Table 4.

Comparison of different methods for fetal brain segmentation in DW-MR images in terms of DSC for different levels of label noise. The highest DSC scores have been highlighted in bold text for each noise level. Our dual CNNs with iterative label update generated the highest DSC scores at small and medium noise levels, whereas the CNN trained with MAE loss generated better results for high noise levels.

	Clean data	noise level 1 (Method 1)	noise level 2 (Method 2)	noise level 3 (Method 3)	noise level 4 (Method 2)	noise level 5 (Method 3)	noise level 6 (Method 2)	noise level 7 (Method 2)
Average DSC of the training labels	1.000	0.949	0.924	0.854	0.807	0.790	0.777	0.742
Baseline CNN	0.878	0.889	0.862	0.846	0.755	0.730	0.736	0.724
Baseline CNN trained with MAE loss	-	0.881	0.864	0.840	0.780	0.741	0.778	0.760
Dual CNNs with iterative label update	-	0.906	0.895	0.886	0.849	0.804	0.773	0.732