Segmentação automática de tecidos cerebrais em imagens de ressonância magnética do tipo fluid-attenuated inversion recovery

Abstract

This thesis proposes a method for brain tissue segmentation on Magnetic Resonance Images of type Fluid Attenuated Inversion Recovery (FLAIR), among White Matter, Gray Matter and Cerebrospinal Fluid. Images of type FLAIR are important for diagnosis and control of diseases such as Multiple Sclerosis or Systemic Lupus Erythematosus, for they show White Matter Lesions, which are characteristic of such diseases, as hyperintense areas. However, any brain area can display hyperintensities, requiring tissue segmentation to confirm the position of lesions. However, this image kind presents low contrast between White matter and Gray matter, which makes segmentation difficult. The T1-weighted modality is the most used one for this operation, as it presents good contrast between brain tissue types. For diseases such as Multiple Sclerosis, the T1 modality can be non-essential from a clinical perspective, representing an extra cost. No published works were found on tissue segmentation directly in images of type FLAIR. The methodology proposed in this thesis uses an Artificial Neural Network, trained with a dataset generated from train FLAIR images, for which either the tissue segmentation is available, or it was obtained from the T1 modality. The classification model is then used to segment tissues in other FLAIR images. The methodology uses both features which are commonly found in literature, as well as new features, proposed in this thesis. The results are promising, being comparable to results of other published works which segment brain tissues using the T1 modality.