Deep Learning Techniques for Content-based Medical Image Retrieval

• Ano de defesa: 2022
• Autor(a) principal: Motta, Cezanne Alves Mendes
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado profundo; Busca por imagem baseada em conteúdo; CBIR; Content-based image retrieval; Deep Learning; Imagens médicas; Medical images; Variational autoencoder
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-15092022-105141/

In the context of Computer-Aided Diagnosis, it is often not enough for the system to produce correct predictions. When physicians are certain about the diagnosis of a given case, they can accept or disregard the systems prediction according to their own conclusions. However, in cases where they are uncertain, the physicians may not trust the system prediction without an explanation for it. In the medical domain, where the users are ethically and legally responsible for their decisions, the system should be able to articulate the reasons for its prediction in some way. One strategy that has been suggested to provide this support for decision is to retrieve images from similar cases that were already diagnosed. The physicians can then compare the retrieved cases to the one under consideration and decide if such diagnosis apply. Traditionally, Content-Based Medical Image Retrieval (CBMIR) has been done with hand-crafted features. Despite showing significant improvements in many other medical images analysis tasks, Deep Learning is not frequently used in CBMIR. Most current approaches to integrate Deep Learning into CBMIR use features obtained from models trained to classify images. These models tend to learn features that are correlated with the classes and ignore the ones that are not. Despite being useful to categorize images, the features learned from such models ignore intra-class variations that may be relevant to finding visually similar images. The ideal would be to retrieve the most visually similar case possible, not just one that belongs to the same class, so that the physicians can have more confidence in their decision. Autoencoders, on the other hand, are Deep Learning models that aim to learn features that describe the intrinsic factors of variations of a dataset. In this work, we investigate and discuss the use of Deep Learning for medical image retrieval, presenting the theoretical foundation and a critical analysis of current approaches found in literature. We also propose an approach for CBMIR based on Variational Autoencoders and show that this approach can yield better results than the ones based solely on classification and can even be used in combination to improve the results of the latter.