PE-QUEST: QUality ESTimation-Driven Sample Selection for Pulmonary Embolism Segmentation
MCML Authors
Christian Wachinger
Prof. Dr.
Core PI
Abstract
Christian Wachinger
Prof. Dr.
Core PI
Abstract
3D pulmonary embolism (PE) segmentation enables quantitative assessment of thrombus burden in computed tomography pulmonary angiography (CTPA), yet public datasets offer only limited cases with voxel-level annotations. We propose PE-QUEST, a quality-guided pseudo-label filtering strategy for annotation-efficient semi-supervised 3D PE segmentation. Pseudo-labels generated from limited annotations are inherently noisy, and conventional confidence-based filtering fails to detect confident yet inaccurate predictions. PE-QUEST addresses this by training an independent quality prediction model that provides an external, image-aware reliability estimate for each pseudo-label. The quality model achieves an MAE of 0.0429 on the private test set and transfers zero-shot across domains without fine-tuning (AUC > 0.93 on public datasets). In semi-supervised experiments with only 20 labeled cases, joint quality-and-uncertainty filtering retains 55% of pseudo-labels and matches fully supervised performance (DSC 0.580 vs. 0.567, NSD 0.712 vs. 0.698).
inproceedings SJL+26a
TIA @MICCAI 2026
3rd International Workshop on Thoracic Image Analysis at the 29th International Conference on Medical Image Computing and Computer Assisted Intervention. Strasbourg, France, Sep 27-Oct 01, 2026. To be published. Preprint available.Authors
Q. Sun • B. Jian • Z. Liu • Y. Li • C. WachingerLinks
URL GitHubResearch Area
BibTeXKey: SJL+26a