CARDIAG Benchmark for Deep Learning in Coronary Angiography
A new benchmark called CARDIAG has been introduced for evaluating deep learning models that classify pixels in coronary angiograms according to SYNTAX classes. The study covers 24 architectures, from classic convnets to state-space-based vision algorithms. The dataset is multi-center, multi-label, and includes SYNTAX labels, binary masks, uncertainty masks, segmentation masks, intermediate frames, and non-sensitive DICOM metadata. Metrics account for diameter error, overlap, centerline quality, and calibration. The best-performing model is ConvNeXt V2 encoder with DeepLab V3 Plus decoder.
Key facts
- CARDIAG is a new benchmark for dense segment classification of coronary angiograms.
- It evaluates 24 deep learning architectures.
- The dataset is multi-center and multi-label.
- SYNTAX classes are used for pixel classification.
- Metrics include diameter error, overlap, centerline quality, and calibration.
- Data includes SYNTAX labels, binary, uncertainty, and segmentation masks.
- Intermediate frames and non-sensitive DICOM metadata are provided.
- ConvNeXt V2 encoder with DeepLab V3 Plus decoder is the best performer.
Entities
Institutions
- arXiv