Segmentação Automática de Endometriose Profunda em Imagens de Ressonância Magnética Baseada em Aprendizado Profundo

• Ano de defesa: 2025
• Autor(a) principal: PINTO, Daniel Moreira
• Orientador(a): SILVA, Aristófanes Corrêa
• Banca de defesa: SILVA, Aristófanes Corrêa , PAIVA, Anselmo Cardoso de , CRUZ, Luana Batista da
• Tipo de documento: Dissertação
• 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/CCET
• Departamento: DEPARTAMENTO DE ENGENHARIA DA ELETRICIDADE/CCET
• País: Brasil
• Palavras-chave em Português: Endometriose; Ressonância Magnética; Aprendizado Profundo; Processamento de Imagens; Segmentação
• Palavras-chave em Inglês: Endometriosis; Magnetic Resonance Imaging; Deep Learning; Image Processing; Segmentation
• Área do conhecimento CNPq: Bioengenharia
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/6071

Endometriosis is a disease characterized by the presence of endometrial tissue outside the uterine cavity. It is classified as deep endometriosis when the lesion extends more than 5 millimeters. This condition affects tens of millions of women of various age groups worldwide. It impacts patients’ quality of life and can present a wide range of symptoms, with the potential to progress into more serious complications. Therefore, effective diagnosis of the disease is crucial for initiating treatment. One of the primary areas where endometriosisrelated lesions occur is the rectum and sigmoid colon. Through magnetic resonance imaging (MRI), specialists can assess the presence of the disease non-invasively, but at a high time cost, as the lesions can appear in various shapes, sizes, and locations. Thus, this diagnostic method is extremely challenging. In this context, computational methods can assist specialists in diagnosis by providing an indication of the possible presence of lesions. However, the literature on deep learning-based segmentation of endometriosis in MRI is extremely limited. Therefore, this study proposes a method based on the application of Swin-Unet, a deep learning network based on transformers, for the segmentation of deep endometriosis in MRI images of the rectum and sigmoid colon. Additionally, a refinement step for the segmentation, based on image processing, is proposed to complement the method. The results achieved were an accuracy of 45.6%, sensitivity of 61.9%, Dice coefficient of 47.7%, and Jaccard index of 36.2%. At least one image per patient was segmented with good quality in 17 out of the 18 patients used for testing. Thus, the developed method shows promise in assisting the diagnosis of the disease.