{06/09} · Imaging
All workPreprocessing and Augmentation for CAD Detection
Do preprocessing and augmentation choices change coronary artery disease detection?
A four phase benchmark of 11 preprocessing methods, 10 augmentations and six classifiers for coronary artery disease.
- 780model fits
- 8,020images from 1,168 patients
- 0.670best AUROC, only 0.014 above baseline
What the record supports
Claim
EfficientNet-B3 gave the most consistent saliency maps.
Evidence
780 model fits on 8,020 images from 1,168 patients.
Limits, stated plainly
Best AUROC was 0.670, only 0.014 above baseline. Frangi and homomorphic filtering fell to near chance.