Table 1.
Summary of studies included in this review
| Study | Source of data (where data were obtained) | # of patients | Benign lesions | Malignant lesions | Image type | DL architecture | Box or contour | Ground truth | Cross validation | Heatmap | AUC | Accu | Sens | Spec |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Adachi [13] | Tokyo Medical and Dental University (Japan) | 72 | 20 | 52 | MIP, DCE | RetinaNet | Bounding box | Rad per path; path, > 1y F/U | Hold-out method | No | 0.93 | Not reported | 0.93 | 0.83 |
| Antropova [24] | University of Chicago | 690 | 212 | 478 | MIP DCE; DCE subtraction | ConvNetVGGN | Manual ROI with bounding box | Rad per path; biopsy proven | Fivefold | No |
0.88, 0.80, 0.83 |
Not reported | Not reported | Not reported |
| Ayatollahi [22] | Radbound University Medical Center (the Netherlands) | 462 | 207 | 365 | DCE | RetinaNet | Bounding box | Path, 2y F/U | Tenfold | No | 0.90 | Not reported | 0.95 | Not reported |
| Feng [24] | Tangdu Hospital and Xi'an International Medical Center Hospital (China) | 100 | 32 | 68 | DCE + DWI | KFLI | Bounding box | Path | Hold-out method | No | 0.86 | 0.85 | 0.85 | 0.86 |
| Fujioka [25] | Tokyo Medical and Dental University (Japan) | 72 |
12 normal 13 benign |
47 | MIP DCE | InceptionResNetV2 Others | None |
Path, > 1y F/U |
Tenfold | Yes | 0.83 | Not reported | 0.75 | 0.96 |
| Haarburger [26] | University Hospital Aachen (Germany) | 408 | 103 | 305 | DCE + T2 |
Mask-R-CNN; Retina U-Net; 3D ResNet18 Naive; ResNet18; Curriculum |
Coarse localization by radiologist Whole breast images |
Path, two-year F/U | Fivefold | Yes |
0.88 0.89 0.50 0.89 0.93 |
0.77 0.82 0.45 0.81 0.93 |
Not reported | Not reported |
| Herent [27] | Journées Francophones de Radiologie (France) | 335 | 212 | 123 |
Post-contrast T1 |
CNN Resnet-50 | Manual bounding boxes | Path | Threefold | No | 0.82 | Not reported | Not reported | Not reported |
| Hu [28] | University of Chicago | 616 | 199 | 728 | DCE, T2 (not coupled) | CNN |
Path, Rad reports |
Fivefold | 0.87 | Not reported | 0.78 | 0.79 | ||
| Li [29] | Zhiejiang Cancer Hospital, China | 143 | 66 | 77 | DCE | 2D vs. 3D CNN | Bounding box | Path verified | No | 0.80 3D |
0.78 3D |
0.74 3D |
0.82 3D |
|
| Liu [30] | Multiple institutions in the US (ISPY-1 data)* | 438 | DCE | CNN | Square cropping in sagittal plane |
Path Rad annotation |
Fivefold | No | 0.92 | 0.94 | 0.74 | 0.95 | ||
| Marrone [31] | University of Naples (Italy) | 42 | 25 | 42 | DCE 4D | AlexNet CNN | Manual ROI | Path | Tenfold | No | 0.76 | 0.76 | 0.83 | 0.79 |
| Rasti [32] | Imaging Center of Milad Hospital (Tehran) | 112 | 59 | 53 | DCE 1st Post-subtraction | ME-CNN | Path | Fivefold | No | 0.99 | 0.96 | 0.98 | 0.95 | |
| Truhn [33] | University of Aachen (Germany) | 447 | 507 | 787 | T2, pre- and post-contrast | CNN ResNet 18 | Manual segmentation |
Path, F/U |
Tenfold | No | 0.88 | Not reported | 0.78 | 0.85 |
| Wu [34] | Beijing University People's Hospital (China) | 130 | 59 | 71 | DCE | CNN | Bounding box | Path | Yes | 0.91 | 0.88 | 0.86 | Not reported | |
| Yurttakal [35] | Haseki Training and Research Hospital (Turkey) | 200 | 98 | 102 | DCE subtraction | CNN |
Cropping tumorous regions Rectang box |
Path, Rad |
No | Not reported | 0.98 | 1.0 | 0.97 | |
| Zheng [36] | Renji Hospital (China) | 72 | 45 | 27 |
DCE + DWI Multi-timepoints |
DC-LSTM, ResNet50 | Labels by radiologists, then cropped 40 × 40x40 | Rad per path; Path for B/M benign vs. mal | Threefold | No | Not reported | 0.85 | Not reported | Not reported |
| Zhou [37] |
Not specified (Likely China) |
133 | 62 | 91 | DCE | CNN ResNet50 | Segment by fuzzy C-means after radiologist indicated location | Path | Tenfold | No | 0.97–0.99 | 0.89 | 0.94 | 0.81 |
| Zhou [38] | The Fifth Medical Center of Chinese PLA General Hospital (China) | 307 | 101 | 206 | DCE |
3D DenseNet GAP 3D DenseNet GMP Ensemble |
Bounding box |
Path, 3y F/U |
Yes |
0.86 0.86 0.86 |
0.81 0.81 0.83 |
0.86 0.92 0.91 |
0.70 0.61 0.69 |
Studies are arranged alphabetically. DCE: dynamic contrast enhancement, DWI: diffusion-weighted imaging, MIP: maximum intensity projection, T2: T2-weighted MRI. F/U: follow-up, Rad: radiology, Path: pathology, CNN: convolutional neural network, AUC: area under the curve, Accu: accuracy, Sens: sensitivity, and Spec: specificity
*public dataset (see text for link)