Uncertainty Quantification from Calcified Plaque to Outer Wall Estimation in CTA

Jiehyun Kim, Huy Q Phi, Grace J. Wang, Brett L. Cucchiara, Elias Johansson, Jae W. Song, Daniel Haehn

Keywords: Calcified Plaque, Outer Wall Estimation, CTA, Uncertainty Quantification

Registration Requirement: Yes

Abstract: Deep learning methods have shown strong performance for calcified plaque segmentation on CTA. However, the reliability of voxel-level predictions remains underexplored. We present an nnUNet-based framework that estimates voxel-wise uncertainty from a 5-fold ensemble. Rejecting low-confidence voxels improves both segmentation accuracy and calibration. We further extend the analysis to outer vessel wall estimation using variance-based tissue characterization, where ground-truth annotation is difficult and inconsistent. Our framework achieves an AUROC of 0.954 for voxel-level error detection and provides an integrated visualization of plaque hounsfield units (HU) values, prediction probability, uncertainty, and outer-wall variance. These results demonstrate that uncertainty-aware analysis supports both reliable plaque segmentation and exploratory vessel wall characterization. All implementations are open source on our GitHub repository: https://github.com/mpsych/CACTAS-UQ.

Reproducibility: https://github.com/mpsych/CACTAS-UQ

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 23