Anotação semântica baseada em ontologia aplicada em imagens médicas

Abstract

Breast cancer is one of the most common types of cancer among women in Brazil and worldwide. Early diagnosis of a breast lesion is a priority for effective treatment at lower costs, with mammography being the most used exam. In this doctoral research, was developed the MUSA method to classify and semantically annotate mammography images based on the fusion of multimodal information, providing a fuller image annotation than the current state of the art. For this, the approach includes a text mining process, an image mining process, and an ontology engineering process. With the text mining process, it is possible to establish the contextual features based on similarity and co-occurrence relations between pairs of words in the same medical report. The image mining process allows classifying a mammography image using an end-to-end deep learning strategy. In the ontology engineering process, we built the AnotaMammo ontology, composed of the annotation ontology and the Mammo domain ontology. AnotaMammo allows the semantic annotation of mammography images. The text mining results were used to generate semantic rules, included in AnotaMammo. The results of the image mining process surpassed or compared with studies published in the literature, reaching more than 92% accuracy for classifying lesions as mass or calcification. The results also demonstrate that the AnotaMammo ontology adequately performed the semantic enrichment of the classification, in addition to having adequately performed the fusion of multimodal information. Finally, the MUSA method adds information to make the result more semantic and interpretable, thus reducing the semantic gap.