Table 2.

Fracture classification performance comparison between the proposed two-stage method (in bold), a one-stage approach which directly classifies full X-ray images using EfficientNet-B3 ensemble model, and other two-stage frameworks using YOLO ensemble model for ROI detection and ensemble classification model with other base networks (ResNet-18, DenseNet169, Swin Transformer, EfficientNet-B0, B1, B2 and B4)

Networks used in the Ensemble model	AUC	Acc	TPR (Sen.)	FPR (1-Spe.)	Precision	Recall	F1-score	Cohen’s kappa score
One-stage framework
EfficientNet-B3	0.69	0.68	0.86	0.80	0.75	0.86	0.80	0.07
Two-stage framework
YOLO + EfficientNet-B3 (proposed)	0.82	0.81	0.83	0.27	0.90	0.83	0.86	0.53
YOLO + ResNet-18	0.42	0.26	0.10	0.27	0.50	0.10	0.16	− 0.10
YOLO + DenseNet169	0.50	0.28	0.02	0.00	1.0	0.02	0.05	0.01
YOLO + Swin Transformer	0.50	0.74	1.00	1.00	0.74	1.0	0.85	0.00
YOLO + EfficientNet-B0	0.80	0.72	0.71	0.27	0.88	0.71	0.79	0.38
YOLO + EfficientNet-B1	0.84	0.65	0.60	0.20	0.89	0.60	0.71	0.30
YOLO + EfficientNet-B2	0.83	0.79	0.83	0.33	0.88	0.83	0.85	0.48
YOLO + EfficientNet-B4	0.81	0.72	0.69	0.20	0.91	0.69	0.78	0.40

Acc. accuracy, TPR true positive rate (sensitivity), FPR false positive rate (1-specificity)