Fig. 2.

Schematic overview of the evaluation setup in a single random-split cross-validation iteration for the single-observer and multi-observer models. For the single-observer models, here illustrated for observer CNN, for both the patients included in the training and in the testing set, each patient appears one time with the segmentation of that single observer. For the multi-observer model, the test set is exactly the same as the single-observer model. However, in the training set, each patient appears three times, each time with a different segmentation from one of the three other observers (STUD2, PhD, and RAD). Hence, in the multi-observer model, the training set size is effectively tripled compared to the single-observer model, while the test set remains unchanged. (Color figure online)