Development and evaluation of a deep learning framework for pelvic and sacral tumor segmentation from multi-sequence MRI: a retrospective study

Table 3 Dice score^a of different input fusions for 2.5D MobilenetV2 + Unet

Input	T1-w	CET1-w	T2-w	DWI
T1-w	0.589
CET1-w		0.804
T2-w			0.804
DWI				0.714
T1-w & CET1-w	0.645	0.767
T2-w & CET1-w		0.810	0.804
DWI & CET1-w		0.805		0.689
T1-w & T2-w	0.655		0.801
DWI & T2-w			0.814	0.723
T1-w & T2-w & DWI	0.635		0.792	0.694
T1-w & T2-w & CET1-w	0.630	0.752	0.786
T1-w & DWI & CET1-w	0.598	0.716		0.710
CET1-w & T2-w & DWI		0.801	0.806	0.701
All-sequence model^b	0.659	0.763	0.819	0.723

Note:
a, Dice score is 2×the area of overlap divided by the total number of pixels in both images
b, All-sequence model represents the fusions of T1-w & CET1-w & T2-w & DWI

ISSN: 1470-7330