ABSTRACT
Development and evaluation of medical image analysis algorithms often require a large number of validated and/or annotated samples. One promising approach to circumvent the Issues related to unbalanced or small datasets is to increase the available number of labeled or annotated samples by artificially supplementing the dataset. An existing dataset can be expanded to include additional synthetic samples with known properties, pathology, or diagnosis using these methods in order to overcome or reduce the shortcomings of the original set of samples. This chapter reviews some of the existing work in this domain and describes in more detail one specific approach for creating additional samples alongside examples from several modalities and application areas: data augmentation. More specifically, it uses this approach to augment datasets for mammography and chest computed tomography (CT) and to show applications of the augmented datasets in CAD training and evaluation of volumetry algorithms.
