Automated brain structure segmentation in magnetic resonance images of multiple sclerosis patients

• Autores: Alexandra González Villà
• Directores de la Tesis: Xavier Lladó Bardera (dir. tes.) , Arnau Oliver i Malagelada (codir. tes.)
• Lectura: En la Universitat de Girona ( España ) en 2019
• Idioma: español
• Tribunal Calificador de la Tesis: Petia Radeva Ivanova (presid.) , Jordi Freixenet Bosch (secret.) , Pierrick Coupé (voc.)
• Enlaces
  • Tesis en acceso abierto en: TDX
• Resumen

  • This thesis is focused on the automated segmentation of the brain structures in magnetic resonance images, applied to multiple sclerosis (MS) patients. This disease is characterized by the presence of demyelinating lesions in the brain, that appear as focal low signal intensity areas in the T1-weighted sequence, which is the most frequently used modality to segment the brain structures. In the first place, we exhaustively analyze the state of the art on this topic, presenting a new classification of the methods based on their segmentation strategy. We further discuss each category’s strengths and weaknesses and analyze its performance in segmenting different brain structures, providing a qualitative and quantitative comparison. From this first analysis, we observe that the vast majority of the reviewed methods are not designed for brains with lesions, such as those encountered in MS patients. Consequently, we also perform a thorough analysis of the effect of MS lesions on three representative state-of-the-art methods, each relying on a different category of the classification: FreeSurfer, FIRST and majority-vote label fusion. This second analysis reveals that the three segmentation approaches are indeed affected by the presence of these lesions, demonstrating that there exists a problem when using automatic methods as a tool to measure the disease progression. Therefore, based on the conclusions of these two studies, we propose a new correspondence search model able to minimize this problem on intensity-based multi-atlas label fusion strategies. Afterwards, we extend the theory of two remarkable label fusion strategies of the literature, i.e. Non-local Spatial STAPLE and Joint Label Fusion, in order to integrate our model into their corresponding estimation algorithms. Furthermore, with the aim of providing fully automated brain structure segmentation algorithms, a whole automated pipeline including lesion segmentation, pre-processing, atlas selection, masked registration and label fusion, is presented. Finally, a second extension of the theory to enable the integration of manual and automatic edits into the segmentation estimation of both strategies is also proposed. The evaluation, carried out in a quantitative and qualitative manner, includes a comparison of the proposed approaches to the original strategies when segmenting the raw images and the lesion-filled images, using both manual and automatically segmented lesion masks. The analysis of the results obtained with the proposed strategies points out a performance improvement on the lesion areas, which is also reflected on the whole brain segmentation performance.