiologists, increasing workload and potential for error. This paper aims to bridge this gap by developing a method that produces robust, well-connected coronary artery wall boundaries.

Methods

The authors propose a two-pronged approach:

  1. Hybrid 3D-2D Network Architecture: To leverage the spatial context of adjacent slices, the network takes a volume of consecutive CCTA slices (specifically, a segment of the coronary artery) as input. A 3D encoder processes this volume to capture inter-slice dependencies, while a 2D decoder then generates the boundary prediction for the central slice. This hybrid design aims to balance the benefits of 3D context with reduced computational complexity compared to fully 3D networks. Skip connections are used to propagate features from the encoder to the decoder.
  2. Contour-Regularized Weighted Hausdorff Distance (WHD) Loss: A key innovation is a novel loss function designed to enforce closed and connected boundaries. Building upon the Weighted Hausdorff Distance (WHD) loss, which measures boundary differences, this new loss incorporates a "contour regularization" term. This term essentially guides the network's output towards forming a closed contour. The final loss is a combination of the WHD loss between the predicted and ground truth boundaries, and the WHD loss between the predicted boundary and a derived closed contour. This encourages the network to produce boundaries that are not only accurate but also topologically sound.

Datasets

The method was evaluated on a non-public CCTA-CPR dataset comprising scans from 34 patients. The dataset was split into training (25 patients, 24,727 slices), validation (4 patients, 3,231 slices), and testing (5 patients, 5,718 slices). The data was pre-processed and augmented, including Hounsfield Unit mapping, random cropping, resizing to 96x96, and random rotations/flips.

Key Insights and Ideas

  • Leveraging 3D Context for 2D Tasks: The hybrid 3D encoder and 2D decoder architecture effectively utilizes information from neighboring slices to improve the segmentation of the target slice, a crucial step for handling the continuity of blood vessels.
  • The Importance of Boundary Connectivity: The paper highlights that simply achieving high pixel-wise accuracy isn't enough for medical image segmentation. For coronary arteries, the continuity and closure of boundaries are paramount for accurate diagnosis, and this is where traditional methods often fail.
  • Contour Regularization as a Powerful Constraint: The proposed contour-regularized WHD loss is a clever way to explicitly enforce topological constraints (closed boundaries) on the network's output, directly addressing the fragmentation issue.
  • Separate Handling of Inner and Outer Boundaries: Recognizing that inner and outer boundaries have different characteristics (e.g., outer boundaries are harder to detect due to partial volume effects), the loss function is applied separately to each, with different weighting ratios, further improving accuracy.

Key Results

  • Quantitative Performance: The proposed method achieved state-of-the-art accuracy, outperforming several baseline methods in terms of Average Boundary Distance (ABD) and 95th percentile Hausdorff Distance (HD95). It was also comparable to the state-of-the-art BESNet in Dice Similarity Coefficient (DSC).
  • Qualitative Improvements: Visual comparisons demonstrated that the proposed method produces significantly smoother and more closed boundaries compared to methods like morphological Snakes, segmentation-based Res-UNet, and even a hybrid network with a standard Cross Entropy loss. The contour-regularized WHD loss successfully generated closed, one-pixel thick boundaries.

Limitations and Future Work

  • Computational Cost: While the hybrid 3D-2D approach is less computationally intensive than fully 3D methods, it still involves 3D convolutions, which can be demanding.
  • Dataset Specificity: The evaluation was performed on a specific CCTA-CPR dataset. Generalizability to other imaging modalities or datasets might require further validation.
  • Contour Derivation: The paper mentions using morphological Snakes or border following to derive the contour for regularization. The robustness and computational cost of these post-processing steps could be a factor.

Future work could explore further optimization of the hybrid architecture, investigate more advanced contour regularization techniques, and test the method on larger and more diverse datasets.

Summary updated

Abstract

Providing closed and well-connected boundaries of coronary artery is essential to assist cardiologists in the diagnosis of coronary artery disease (CAD). Recently, several deep learning-based methods have been proposed for boundary detection and segmentation in a medical image. However, when applied to coronary wall detection, they tend to produce disconnected and inaccurate boundaries. In this paper, we propose a novel boundary detection method for coronary arteries that focuses on the continuity and connectivity of the boundaries. In order to model the spatial continuity of consecutive images, our hybrid architecture takes a volume (i.e., a segment of the coronary artery) as input and detects the boundary of the target slice (i.e., the central slice of the segment). Then, to ensure closed boundaries, we propose a contour-constrained weighted Hausdorff distance loss. We evaluate our method on a dataset of 34 patients of coronary CT angiography scans with curved planar reconstruction (CCTA-CPR) of the arteries (i.e., cross-sections). Experiment results show that our method can produce smooth closed boundaries outperforming the state-of-the-art accuracy.
BibTeX
@inproceedings{huang2020coronary,
  author = {Kaikai Huang and 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 = {{Coronary Wall Segmentation in CCTA Scans Via a Hybrid Net with Contours Regularization}},
  booktitle = {ISBI},
  year = {2020},
  url = {https://openreview.net/forum?id=oWXtFTn5R2pE},
}

Loading related content...

Figure