Evaluation of Multimodal Algorithms for the Segmentation of Multiparametric MRI Prostate Images

Table 5

Comparison of previously reported DSC for prostate segmentation using DL networks trained with the stated image inputs and the number of training subject data as reported in the literature. AAM-CNN = active appearance model followed by a CNN. ^#With an endorectal coil. ^&With surface coil. ^$Training and test data are from the same datasets. ^%Averaged across 2 datasets. ^ΒSlices instead of subjects. NCI-ISBI 2013 Challenge dataset consisted of PROSTATE-DIAGNOSIS and Prostate-3T datasets (refer to Supplementary Table 1).

Networks	Input images	No. of training iterations	No. of training subjects	No. of test subjects	DSC (WG)	DSC (PZ)	Ref.
UNet	T2w	15,000	173	59		—	[6]
VNet						—
HighRes3DNet						—
HolisticNet						—
DenseVNet						—
Adapted UNet						—
ConvNet	T2w	80	141	12		—	[8]
Cascaded 2D UNet	DWI ( value = 1000 s/mm^2, preprocessing), T2w	100	76	51			[18]
DSCNN	T2w		77	4	0.885	—	[19]
PSFCN	T2w	80,000	—	20		—	[20]
Volumetric	T2w	10,000	50	30	0.894	—	[21]
ConvNet
SegNet	T2w	—	19	4	0.73	—	[22]
AAM-CNN	T2w^#	—	100	20	0.925	—	[23]
3D Multistream	T2w	—	220 (GE)	22	^$	^$	[24]
UNet	(axial, sagittal, coronal)		330 (Siemens); 550 (Combined)	33; 55	^$; ^%	^$; ^%
FCN	T2w	—	40 (542^Β)	82^Β			[25]
SegNet	T2w^&	—	11 (229^Β)	72^Β
UNet
DeepLabV3+
UNet	T2w	36,952	141	47			[26]
Cascaded UNet		—
PSPNet		—
Dense-2 UNet		35,760