Table 1.

Results from training on the CVC-ClinicDB for the U-Net, Attention U-Net and Focus U-Net. The best result is seen with the addition of the Focus Gate, short skip connections, deep supervision and use of the Hybrid Focal loss.

Model	Loss function	Focal parameter γ	Short skip connections	Deep supervision	mDSC	mIoU	Recall	Precision	FPS
U-Net	DSC + CE	–	✗	✗	0.828 ± 0.021	0.747 ± 0.022	0.817 ± 0.024	0.877 ± 0.021	27
Attention U-Net	DSC + CE	–	✗	✗	0.801 ± 0.019	0.705 ± 0.023	0.799 ± 0.012	0.844 ± 0.030	25
Focus U-Net	DSC + CE	1	✗	g ✗	0.838 ± 0.018	0.755 ± 0.018	0.833 ± 0.016	0.876 ± 0.028	25
Focus U-Net	DSC + CE	1.25	✗	✗	0.844 ± 0.011	0.762 ± 0.011	0.844 ± 0.018	0.876 ± 0.020	25
Focus U-Net	DSC + CE	1.5	✗	✗	0.842 ± 0.025	0.755 ± 0.027	0.845 ± 0.038	0.866 ± 0.016	25
Focus U-Net	DSC + CE	2	✗	✗	0.817 ± 0.025	0.728 ± 0.030	0.817 ± 0.023	0.859 ± 0.026	25
Focus U-Net	DSC + CE	3	✗	✗	0.825 ± 0.022	0.736 ± 0.022	0.820 ± 0.035	0.863 ± 0.022	25
Focus U-Net	DSC + CE	1.25	✓	✗	0.867 ± 0.018	0.800 ± 0.011	0.852 ± 0.023	0.908 ± 0.017	25
Focus U-Net	HFL	1.25	✓	✗	0.869 ± 0.013	0.797 ± 0.014	0.870 ± 0.017	0.892 ± 0.010	2
Focus U-Net	HFL	1.25	✓	✓	0.875 ± 0.016	0.801 ± 0.017	0.878 ± 0.013	0.889 ± 0.018	25