Normalizing Flows for Out-of-Distribution Detection: Application to Coronary Artery Segmentation
Coronary computed tomography angiography (CCTA) is an effective imaging modality, increasingly accepted as a first-line test to diagnose coronary artery disease (CAD). The accurate segmentation of the coronary artery lumen on CCTA is important for the anatomical, morphological, and non-invasive func...
Gespeichert in:
Veröffentlicht in: | Applied sciences 2022-04, Vol.12 (8), p.3839 |
---|---|
Hauptverfasser: | , , , , , , , , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Coronary computed tomography angiography (CCTA) is an effective imaging modality, increasingly accepted as a first-line test to diagnose coronary artery disease (CAD). The accurate segmentation of the coronary artery lumen on CCTA is important for the anatomical, morphological, and non-invasive functional assessment of stenoses. Hence, semi-automated approaches are currently still being employed. The processing time for a semi-automated lumen segmentation can be reduced by pre-selecting vessel locations likely to require manual inspection and by submitting only those for review to the radiologist. Detection of faulty lumen segmentation masks can be formulated as an Out-of-Distribution (OoD) detection problem. Two Normalizing Flows architectures are investigated and benchmarked herein: a Glow-like baseline, and a proposed one employing a novel coupling layer. Synthetic mask perturbations are used for evaluating and fine-tuning the learnt probability densities. Expert annotations on a separate test-set are employed to measure detection performance relative to inter-user variability. Regular coupling-layers tend to focus more on local pixel correlations and to disregard semantic content. Experiments and analyses show that, in contrast, the proposed architecture is capable of capturing semantic content and is therefore better suited for OoD detection of faulty lumen segmentations. When compared against expert consensus, the proposed model achieves an accuracy of 78.6% and a sensitivity of 76%, close to the inter-user mean of 80.9% and 79%, respectively, while the baseline model achieves an accuracy of 64.3% and a sensitivity of 48%. |
---|---|
ISSN: | 2076-3417 2076-3417 |
DOI: | 10.3390/app12083839 |