%0 Conference Proceedings
%4 sid.inpe.br/sibgrapi/2015/07.31.16.36
%2 sid.inpe.br/sibgrapi/2015/07.31.16.36.52
%T Active Learning with Interactive Response Time and its Application to the Diagnosis of Parasites
%D 2015
%A Saito, Priscila T. M.,
%A de Rezende, Pedro J.,
%A Falcão, Alexandre Xavier,
%@affiliation Federal University of Technology - Parana
%@affiliation University of Campinas
%@affiliation University of Campinas
%E Segundo, Maurício Pamplona,
%E Faria, Fabio Augusto,
%B Conference on Graphics, Patterns and Images, 28 (SIBGRAPI)
%C Salvador, BA, Brazil
%8 26-29 Aug. 2015
%I Sociedade Brasileira de Computação
%J Porto Alegre
%S Proceedings
%K active learning, pattern recognition, automated diagnosis of intestinal parasites, microscopy image analysis, optimum-path forest classifiers.
%X We have developed an automated system for the diagnosis of intestinal parasites from optical microscopy images. Each exam produces about 2,000 images with hundreds of objects in each image for classification as one out of the 15 most common species of parasites or impurity. As the number of exams increases, a dataset with unlabeled samples for classification grows in size. Impurities are numerous and diverse, with similar features to several species of parasites. Some species are also difficult to be differentiated. In this context, datasets are large and unbalanced, making the identification of the best samples for expert supervision crucial for the design of an effective classifier. We have addressed the problem by proposing a new paradigm for active learning, in which the dataset can be a priori reduced and/or organized to make that process realistic (efficient) for user interaction and yet more effective. We have also proposed several active learning methods under this paradigm and evaluated them for the diagnosis of intestinal parasites and other applications. Data reduction and/or organization avoid to reprocess the large dataset at each learning iteration, enabling to halt sample selection after a desired number of samples per iteration, which yields interactive response times. The proposed methods were validated in comparison with state-of-the-art approaches. Experiments included three datasets with parasites and/or impurities. One with 1,944 parasites (without impurities) and another with almost 6,000 labeled objects were used to develop the methods. A more realistic one, with over 140,000 unlabeled objects, unbalanced classes, absence of classes, and considerably higher number of impurities, was used for final validation by an expert in Parasitology.
%@language en
%3 2015-wtd-sibgrapi-camera-ready-submitted.pdf