Attention Mechanisms in the Classification of Histological Images

• Autor(a) principal: Gomes, Mário Alexandre Neves
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/148925
• topic: Machine Learning; Attention Mechanisms; Deep Learning; Histopathological Images; Diagnosis; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Recently, there has been an increase in the number of medical exams prescribed by medical doctors, not only to diagnose but also to keep track of the evolution of pathologies. In this sense, one of the medical specialties where the mentioned increase in the prescription rate has been observed is oncology. In this regard, not only to efficiently diagnose but also to monitor the evolution of the mentioned diseases, CT (Computed Tomography) scans, MRIs (Magnetic Resonance Imaging), and Biopsies are imaging techniques commonly used. After the exams are performed and the results retrieved by the respective health professionals, their analysis and interpretation are mandatory. This process, carried out by medical experts, is usually a time-consuming and tiring task. In this sense and to reduce the workload of these experts and support decision making, the research community start proposing several computer-aided systems, whose primary goal is to efficiently distinguish between healthy images and tumoral ones. Despite the success achieved by these methodologies, it become evident that the distinction of the two mentioned image categories (healthy and not-healthy) was associated with small regions of the images, and therefore not all image regions were equally important for diagnostic purposes. In this line of thinking, attention mechanisms start being considered to highlight important regions and neglect unimportant ones, leading to more correct predictions. In this thesis, we aim to study the impact of such mechanisms in the extraction of features from histopathological images of the epithelium from the oral cavity. In order to access the quality of the generated features for diagnostic purposes, those features were used to distinguish healthy from cancerous histopathological images.