Analysis of medical images to support decision-making in the musculoskeletal field

• Ano de defesa: 2021
• Autor(a) principal: Ramos, Jonathan da Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Fratura vertebral; Image segmentation; Magnetic resonance imaging; Ressonância magnética; Segmentação de imagens; Vertebral fragility fracture
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-17082021-102307/

Computer-aided diagnosis, computer-based image retrieval systems, and the radiomics approach are great allies to aid in decision making. However, an ordinary and laborious step in those approaches is the segmentation of a region of interest, for example, a vertebral body. Unfortunately, manually drawing accurate and precise boundaries is time-consuming and impractical to perform for many exams. Consequently, semi-automatic segmentation tools, with minimal interaction, pose high and attractive demand to the computational end. The greater goal is that, at some point, the physician intervention in the segmentation would be minimal with just a few or even no manual corrections. This doctorate research has the following hypothesis: The segmentation of vertebral bodies in MRI can be performed computationally faster with easier manual interaction and, at the same time, producing accurate results. We evaluate this hypothesis in three application scenarios as follows. First, we dealt with the challenging task of segmenting vertebral compression fractures in single MRI slices. We proposed Balanced Growth (BGrowth), which achieved 96.1% accuracy while keeping fast run-time performance. Second, we stepped into the segmentation of volumetric spine MRI exams, which is even more challenging due to several slices in the exams. We came up with a family of segmentation methods, presenting faster approaches with less manual interaction. Our final solution required annotating only two or three slices (among about 100 slices) and achieved 94% of F-Measure. To do so, we proposed the Estimation of ANnotation on Intermediary Slices (EANIS) along with the Fast Clever Segmentation (FastCleverSeg) method. Our approach was the fastest one and, at the same time, presented results similar to or better than the competitors. Finally, we assessed patients with bone fragility fractures using spine MRI and the radiomics approach. We proposed BonE Analysis Using Texture (BEAUT), which achieved 97% AUC for differentiating patients with and without vertebral body fragility fracture. Therefore, this doctorate research contributed to the state-of-the-art by introducing segmentation methods that presented promising results in the three scenarios mentioned above. We firmly believe that our contributions have a high potential to aid in the decision-making process and producing more ground truths for the training of deep learning models.