SwinD-Net: a lightweight segmentation network for laparoscopic liver segmentation

Figures & data

Figure 1. Swin Transformer Block.

Figure 2. The proposed lightweight network architecture.

Figure 3. The liver segmentation result comparison of our model and other SOTA methods.

Table 1. The five cross-validation comparisons result in a dice coefficient.

Model	Fold0	Fold1	Fold2	Fold3	Fold4	Average
U-Net	0.9802	0.9797	0.9804	0.9805	0.9805	0.9803
DeepLabv3+	0.9812	0.9818	0.9807	0.9818	0.9810	0.9813
U-Net++	0.9802	0.9806	0.9808	0.9798	0.9801	0.9803
Attention U-Net	0.9797	0.9796	0.9792	0.9796	0.9798	0.9796
TransUNet	0.9774	0.9776	0.9769	0.9772	0.9770	0.9772
MobileNetV3	0.9692	0.9684	0.9694	0.9692	0.9697	0.9692
UNeXt	0.9723	0.9720	0.9712	0.9718	0.9722	0.9719
Our Model	0.9732	0.9735	0.9724	0.9734	0.9726	0.9730

Table 2. The segmentation performance in dice coefficient, FLOPs and params of our model and other SOTA model comparison. All p-values of paired tests were greater than 0.05.

Model	IoU	Dice	FLOPs(M)	Params(M)
U-Net	0.9623 ± 0.0326	0.9803 ± 0.0227	203057.79	32.09
DeepLabv3+	0.9648 ± 0.0291	0.9819 ± 0.0165	69299.37	40.35
U-Net++	0.9632 ± 0.0327	0.9806 ± 0.0159	554388.95	36.63
Attention U-Net	0.9646 ± 0.0322	0.9799 ± 0.0237	266290.59	34.88
TransUNet	0.9638 ± 0.0329	0.9778 ± 0.0253	128715.06	93.19
MobileNetV3	0.9513 ± 0.0327	0.9709 ± 0.0238	166916.52	14.09
UNeXt	0.9524 ± 0.0314	0.9723 ± 0.0203	2281.79	1.47
Our Model	0.9541 ± 0.0287	0.9762 ± 0.0212	98.82	0.52

Table 3. The inference time comparison of our model and other SOTA methods on GPU and CPU.

Model	GPU Inference Time (ms)	CPU Inference Time (ms)
U-Net	31.97	1878.21
DeepLabv3+	22.30	660.93
U-Net++	102.41	3753.65
Attention U-Net	53.65	1848.88
TransUNet	35.82	1935.62
MobileNetV3	19.89	164.89
UNeXt	17.83	115.71
OurModel	9.61	47.49

Table 4. The ablation experiment result of our model.

Model	IoU	Dice
U-Net	0.9623 ± 0.0326	0.9803 ± 0.0227
Our Model without Depthwise separable convolutions	0.9548 ± 0.0275	0.9765 ± 0.0237
Our Model without Swin Transformer Block	0.8763 ± 0.0289	0.8978 ± 0.0273
Our Model	0.9541 ± 0.0287	0.9762 ± 0.0212

Figure 4. Some segmentation failure cases in the DSAD dataset. The ground truth segmentation is colored in green, and our model’s segmentation results are colored in red.