Table 3.

Segmentation performances of Jaccard coefficient (JC), Dice similarity coefficient (DSC), precision (PR), and recall (RC) for the maxillary sinus (MS), maxillary sinus air region (MSA), and the lesions inside the maxillary sinus (MSL) by a 2.5D (U-net + +) network with different ensemble methods (unanimous, affirmative, and majority) (*: significant differences among 2.5D (U-net + +) networks with unanimous, affirmative and majority methods, and †: between 2.5D (U-net + +) of ResNet101 and Densenet169 with the majority method in JC and DSC (p < 0.05))

	MSA				MS				MSL
	JC	DSC	PR	RC	JC	DSC	PR	RC	JC	DSC	PR	RC
ResNet101
Unanimous	0.931 ± 0.041	0.964 ± 0.023	0.992 ± 0.003	0.937 ± 0.041	0.892 ± 0.062	0.942 ± 0.036	0.993 ± 0.004	0.898 ± 0.062	0.652 ± 0.170	0.775 ± 0.135	0.883 ± 0.092	0.698 ± 0.160
Affirmative	0.935 ± 0.029	0.966 ± 0.016	0.948 ± 0.020	0.985 ± 0.017	0.932 ± 0.024	0.965 ± 0.013	0.947 ± 0.022	0.984 ± 0.010	0.767 ± 0.121	0.862 ± 0.084	0.857 ± 0.062	0.871 ± 0.113
Majority	0.947 ± 0.031	0.973 ± 0.017	0.978 ± 0.007	0.968 ± 0.029	0.940 ± 0.023*	0.969 ± 0.012*	0.979 ± 0.010	0.959 ± 0.021	0.770 ± 0.132*	0.864 ± 0.091*	0.904 ± 0.054	0.831 ± 0.122
Densenet169
Unanimous	0.931 ± 0.039	0.964 ± 0.022	0.991 ± 0.004	0.939 ± 0.039	0.921 ± 0.027	0.959 ± 0.015	0.991 ± 0.006	0.929 ± 0.028	0.715 ± 0.162	0.822 ± 0.124	0.888 ± 0.079	0.771 ± 0.151
Affirmative	0.932 ± 0.032	0.964 ± 0.018	0.943 ± 0.025	0.987 ± 0.018	0.927 ± 0.030	0.962 ± 0.017	0.938 ± 0.031	0.988 ± 0.008	0.759 ± 0.118	0.858 ± 0.082	0.847 ± 0.055	0.874 ± 0.120
Majority	0.948 ± 0.03	0.973 ± 0.016	0.977 ± 0.008	0.970 ± 0.029	0.947 ± 0.018*†	0.973 ± 0.010*†	0.974 ± 0.013	0.971 ± 0.015	0.787 ± 0.127*†	0.875 ± 0.086*†	0.897 ± 0.053	0.858 ± 0.120