A progressive teacher–student framework for semi-supervised coronary artery segmentation in X-ray angiography via robust signal refinement
Journal
Biomedical Signal Processing and Control
Series/Report No.
Biomedical Signal Processing and Control
Journal Volume
126
Pages
110892
ISSN
1746-8094
Date Issued
2026-10-15
Author(s)
Hsu, Jung-Cheng
Chan, Chien-Lung
Lyu, Yun-Huan
Chang, Wei-Hao Martin
Chen, Shu-Lu
Chuang, Wen-Po
Tu, Chung-Ming
Wu, Yen-Wen
Abstract
Background and Objective: X-ray coronary angiography provides high-resolution vessel signals but is degraded by quantum noise, catheter artifacts, and Poisson-distributed scatter, making manual annotation costly. We propose a noise-resilient teacher-student framework with a dual-criteria pseudo-label filter combining structural fidelity (Dice ≥ 0.87) and anatomical signal-density prior (vessel-pixel count > 1480) for semi-supervised learning under label scarcity. Methods: An U-Net with EfficientNet backbone was embedded in a generational SSL pipeline. Two teacher models (EfficientNet-B6, ∼43 M parameters) were trained on 339 labeled frames from 103 patients; unlabeled sequences (3060 frames) were processed for pseudo-labels meeting filter criteria. Three student generations (EfficientNet-B7, ∼66 M parameters) were trained on progressively cleaner labels. Evaluation used Dice Similarity Coefficient (DSC) with 95 % confidence intervals (CIs), Intersection over Union (IoU), 95th-percentile Hausdorff Distance (HD95), and SNR. Comparisons were made against SSL paradigms (FixMatch, Mean-Teacher, Noisy-Student). Results: The final model achieved Dice similarity coefficients of 0.915 for the right coronary artery (RCA), 0.851 for the left circumflex artery (LCX), and 0.773 for the left anterior descending artery (LAD), with a mean DSC of 0.847. After Bonferroni correction, statistically supported improvement was observed mainly from the teacher model to Student 2 and Student 3, whereas later-generation gains were smaller and consistent with diminishing returns. Additional evaluation using Intersection over Union (IoU) and 95th-percentile Hausdorff Distance (HD95) further supported improved segmentation accuracy and boundary precision. The framework substantially reduced manual annotation requirements by approximately 80–90 % while maintaining strong RCA segmentation performance. Conclusion: This study demonstrates the feasibility of a semi-supervised learning approach for coronary artery segmentation in X-ray angiography. The proposed framework provides a practical foundation for data-efficient AI-assisted QCA development and represents preliminary single-center proof-of-concept evidence rather than definitive evidence of clinical robustness. Vessel-specific applicability was strongest for RCA segmentation, whereas LAD and LCX performance still require further methodological improvement. External multicenter validation will be necessary to establish generalizability and clinical reliability before clinical use can be considered. © 2026
Publisher
Elsevier BV
Type
journal article