Table 3.

Segmentation results of different CNN architectures with dilated convolution for the segmentation of proximal femur.

Network	Dilation Rate (r)	AP ↑	DSC ↑	Precision ↑	Recall ↑	ASD [mm] ↓	MSD [mm] ↓
3D CNN F:32, L:4	1†	0.986 (0.002)	0.945 ± 0.029	0.948 ± 0.023	0.944 ± 0.052	0.45 ± 0.25	13.44 ± 13.14
	1, 2	0.988 (0.001)	0.950 ± 0.020	0.951 ± 0.017	0.949 ± 0.039	0.43 ± 0.23	8.77 ± 6.32
	1, 4	0.988 (0.002)	0.949 ± 0.019	0.948 ± 0.022	0.950 ± 0.036	0.43 ± 0.21	8.10 ± 4.73
	1, 8	0.988 (0.003)	0.948 ± 0.015	0.949 ± 0.023	0.948 ± 0.031	0.44 ± 0.19	8.05 ± 4.25
	1, 2, 4	0.988 (0.002)	0.948 ± 0.023	0.948 ± 0.021	0.949 ± 0.039	0.43 ± 0.24	8.14 ± 5.68
	1, 2, 4, 8	0.992 (0.002)	0.953 ± 0.016	0.954 ± 0.017	0.953 ± 0.030	0.39 ± 0.19	7.88 ± 4.33
2D CNN F:64, L:4	1†	0.979 (0.003)	0.937 ± 0.026	0.932 ± 0.036	0.943 ± 0.036	2.13 ± 1.22	42.22 ± 5.49
	1, 2	0.978 (0.002)	0.939 ± 0.025	0.937 ± 0.034	0.944 ± 0.037	2.04 ± 1.33	40.10 ± 7.34
	1, 4	0.984 (0.000)	0.943 ± 0.022	0.944 ± 0.026	0.944 ± 0.035	1.85 ± 1.13	39.70 ± 7.38
	1, 8	0.984 (0.002)	0.941 ± 0.023	0.941 ± 0.030	0.943 ± 0.034	2.04 ± 1.31	42.41 ± 7.60
	1, 2, 4	0.984 (0.005)	0.941 ± 0.025	0.941 ± 0.033	0.942 ± 0.034	2.02 ± 1.39	40.27 ± 7.76
	1, 2, 4, 8	0.986 (0.002)	0.946 ± 0.022	0.948 ± 0.024	0.944 ± 0.034	1.75 ± 1.24	40.03 ± 8.37

The best performing 2D CNN and 3D CNN models (based on cross-validation AP on Fig. 3) were used as a baseline for our experiments. AP values are derived for the individual segmentations and performance metrics are calculated from individual subject segmentations. F is the number of initial feature maps, L is the number of layers. ^† indicates the original CNNs presented in Table 1. Statistical analysis for comparing models using AP is presented in Table 4. AP data is presented using the median and interquartile range in parentheses.