Inteligência artificial aplicada na quantificação de nódulos pulmonares em imagens de tomografia computadorizada

• Ano de defesa: 2025
• Autor(a) principal: Camargo, Thiago Fellipe Ortiz de
• Orientador(a): Calixto, Wesley Pacheco
• Banca de defesa: Calixto, Wesley Pacheco , Martins, Marcella Scoczynski Ribeiro , Siqueira, Hugo Valadares , Ribeiro, Guilherme Alberto Souza , Cruz Junior , Gelson da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/38995/001300000zzw1
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Engenharia Elétrica e da Computação (EMC)
• Departamento: Escola de Engenharia Elétrica, Mecânica e de Computação - EMC (RMG)
• País: Brasil
• Palavras-chave em Português: Nódulo pulmonar; Aprendizado profundo; Tomografia computadorizada; Quantificação; Segmentação semântica
• Palavras-chave em Inglês: Lung nodule; Deep learning; Computed tomography; Quantification; Semantic segmentation.
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://repositorio.bc.ufg.br/tede/handle/tede/14733

Accurate quantification of pulmonary nodules in computed tomography remains challenging due to interobserver variability and the lack of scalable methods capable of generalizing across heterogeneous datasets. This study proposes an automated solution that integrates deep learning to generate nodule segmentation masks and compute volumetric measurements using spatial information from the images. The methodology is structured into three main stages: data preparation and pre-processing, model training and validation, and performance evaluation. The use of the nnU-Net architecture, which automates pre-processing, segmentation, and post-processing, provides scalability and dynamic adaptation to the workflow, enhancing the clinical applicability of the solution. The results indicate consistent volume and diameter measurements across successive scans and strong agreement with consensus masks, even for anatomically complex nodules. The 3D U-Net architecture achieved a mean Dice coefficient of DC = 0.7846 with a standard deviation of σ = 0.18, outperforming the interobserver Dice index of DC = 0.5218 and exhibiting a low volumetric deviation between acquisitions. The proposed methodology advances automated quantification of pulmonary nodules, offering a resilient and adaptable solution to support medical diagnosis in real-world clinical scenarios.