Abstract: In classification, categories are typically treated as independent of one-another. In many problems, however, this neglects the natural relations that exist between categories, which are often dictated by an underlying biological or physical process. In this work, we propose novel formulations of the classification problem, aimed at reintroducing class relations into the training process. We demonstrate the benefit of these approaches for the classification of intravenous contrast enhancement phase in CT images, an important task in the medical imaging domain. First, we propose manual ways reintroduce knowledge about problem-specific interclass relations into the training process. Second, we propose a general approach to jointly learn categorical label representations that can implicitly encode natural interclass relations, alleviating the need for strong prior assumptions or knowledge. We show that these improvements are most significant for smaller training sets, typical in the medical imaging domain where access to large amounts of labelled data is often not trivial.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: both
Source Latex: zip
Primary Subject Area: Learning with Noisy Labels and Limited Data
Secondary Subject Area: Application: Radiology