Publicação
Brain semantic segmentation: a deep learning approach in human and Rat MRI studies
| Resumo: | Magnetic Resonance Imaging (MRI) provides information about anatomy and pathology. This type of technique is the most popular used for the study of rat and human brain. Classifying voxels according to the presence of relevant anatomic features is an important step in the pre-processing of the data. A precise delineation and automatic segmentation of the brain structures is required in preclinical rodent imaging field and can substitute the manual segmentation where time consuming or human-error problems can occur. Current solutions are based on traditional segmentation algorithms that raise accuracy issues and generally need human intervention during or after the segmentation process. In the humans’ field, most of the tools created in DL (deep learning) are used in tumour or lesion segmentation. Brain segmentation tissues are not as explored as oncology problems and lesions complications. In the rats’ field, there are no segmentation studies in DL. It was decided to use a DL approach in Rats to solve some of the old techniques’ problems. This dissertation will present an approach on semantic segmentation of white matter and gray matter in Human’s images, evaluate the algorithm’s performance with outliers. It will also present an FCN (fully convolutional network) solution for on semantic segmentation using rat’s and human’s MRI of anatomical features. A two-dimensional convolution (slice-by-slice) approach and a three-dimensional (volume) convolutions approach were evaluated. At the end, the results found, using FCN U-NET in rats’ MRI for a 2D convolutions approach, DSC were 94.65 % for WM, 91.03% for GM and 76.89 % for cerebrospinal fluid. Using the 3D convolutions approach, the results using DSC found are 93.81 % for WM, 89.69 % for GM and 74.68 % for cerebrospinal fluid. The results using humans’ MRI using DSC were 91.59% for WM and 84,58% for GM. |
|---|---|
| Autores principais: | Rodrigues, Mariana Fontainhas |
| Assunto: | Engenharia e Tecnologia::Engenharia Médica |
| Ano: | 2018 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Magnetic Resonance Imaging (MRI) provides information about anatomy and pathology. This type of technique is the most popular used for the study of rat and human brain. Classifying voxels according to the presence of relevant anatomic features is an important step in the pre-processing of the data. A precise delineation and automatic segmentation of the brain structures is required in preclinical rodent imaging field and can substitute the manual segmentation where time consuming or human-error problems can occur. Current solutions are based on traditional segmentation algorithms that raise accuracy issues and generally need human intervention during or after the segmentation process. In the humans’ field, most of the tools created in DL (deep learning) are used in tumour or lesion segmentation. Brain segmentation tissues are not as explored as oncology problems and lesions complications. In the rats’ field, there are no segmentation studies in DL. It was decided to use a DL approach in Rats to solve some of the old techniques’ problems. This dissertation will present an approach on semantic segmentation of white matter and gray matter in Human’s images, evaluate the algorithm’s performance with outliers. It will also present an FCN (fully convolutional network) solution for on semantic segmentation using rat’s and human’s MRI of anatomical features. A two-dimensional convolution (slice-by-slice) approach and a three-dimensional (volume) convolutions approach were evaluated. At the end, the results found, using FCN U-NET in rats’ MRI for a 2D convolutions approach, DSC were 94.65 % for WM, 91.03% for GM and 76.89 % for cerebrospinal fluid. Using the 3D convolutions approach, the results using DSC found are 93.81 % for WM, 89.69 % for GM and 74.68 % for cerebrospinal fluid. The results using humans’ MRI using DSC were 91.59% for WM and 84,58% for GM. |
|---|