Full text
| |
| Author(s): |
Fabio Augusto Menocci Cappabianco
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2010-02-24 |
| Examining board members: |
Alexandre Xavier Falcão;
Bruno Motta de Carvalho;
Luciano da Fontoura Costa;
Fernando Cendes;
Anderson de Rezende Rocha
|
| Advisor: | Alexandre Xavier Falcão; Guido Costa Souza de Araújo |
| Abstract | |
Segmentation of brain tissues from MR-images has become crucial to advance research, diagnosis and treatment in Neurology. Despite the large number of contributions, brain tissue segmentation is still a challenge, due to problems in image acquisition, large data sets, and anatomical variations caused by surgery, pathologies and differences in sex and age. Another difficulty is to create reliable ground truths for evaluation, which also requires suitable metrics. In this work, we review the most important pre-processing operations, as well as the most popular brain tissues segmentation methods. We also propose a new approach based on optimum-path forest clustering, which improves previous works on various aspects: speed, robustness, accuracy, intuitive tuning of parameters and adaptability to different imaging modalities and anatomies. The effectiveness of the approach can be noticed in both inhomogeneity correction and in white matter, gray matter and cerebral-spinal fluid segmentation. The method is evaluated quantitatively and qualitatively by taking into account two other popular methods, five datasets from diferent modalities, an operational range of parameters for each method and scores from distinct specialists. The results reveal a signiicant contribution to the state-of-the-art and emphasize the importance of suitable evaluation metrics in medical image analysis (AU) | |