CACTAS-UQ

This repository presents the uncertainty quantification (UQ) pipeline for calcified plaque segmentation in carotid CTA images. Segmentation and probability maps were generated using nnUNet, and voxel-level uncertainty was quantified from the probability maps.

---

Pipeline Overview

---

Steps

1. Compute baseline metrics

   • Intersection over Union (IoU)
   • Expected Calibration Error (ECE)

2. Apply uncertainty-based rejection

   • Recalculate IoU and ECE after removing uncertain voxels.

3. Report extended performance metrics

   • IoU, Dice, F1-score
   • Confusion matrix components (TP, FP, FN)

4. Visualize representative plaque cases

   • If plaque exists on one side, show a single example.
   • If plaques exist on both sides, visualize both.

5. Illustrate voxel-level FP/FN distributions

   • Compare before and after uncertainty-based rejection.

---

Uncertainty Map Example

The uncertainty maps highlight and remove ambiguous regions: