1320-GRP Uncertainty-Guided Conservative Propagation for Structured Inference in Vessel Segmentation
Location
https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php
Document Type
Event
Start Date
20-11-2026 2:00 PM
Description
Accurate vessel segmentation is critical for vascular analysis, but thin branches and ambiguous imaging regions often produce fragmented predictions. Existing segmentation networks typically lack an explicit mechanism that jointly considers prediction reliability, local structure, and refinement stability. We propose Uncertainty-Guided Conservative Propagation (UGCP), a differentiable, model-integrated module that treats segmentation logits as an evolving state. UGCP performs finite-step local updates in which relative uncertainty guides information flow, structure-aware modulation regulates neighboring interactions, and a source term anchors updates to the initial prediction. We evaluate UGCP using five-fold cross-validation on four public vascular datasets covering 2D invasive coronary angiography (ICA), retinal fundus imaging (FIVES), 3D coronary CT angiography (ImageCAS), and brain TOF-MRA (COSTA). Across 12 dataset–backbone configurations encompassing convolutional networks, Transformers, and vessel-specific architectures, UGCP improves centerline Dice (clDice), HD95, and Betti-0 error (B0E) in every configuration. clDice improvements range from 0.0032 to 0.0253, while B0E decreases by 10.9%–60.1%. These results demonstrate the potential of uncertainty-guided structured refinement to improve vessel continuity across diverse imaging modalities and network architectures.
1320-GRP Uncertainty-Guided Conservative Propagation for Structured Inference in Vessel Segmentation
https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php
Accurate vessel segmentation is critical for vascular analysis, but thin branches and ambiguous imaging regions often produce fragmented predictions. Existing segmentation networks typically lack an explicit mechanism that jointly considers prediction reliability, local structure, and refinement stability. We propose Uncertainty-Guided Conservative Propagation (UGCP), a differentiable, model-integrated module that treats segmentation logits as an evolving state. UGCP performs finite-step local updates in which relative uncertainty guides information flow, structure-aware modulation regulates neighboring interactions, and a source term anchors updates to the initial prediction. We evaluate UGCP using five-fold cross-validation on four public vascular datasets covering 2D invasive coronary angiography (ICA), retinal fundus imaging (FIVES), 3D coronary CT angiography (ImageCAS), and brain TOF-MRA (COSTA). Across 12 dataset–backbone configurations encompassing convolutional networks, Transformers, and vessel-specific architectures, UGCP improves centerline Dice (clDice), HD95, and Betti-0 error (B0E) in every configuration. clDice improvements range from 0.0032 to 0.0253, while B0E decreases by 10.9%–60.1%. These results demonstrate the potential of uncertainty-guided structured refinement to improve vessel continuity across diverse imaging modalities and network architectures.