|
| Publication (year) | Approach | Pros (+)/cons (−) |
|
| [51] (2016) | (1) An 8-layered CNN | +parameter initialization |
| (2) SVM-based decision mechanism | +decision mechanism |
| (3) Compared to ML- and CNN-based models | −256 mid- and 2048 high-level features |
|
| [52] (2016) | (1) A 3-layered CNN | +medical instances for training |
| (2) SVM-based classification | −17 low- and 400 high-level features |
| (3) Compared to ML- and CNN-based models | −a shallow CNN model |
|
| [53] (2016) | (1) Transferred AlexNet | +parameter initialization |
| (2) SVM-based classification | +soft-voting-based decision mechanism |
| (3) Classifier-based soft voting | −29 low- and 3795 high-level features |
|
| [54] (2017) | (1) Modified LeNet | +semisupervised learning |
| (2) Graph based semisupervised learning | +a few labeled data used for training |
| (3) Feature dimension reduction | +less sensitive to initial labeled data |
| (4) Using unlabeled data | |
|
| [55] (2017) | (1) Modified AlexNet | +parameter initialization |
| (2) Multitask transfer learning | +improved generalizability |
|
| [56] (2017) | (1) Transferred the VGG | +parameter initialization |
| (2) SVM-based classification | +decision mechanism |
| (3) Compared to ML- and CNN-based models | −38 low- and 1472 high-level features |
|
| [57] (2017) | (1) R-CNN for detection and diagnosis | +minimal user intervention in image analysis |
| (2) Feature regression | −781 low-level features for CNN feature regression |
| (3) RF-based classification | |
|
| [58] (2017) | (1) A 4-layered CNN | +medical instances for training |
| | −a shallow CNN model |
|
| [59] (2017) | (1) A 3-layered CNN | +medical instances for training |
| (2) SVM-based classification | +image analysis in transformed domain |
| (3) Data augmentation | −a shallow CNN model |
|
| [60] (2018) | (1) VGG for feature extraction | +2 features selected for diagnosis |
| (2) Stepwise feature selection | |
| (3) SVM-based classification | |
|
| [61] (2018) | (1) Transferred AlexNet | +parameter initialization |
| (2) Data augmentation | |
| (3) Compared to CNN models | |
|
| [62] (2018) | (1) A 7-layered CNN | +parameter initialization |
| (2) Parasitic metric learning | +parasitic metric learning |
|
| [63] (2018) | (1) Transferred VGG | +parameter initialization |
| (2) Compared to CNN-based models | |
|
| [64] (2018) | (1) Transferred VGG/ResNet/Inception | +parameter initialization |
| (2) Comparison on 3 databases | +systematic comparison |
| | −time consuming |
|
| [65] (2018) | (1) YOLO and tensor structure | +medical instances for training |
| (2) Data augmentation | +simultaneous detection and classification |
|
| [66] (2018) | (1) Faster R-CNN and VGG | +medical instances for training |
| (2) Pretrained with the DDSM | +both detection and diagnosis |
| | +evaluated on a large-scale screening dataset |
|
| [67] (2018) | (1) GoogLeNet for feature extraction | +medical instances for training |
| (2) Attention mechanism for feature selection | +multiview and clinical information fusion |
| (3) LSTM for feature fusion | |
|
| [68] (2018) | (1) Transferred AlexNet/GoogLeNet | +parameter initialization |
| (2) Data augmentation | |
| (3) Compared to ML- and CNN-based models | |
|
