Uncertainty-Aware Meta-Learning for Robust Stent Segmentation via Selective Ensemble

Hyojin Ko1, Jihoon Kweon2, Jung-Min Ahn2, Young-Hak Kim2
1Asan Medical Center, University of Ulsan College of Medicine, 2Asan Medical Center


Abstract

Assessing stent expansion on fluoroscopic images during Percutaneous Coronary Intervention (PCI) requires coronary stent segmentation, but achieving consistent performance with a single model is challenging due to low stent visibility and morphological variability. We propose a selective ensemble of five EfficientNet-B7/Feature Pyramid Network (FPN) base models, each trained with BCEDice (binary cross-entropy + Dice) plus one auxiliary loss (clDice, Hausdorff distance, perimeter regularity, TopK, or Wasserstein topological homology) and search-optimized weights. A Convolutional Block Attention Module (CBAM) meta-learner integrates the fluoroscopic image, base model probability maps, and pixel-wise uncertainty through two-phase label-free training (random-fit and nearest-neighbor-fit). Internal five-fold cross-validation (942 frames, 826 patients) achieved a Dice Similarity Coefficient (DSC) of 90.43%, exceeding the best-performing individual base model and hard voting by 1.08 and 0.21 percentage points, respectively. External validation (361 frames, 212 patients) reached a DSC of 83.55%, up to 3.68 percentage points above the lowest-performing base model. This uncertainty-aware framework enables robust stent segmentation without additional annotation, supporting automated stent analysis during PCI.