Automated classification of hip fractures using deep convolutional neural networks with orthopedic surgeon-level accuracy: ensemble decision-making with antero-posterior and lateral radiographs

Figures & data

Figure 1. Image preprocessing for the convolutional neural network model training and validation. We cropped images to a minimum region containing the femoral head and the greater and lesser trochanters in both the AP (A) and lateral (B) hip radiographs. On the AP radiographs, the fractured hip (left white box) was cropped and the side contralateral from the fractured hip (right white box) was cropped as the non-fractured hip. AP = anteroposterior.

Table 1. Baseline patient characteristics

Factor	Femoral neck fracture n = 569	Trochanteric fracture n = 466	Non-fracture n = 234
Age, mean (SD)	81.3 (11.4)	85.2 (10.0)	68.8 (16.2)
Sex (M/F), n	136/433	105/361	81/153

Figure 2. Comparison of the accuracy between the AP, lateral, and both views of the CNN and the 4 orthopedic surgeons. In the CNN model, the accuracy of the diagnosis based on both views was statistically better than the AP view alone and the lateral view alone. The accuracy of diagnosis based on the AP view alone was statistically better than the lateral view alone. The same trend was also seen with the board-certified orthopedic surgeons. AP = anteroposterior; CNN = convolutional neural network.

Table 2. Accuracy, p-value of the accuracy compared with the CNN, average recall, precision, and F1 score of the diagnostic performance of the CNN and the 4 orthopedic surgeons based on both the anteroposterior and the lateral radiographs

CNN/ surgeon	Accuracy (CI)	p-value ^a	Average recall	Average precision	Average F1 score
CNN	0.98 (0.96–1.00)	–	0.98	0.98	0.98
Board certified
1	0.92 (0.88–0.96)	0.01	0.92	0.92	0.92
2	0.95 (0.91–0.98)	0.1	0.95	0.95	0.95
Resident
1	0.87 (0.82–0.93)	0.0006	0.87	0.89	0.88
2	0.78 (0.71–0.85)	< 0.0001	0.78	0.82	0.80

^a compared with CNN

CI = 95% confidence interval;

CNN = convolutional neural network.

Table 5. Diagnostic performance of the CNN and the 4 orthopedic surgeons based on both the anteroposterior and the lateral radiographs

CNN/	Femoral neck fracture			Trochanteric fracture			Non-fracture
surgeon	Recall	Precision	F1 score	Recall	Precision	F1 score	Recall	Precision	F1 score
CNN	1.00	0.98	0.99	0.94	1.00	0.97	1.00	0.96	0.98
Board certified
1	0.96	0.89	0.92	0.94	0.92	0.93	0.86	0.96	0.91
2	0.96	0.96	0.96	0.96	0.96	0.96	0.92	0.92	0.92
Resident
1	0.96	0.81	0.88	1.00	0.86	0.93	0.66	1.00	0.80
2	0.90	0.71	0.80	0.96	0.77	0.86	0.48	0.96	0.64

CNN = convolutional neural network.

Table 8. Interrater reliability presented with Cohen’s kappa of the orthopedic surgeons

	Board certified		Resident
	orthopedic	surgeon	orthopedic	surgeon
Surgeon	1	2	1	2
Board certified
1	–	0.85	0.78	0.73
2	0.85	–	0.76	0.78
Resident
1	0.78	0.76	–	0.66
2	0.73	0.78	0.66	–

Figure 3. Representative radiographs of hip fractures. The AP (A) and lateral (B) radiographs of a trochanteric fracture, which the CNN misdiagnosed as a non-fracture, but all the orthopedic surgeon diagnosed correctly. The AP (C) and lateral (D) radiographs of a neck fracture, which 3 of the 4 orthopedic surgeons misdiagnosed as a non-fracture, but the CNN diagnosed correctly. The AP (E) and lateral (F) radiographs of a trochanteric fracture, which 3 of the 4 orthopedic surgeons misdiagnosed as a non-fracture or a neck fracture, but the CNN diagnosed correctly. AP = anteroposterior; CNN = convolutional neural network.