Abordagem computacional baseada em deep learning para a classificação e segmentação de endometriose profunda no retossigmoide através de imagens de ressonância magnética

• Ano de defesa: 2025
• Autor(a) principal: FIGUEREDO, Weslley Kelson Ribeiro
• Orientador(a): SILVA, Aristófanes Corrêa
• Banca de defesa: SILVA, Aristófanes Corrêa , OLIVEIRA, Marco Aurelio Pinho de , PAIVA, Anselmo Cardoso de , CARVALHO FILHO, Antonio Oseas de , CUNHA, António Manuel Trigueiros da Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO (ASSOCIAÇÃO UFMA/UFPI)
• Departamento: DEPARTAMENTO DE ENGENHARIA DA ELETRICIDADE/CCET
• País: Brasil
• Palavras-chave em Português: endometriose; aprendizado profundo; classificação; segmentação; ressonância magnética.
• Palavras-chave em Inglês: endometriosis; deep learning; classification; segmentation; magnetic resonance images.
• Área do conhecimento CNPq: Processamento Gráfico; Ginecologia
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/5941

Endometriosis is a condition that primarily affects the pelvic region, significantly impacting the quality of life of women, especially those of reproductive age. This study proposes an automated method for classifying patients with endometriosis, localizing, and segmenting lesions in magnetic resonance images of the rectosigmoid region. The method aims to assist in diagnosis, reduce the need for invasive procedures, and map the extent of the lesions. It combines image processing techniques and deep learning, focusing on three main objectives: patient classification, lesion localization, and lesion segmentation. For patient classification, an F1-score of 94.74% was achieved using an ensemble of convolutional neural networks, including a proposed modification of the VGG-16 architecture. Lesion localization reached a sensitivity of 98.70% through an initial segmentation step and a region of interest extraction, employing a combination of image processing and the TransUNet neural network model. Lesion segmentation achieved a Dice score of 66.55%, also using TransUNet. Additionally, the study introduces a novel application of active learning for selecting training images to enhance model performance. These results demonstrate the potential of the proposed method to support the clinical management of endometriosis, particularly in diagnosis.