Vessel segmentation for X-ray coronary angiography using ensemble methods with deep learning and filter-based features

Table 4 A comparison of model performance

	Precision	Sensitivity	Specificity	F1 Score	AUROC	IoU
U-Net	0.867 ± 0.073	0.810 ± 0.122	0.993 ± 0.005	0.831 ± 0.082	0.902 ± 0.060	0.719 ± 0.115
DeepLabV3+	0.862 ± 0.082	0.828 ± 0.096	0.992 ± 0.006	0.838 ± 0.081	0.909 ± 0.047	0.726 ± 0.088
Inception- ResNet-v2 U-Net	0.904 ± 0.072	0.805 ± 0.133	0.995 ± 0.004	0.842 ± 0.089	0.900 ± 0.066	0.737 ± 0.120
DenseNet121 U-Net	0.891 ± 0.053	0.824 ± 0.145	0.994 ± 0.004	0.845 ± 0.091	0.909 ± 0.071	0.741 ± 0.117
Resnet101 U-Net	0.865 ± 0.072	0.819 ± 0.122	0.992 ± 0.005	0.832 ± 0.068	0.906 ± 0.060	0.718 ± 0.095
Unsupervised with Deep forest	0.832 ± 0.073	0.911 ± 0.096	0.990 ± 0.005	0.863 ± 0.048	0.95 ± 0.046	0.762 ± 0.071
Tomek Links with Deep Forest	0.884 ± 0.061	0.867 ± 0.124	0.993 ± 0.004	0.867 ± 0.066	0.930 ± 0.061	0.770 ± 0.094
Cluster centroid with Deep forest	0.868 ± 0.067	0.873 ± 0.107	0.993 ± 0.004	0.864 ± 0.062	0.933 ± 0.053	0.765 ± 0.087
Unsupervised with GBDT	0.864 ± 0.066	0.894 ± 0.104	0.992 ± 0.004	0.872 ± 0.051	0.943 ± 0.051	0.776 ± 0.075
Tomek Links with GBDT	0.885 ± 0.06	0.872 ± 0.123	0.994 ± 0.004	0.870 ± 0.066	0.933 ± 0.060	0.775 ± 0.094
Cluster Centroid with GBDT	0.857 ± 0.073	0.902 ± 0.084	0.992 ± 0.004	0.874 ± 0.048	0.947 ± 0.041	0.779 ± 0.072
DenseNet121 U-Net [43]	0.858 ± 0.071	0.873 ± 0.109	0.991 ± 0.006	0.858 ± 0.057	0.926 ± 0.068	0.755 ± 0.082

Bold values are denotes the best-performing statistic of a metric among all models tested.

ISSN: 1471-2342