Current Status and Future Perspectives of Artificial Intelligence in Magnetic Resonance Breast Imaging

Table 2

Segmentation applications in breast MRI.

DL technique	Evaluation results	Used dataset	Reference
2D U-net applied slice-by-slice	Dice = 95.90 ± 0.74; Acc = 98.93 ± 0.15 Sn = 95.95 ± 0.69; Sp = 99.34 ± 0.17	42 patients DCE-MRI	[69]
3TP U-net	Dice = 61 ± 11.84; Acc = 99 ± 0.01; Sn = 68.28 ± 9.73; Sp = 100 ± 9.73	35 DCE-MRI 4D data	[60]
GOCS-DLP shape prior based on semantic segmentation based on DL	Dice = 77 ± 13	117 patients DCE-MRI, T2- and T1-weighted images	[70]
2D U-net applied slice-by-slice	Dice = 97	50 DCE-MR images	[71]
Hierarchical multistage U-net with dice loss	Dice = 72 ± 24; Sn = 75 ± 23	Training set: 224 DCE-MRI cases; test set: 48 DCE-MRI cases	[72]
Comparison of 2D U-net and 2D SegNet models with transfer learning from DCE-MRI to DWI	Dice = 72 ± 16	Training: 39 DCE-MR cases and 15 DWI-MR cases; testing: 10 representative DWI-MR slices	[73]
2D U-net applied slice-by-slice to multiplanar sections followed by voxel-level fusion	Dice = 96 ± 0.3; Acc = 99.16 ± 0.13; Sn = 96.85 ± 0.47; Sp = 96.85 ± 0.47	Training: 42 + 88 T1-weighted MRI series (10-fold cross-validation)	[61]

The most common performance measures are the Dice coefficient and the by-voxel accuracy (ACC), sensitivity (Sn), and specificity (Sp). All performance values reported are percentages.