Divergências de Bregman aplicadas na recuperação de imagens por conteúdo em displasias orais

Abstract

CBIR systems aim perform image retrieval based on their visual content, using algorithms that describe visual characteristics in the form of numerical features. These features allow finding and displaying similar images, seeking to replicate or surpass human judgment about similarity. The challenge of dealing with unstructured data in large sets of images creates opportunities to develop techniques that identify similarities or dissimilarities between images, depending on the comparison method applied. Most of the work dedicated to the development of structures and algorithms that apply similarity measures uses metric spaces, because metric measures satisfy the basic properties of the metric space. For cases that do not follow these properties, using non-metric similarity measures contributes to the solution of problems in more complex types of data. Bregman Divergence allows the development of effective and applicable methods that use a set of inequality functions, instead of relying on a single objective. This can be successful when metric features do not match human judgment. Obtaining superior results lies in the ability to choose the measure that best fits the problem in question, being possible to use generalizations such as: Mahalanobis and Kullback-Leibler (KL). In this study, the objective is to investigate the similarity or dissimilarity of images related to oral cavity cancer, where features were extracted from slides containing experimentally induced lesions in the tongue of mice of the C57Bl/6 strain, using the carcinogen 4NQO. The use of medical images and image retrieval plays a crucial role in the diagnosis of dysplasias, especially in fields such as radiology, pathology and medicine. For the study, a comparison between metric distances and Bregman Divergences (KL and Mahalanobis) was performed, using vectors of morphological and non-morphological features. The results of the CBIR system using different distance functions were evaluated by the measures of Precision and Recall. Through the results, Bregman Divergences, especially KL, demonstrate better performance compared to metric distances in most of the tested cases.