Figure 4.

Comparison between traditional image segmentation workflows and machine learning workflows. Left: the traditional approach, where the statistical algorithm to solve the classification or segmentation problem is designed by manually finding rules in the dataset which are then applied and tested on the whole dataset. If no satisfying outcome is observed, the process is repeated with another choice of parameters. Right: the machine learning approach. Researchers must decide on network architecture and hyperparameters, such as depth, learning rate and optimization algorithm. Once the network is designed, it is fed with training data to learn accurate representations and transformative functions on its own. If the results are not satisfying, we can either keep training the network or adapt its architecture.