Impact of Connectivity-Preserving Loss Functions on the Segmentation of Thin Tubular Structures: Application to Coronary Arteries From CT Angiography Data

Coronary artery disease remains a leading cause of mortality worldwide. Coronary Computed
Tomography Angiography (CCTA) provides a non-invasive basis for diagnosis; however, an
accurate and connectivity-preserving segmentation of the coronary artery tree is essential for
robust automatic and quantitative analyses. Convolutional Neural Networks (CNNs)-based
architectures, in particular U-Net and no-new-U-Net (nnU-Net), have shown outstanding
performance across a wide range of medical image segmentation benchmarks, yet they may
frequently produce fragmented vessel trees when segmenting thin, tubular structures such as
coronary arteries. Recent studies indicate that connectivity-aware loss functions can mitigate
these discontinuities by explicitly penalizing missing centerline segments, but their efficacy
for coronary artery tree segmentation remains to be demonstrated.
This thesis quantifies the benefits and challenges of integrating connectivity-preserving loss
functions into an nnU-Net-based pipeline for one-step coronary artery tree segmentation from
CCTA images. Performance is assessed using complementary metrics covering vessel mask
... mehr
accuracy, vessel accuracy, centerline completeness, and runtime, capturing volumetric overlap,
connectivity-related effects, and computational cost. To contextualize the quantitative
results, we conduct qualitative case studies with targeted visualizations.
The comparison of connectivity-preserving losses reveals that the Skeleton Recall (SR)
loss provides the most consistent improvements in connectivity metrics while incurring
substantially lower training time than the other connectivity-preserving loss formulations. In
the final statistical analysis, augmenting the generic loss with a SR term improves coronary
artery tree connectivity in a statistically significant and practically relevant manner, without
substantially degrading volumetric overlap and with negligible computational overhead.
These findings identify SR as an effective, computationally efficient, and straightforward-to-implement loss function, making it a practically viable choice for accurate and connectivitypreserving coronary artery tree segmentation.