Sample Size for Training and Testing: Segment Anything Models and Supervised Approaches

The problem of determining the minimum amount of data required to train and test an artificial intelligence model has received substantial attention in the literature. In this chapter, we first review key concepts on the topic, then we survey selected theoretical and experimental results from the open literature, and in the end we present, as a case study, experiments we performed ourselves on the semantic segmentation of radiology images. A discussion from both a theoretical and an experimental point of view is required because the two approaches have complementary insights to offer. Theory provides general guidelines to avoid pitfalls during all phases of design: data collection, model design, training, and testing. Experimental results show what the current state of the art is in terms of performance and provide practical advice on which techniques have proven to be the most effective; for a more comprehensive study, we tested both supervised and zero-shot segmentation approaches, such as the “Segment Anything Model” (better known as SAM).