A hierarchy-aware multi-label framework for anatomically consistent intravascular ultrasound segmentation
Intravascular ultrasound (IVUS) segmentation is commonly formulated as a single-label multi-class task, where each pixel is assigned to one class. This setting does not match the nested anatomy of coronary arteries, because the lumen is enclosed by the media and plaque components are located in the annular region between them. We propose a hierarchy-aware multi-label framework for IVUS vessel and plaque segmentation. The dataset included 2079 IVUS frames from 113 patients. Seven foreground categories were annotated, six of which were used for primary evaluation. Side branch was retained as an auxiliary category and excluded from mDice and mIoU. All experiments used patient-level five-fold cross-validation. In controlled DeepLabV3+ ablation experiments, the single-label cross-entropy plus Dice baseline achieved a mean Dice of 0.5893 ± 0.0019 and a raw anatomical violation rate of 7.09 % ± 0.52 % . Multi-label variants increased the mean Dice to 0.7257–0.7372 and reduced the raw violation rate to 0.11%–0.19%. Among 16 architectures evaluated under the same multi-label setting, Attention U-Net achieved the highest mean Dice of 0.7381 ± 0.0059. The proposed label formulation improved segmentation accuracy and anatomical consistency in this internal validation, but its generalizability requires external validation.
Authors
- Xin Ning (ORCID: https://orcid.org/0000-0001-7897-1673)
- Meilan Hao (ORCID: https://orcid.org/0000-0002-5265-4992)
- Hong Wang
- Yujun Zhao
- Ruilan Hao
- Kai Zhou
Publication Details
- Journal
- Engineering Applications of Artificial Intelligence
- Published
- 2026-09-14
- DOI
- https://doi.org/10.1016/j.engappai.2026.116234
- Primary Topic
- Coronary Interventions and Diagnostics
- Type
- article
- Field-Weighted Citation Impact
- 0.00