Use Transfer Learning, or pre-train the U-Net on other datasets to improve performances in small dataset scenarios. Question: Should the U-Net necessarily be retrained from scratch after each iteration ?