Also from the paper for more context:

> The dataset was randomly divided into 70% training subset (1420 for prostate and 692 for pancreatic), 15% validation subset (304 and 148), and 15% test subset (304 and 148).