Classification of medical images: instance space optimization models for Multiple Instance Learning

In the thesis work we focused on the study of Multiple Instance Learning (MIL) techniquesapplied to the binary classification of medical images. Unlike the standard classification,which consists in discriminating some points by assigning each of them to a class, a MILproblem consists in classifying different sets of points: these sets are calledbagsand thepoints inside them are namedinstances. In particular, differently from the classical super-vised classification, in the learning phase of a MIL problem only the labels of the bags areknown, whereas the labels of the instances inside them remain unknown. Problems of thistype fit very well for image classification, where the images are represented by the bags andthe sub-regions inside them correspond to the instances. In the medical field, for example,the image of a CT scan identifies a pathology on the basis not of the entire image, but onthe basis of some portions of it. In the binary case, where the aim is to discriminate betweenpositive and negative bags, the Multiple Instance Learning problems are based on the fol-lowing standard assumption (used when also the instances can belong to only two differentclasses): a bag is positive if at least one of its instances is positive and, vice-versa, it is nega-tive if all its instances are negative