Table 1.
Brief overview of common data-driven techniques used in breast MRI.
| Technique | Advantages | Disadvantages | References |
|---|---|---|---|
| Supervised learning | |||
| Ensemble of decision trees | Decision using branches Variable significance and feature selection are included |
Prone to overfitting | [12–14] |
| [15, 16] | |||
| Random forest | High performance Compared to decision trees |
Prone to overfitting | [14, 17, 18] |
| [19] | |||
| Support vector machines | Transforms nonlinear classification problem into linear one High accuracy |
Difficult computation in high-dimensional data space | [20, 21] |
| [22, 23] | |||
| [24] | |||
| Neural networks | Weights need to be adapted for training Multiclass classification |
No strategy to determine network structure | [25–27] |
| [28, 29] | |||
| [30, 31] | |||
| Deep learning | State-of-the-art in image-derived features | Computationally intensive Hard to interpret |
[32, 33] |
| [34–36] | |||
| [37–39] | |||
|
| |||
| Unsupervised learning | |||
| Clustering (k-means) | Brief training duration | Number of clusters must be known in advance | [40, 41] |
| Topological data analysis | Interpretable data mapping Discovery of variable relationships |
Divided clusters due to mapping | [28, 42, 43] |