Review Article

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

Table 4

Selected studies reporting classification of breast lesions in breast MRI using DL.

DL techniqueEvaluation resultsDatasetReference

3D CNN from scratchAUC = 0.739 (2D)
AUC = 0.801 (3D)
143 DCE-MR cases (M: 77, B: 66)[35]
CNN (ResNet50) fine-tunedAUC = 0.97–0.99Training: 33 patients with 153 lesions (M: 91, B: 62)
Testing: 74 patients with 74 lesions (M: 48, B: 26)
[90]
Cross-modal DL (mammography and MR), trained from scratchAcc = 94
AUC = 0.98
123 DCE-MR + T1W 282 mammography images[91]
Dense convolutional LSTMAcc = 0.847
Precision = 78.2
Sn = 81.5
72 lesions (M: 27, B: 45) DCE-MRI and DWI-MRI[92]
DenseNetAUC = 0.811576 lesions (M: 368, B: 149, FU: 59)
Ultrafast DCE-MRI, T2, and DWI
[93]
CNN (AlexNet) fine-tuned from ImageNet on the second postcontrast frame, LSTM model for final predictionAcc = 76
AUC = 0.76
42 DCE-MR images, 67 lesions (M: 42, B: 25) 10-fold cross-validation[59]
CNN (ResNet34) fine-tuned best three inputs out of 85 combinationsAUC = 0.88 (95% confidence interval: 0.86–0.89)447 patients, 1294 lesions (M: 787, B: 507) multiparametric DCE-MR + T2W
10-fold cross-validation
[94]
MIP + off-the-shelf CNN (VGG) + SVMAUC = 0.88 ± 0.01690 DCE-MR cases (M: 478, B: 212)
5-fold cross-validation
[32]
Multiscale 3D CNN (trained from scratch) inputs: five timepoints T1W DCE-MR and T2WAUC = 0.89 ± 0.01408 patients (M: 305, B: 103) multiparametric
DCE-MR
5-fold cross-validation
[95]
Off-the-shelf CNN (VGG) + SVM target: different molecular subtypesAUC = 0.65 (pretrained)
AUC = 0.58 (from scratch)
270 DCE-MR images (90 luminal A, 180 other 3 subtypes)
10-fold cross-validation
[96]
3TP-CNN pretrained on ImageNetAcc = 74
AUC = 0.81
F1 = 0.78
39 lesions (M: 36, B: 22)
DCE-MRI sequences
10-fold cross-validation
[97]
Three-channel (pre- and postcontrast) CNN (VGG) fine-tuned for classificationAUC = 0.88
(CNN + LSTM)
AUC = 0.84 (CNN-only)
703 DCE-MR dataset (M: 482, B: 221)
80% training + validation, 20% testing
[58]
3D ResNet trained from scratch with ad hoc embedding loss weakly supervised localization with feature correlation attention mapAcc = 85.5
AUC = 0.902
1715 subjects (M: 1137, B: 578)
Training: 1204 subjects
Testing: 346 subjects
[98]

For each study, we report the number of histologically verified benign (B) and malignant (M) lesions or cases; benign lesions without biopsy with at least 12-month follow-up (FU) are also indicated. Histology is used as ground truth in all studies.