Quantifying Anatomical Bias in Coronary Segmentation: Why Your Model Prefers the LCA More Than the RCA
Baumgart S, Deubner N, Kist A (2026)
Publication Type: Conference contribution
Publication year: 2026
Journal
Publisher: Springer Science and Business Media Deutschland GmbH
Pages Range: 70-75
Conference Proceedings Title: Informatik aktuell
Event location: Lübeck
ISBN: 9783658510992
DOI: 10.1007/978-3-658-51100-5_12
Abstract
Deep learning-based segmentation of coronary arteries in X-ray angiography supports stenosis assessment via the Quantitative Flow Ratio. However, traditional metrics like the Dice coefficient neglect howimage acquisition parameters, particularly vessel type and projection angle, affect model accuracy. This study evaluated four U-Net-based models (vanilla U-Net, nnUNet, and U-Nets using MobileNetV2 or InceptionResNetV2 as encoders) on 599 patients covering twelve common projection angles. Results show that projection angles, including vessel overlap, have a stronger impact on segmentation quality than vessel type. InceptionResNetV2 achieved the highest overall Dice scores, while nnU-Net better captured capillaries and catheters. Distal branches remained challenging for all models. Our findings highlight the need to consider projection-angle diversity and segment-level evaluation in datasets and benchmarks to ensure clinically reliable coronary segmentation.
Authors with CRIS profile
Selina Baumgart
Juniorprofessur für Artificial Intelligence in Communication Disorders
Andreas Kist
Juniorprofessur für Artificial Intelligence in Communication Disorders
How to cite
APA:
Baumgart, S., Deubner, N., & Kist, A. (2026). Quantifying Anatomical Bias in Coronary Segmentation: Why Your Model Prefers the LCA More Than the RCA. In Heinz Handels, Katharina Breininger, Thomas Deserno, Andreas Maier, Klaus Maier-Hein, Christoph Palm, Thomas Tolxdorff (Eds.), Informatik aktuell (pp. 70-75). Lübeck, DE: Springer Science and Business Media Deutschland GmbH.
MLA:
Baumgart, Selina, Nikolas Deubner, and Andreas Kist. "Quantifying Anatomical Bias in Coronary Segmentation: Why Your Model Prefers the LCA More Than the RCA." Proceedings of the Bildverarbeitung für die Medizin Workshop, BVM 2026, Lübeck Ed. Heinz Handels, Katharina Breininger, Thomas Deserno, Andreas Maier, Klaus Maier-Hein, Christoph Palm, Thomas Tolxdorff, Springer Science and Business Media Deutschland GmbH, 2026. 70-75.
BibTeX: Download