An Intra- and Cross-frame Topological Consistency Scheme for Semi-supervised Atherosclerotic Coronary Plaque Segmentation

Ziheng Zhang, Zihan Li, Dandan Shan, Yuehui Qiu, Qingqi Hong, Qingqiang Wu

Introduction

Accurate segmentation of atherosclerotic plaques in Computed Tomography Angiography (CTA) is a cornerstone of diagnosing Coronary Artery Disease (CAD). However, automated segmentation faces two primary hurdles: visual ambiguity, where plaque boundaries are indistinct from surrounding tissues, and the annotation bottleneck, caused by the intensive labor required for experts to label 3D vascular structures.

While semi-supervised learning (SSL) aims to mitigate the data shortage by utilizing unlabeled images, conventional SSL methods often struggle to maintain the specific "tube-like" geometry and spatial continuity inherent to human anatomy. This paper introduces a novel framework designed to enforce these geometric rules through topological consistency.

Methodology: The ICTC Framework

The proposed Intra- and Cross-frame Topological Consistency (ICTC) framework leverages the structural properties of vessels as anchors for learning. It operates on two primary levels of consistency:

1. Intra-frame Topological Consistency (ITC)

To capture the internal geometry of a single slice, the model employs a dual-task architecture. Beyond standard segmentation, a secondary branch predicts a Skeleton-aware Distance Transform (SDT). This transform maps the distance of every pixel to the vessel's center (the skeleton) and its edge (the boundary). By forcing the segmentation and distance transform tasks to align, the model develops a robust internal representation of the vessel's tubular shape, even when visual boundaries are unclear.

2. Cross-frame Topological Consistency (CTC)

Blood vessels are continuous structures across 3D space. To ensure this continuity, the framework utilizes an unsupervised Flow Estimator to analyze the "pixel flow" between adjacent frames. This mechanism prevents the model from making erratic predictions where vessel structures might otherwise appear or disappear inconsistently between sequential slices.

Dual-Consistency Loss

The training process is guided by a dual-consistency loss that minimizes discrepancies between these topological predictions. This approach effectively uses the geometric "rules" of anatomy to supervise the model's performance on unlabeled data.

Experimental Evaluation

The ICTC framework was rigorously tested on three datasets to evaluate performance and generalization:

  • CAA-Seg: A coronary artery atherosclerosis dataset (20 patients, 60 segments).
  • AD-CTA: A larger coronary dataset (40 patients, 98 segments).
  • ACDC: A public cardiac MRI dataset used to test the framework's versatility beyond coronary arteries.

Key Results

The ICTC method consistently outperformed established semi-supervised benchmarks such as CPS, MT, and URPC:

  • CAA-Seg: Achieved a Dice score of 74.58%, significantly higher than the previous best SSL method (CPS at 63.71%) and even surpassing the fully supervised nnUNet baseline (73.69%).
  • AD-CTA: Reached a Dice score of 73.34%, leading the semi-supervised field.
  • Generalization: On the ACDC dataset, the model achieved an 88.22% Dice score, demonstrating that topological consistency is a robust regularizer across different imaging modalities and organs.

Discussion and Insights

The core innovation of this research is the treatment of topology as a prior. By explicitly modeling the skeleton and boundary of tubular structures, the researchers provide the AI with a geometric framework that compensates for the lack of manual labels. Unlike many SSL methods that rely on adding noise (perturbation), this "task mapping" approach leverages the inherent anatomical logic of the human body.

Limitations and Future Directions

While effective, the framework introduces higher computational complexity due to the dual decoders and flow estimator. Additionally, the flow estimator may face challenges in cases of total vascular occlusion where the linear continuity of a vessel is disrupted. Future work may extend this topological approach to other complex anatomical structures, such as the bronchial tree or cerebral nerve fibers.

Summary updated

BibTeX
@inproceedings{zhang2025intra,
  author = {Ziheng Zhang and Zihan Li and Dandan Shan and Yuehui Qiu and Qingqi Hong and Qingqiang Wu},
  title = {{An Intra- and Cross-frame Topological Consistency Scheme for Semi-supervised Atherosclerotic Coronary Plaque Segmentation}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10890181},
}
Figure