Table 1 |.

Comparison of hand-crafted feature engineering and DL-based approaches

Aspect	Hand-crafted feature engineering	DL-based approaches
Development	Usually developed in close collaboration with expert pathologists and oncologists; these approaches are based on a given aspect of domain knowledge, and thus their development can be complicated and time consuming	Developed though unsupervised feature learning, dependent on the existence of learning sets and annotated exemplars from the categories of interest; network design usually involves a focus on fine-tuning the algorithm to maximize accuracy while minimizing processing time
Generalizability	Approaches usually tailored for a specific cancer subtype and/ or tissue of origin; for example, features relating to glandular morphology would only apply to diseases with an abundance of tubules and glands	Using approaches such as transfer learning, a network trained on a particular disease subtype could be applied to other subtypes as well
Training datasets	The particular set of features or markers of interest are known, and thus small-sized training datasets are needed for feature engineering	Large amounts of well-annotated data with several exemplars from the different target categories or classes of interest are often required
Interpretability	Feature engineering is typically associated with attributes of the disease domain, and thus the features tend to have greater interpretability and a stronger morphological and biological underpinning in comparison with DL-based approaches	Representations can often be difficult to interpret and, despite new emerging approaches aimed at providing clarity (such as visual attention maps), these approaches are still largely considered ‘black-box' methods
Clinical deployment	On account of being more interpretable than DL-based approaches, hand-crafted features might be more likely to be used for high-level decision-making, such as that relating to disease prognosis or prediction of benefit from therapy	Might be more appropriate in situations in which the need to ‘explain the decision' is reduced; such situations could include low-level tasks such as object detection or segmentation

DL, deep learning.