Skip to main content
. 2021 Mar 24;11:644703. doi: 10.3389/fonc.2021.644703

Table 3.

Performance of the proposed network with different CNN components and configurations.

CNN Width Kernel Size Dropout Rate ROI Attention Skip Attention Residual Module Activation Function SSIM R SSIM+R Percent Difference
3 × 32 5 × 5 ×5 0.1 1 1 1 PReLU 0.6757 0.6305 1.3062 -4.68
4 × 32 5 × 5 × 5 0.1 1 1 1 PReLU 0.7052 0.6651 1.3703 0
5 × 32 5 × 5 × 5 0.1 1 1 1 PReLU 0.6831 0.6272 1.3103 -4.38
4 × 32 5 × 5 × 5 0 1 1 1 PReLU 0.6884 0.6236 1.3120 -4.25
4 × 32 5 × 5 × 5 0.2 1 1 1 PReLU 0.7014 0.6634 1.3648 -0.40
4 × 32 5 × 5 × 5 0.3 1 1 1 PReLU 0.7006 0.6519 1.3525 -1.30
4 × 32 3 × 3 × 3 0.1 1 1 1 PReLU 0.6867 0.6361 1.3228 -3.47
4 × 32 5 × 5 × 5 0.1 1 1 1 LReLU 0.6911 0.6458 1.3369 -2.44
4 × 32 5 × 5 × 5 0.1 1 1 1 ReLU 0.6882 0.6012 1.2894 -5.90
4×32 5×5×5 0.1 / 1 1 PReLU 0.6819 0.6276 1.3095 -4.44
4×32 5×5×5 0.1 1 / 1 PReLU 0.6879 0.6400 1.3279 -3.09
4×32 5×5×5 0.1 1 1 / PReLU 0.6877 0.6085 1.2962 -5.41

The bold values indicates the optimal configurations.