Identificação automática e caracterização de lesões em substância branca no cérebro em imagens volumétricas de ressonância magnética

Author(s)

Mariana Pinheiro Bento

• Document type: Doctoral Thesis
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Elétrica e de Computação
• Defense date: 2016-12-21

Examining board members

Leticia Rittner; Ana Maria Marques da Silva; Paulo Manzzoncini de Azevedo Marques; Romis Ribeiro de Faissol Attux; Gabriela Castellano

• Advisor: Leticia Rittner

Abstract

Lesions in the brain white matter can cause significant functional deficit, and are often related to psychiatric and neurological diseases. The analysis of these lesions is manually performed by specialists on magnetic resonance images, and represents a non-trivial, time-consuming and subjective task. This work aims to develop and to validate methods to perform the identification and characterization of lesions in the brain white matter, helping the specialists in diagnosis and follow-up of patients over time based on structural metrics extracted from identified lesions using T2-weighted magnetic resonance images. The identification aims to distinguish normal from lesioned regions, regardless of lesion location, size or etiology. The characterization aims to analyze lesions according to their etiology, ischemic or demyelinating. The developed methods combine techniques from image processing and pattern recognition, such as image quantization, normalization, two dimensional and three-dimensional texture analysis (statistics from histogram, co-occurrence matrix, run-length matrix, local binary pattern and gradients), mathematical morphology, feature selection and supervised classifiers, such as support vector machines, k-nearest neighbors and linear discriminant analysis. This research project also aims to construct a dataset containing images acquired on multiple research centers, presenting lesions with different etiologies, i.e. ischemic and demyelinating, with varying shape and location. Experiments performed in this dataset evaluated not only the accuracy and efficiency of the developed methods, but also their robustness to process images with different acquisition parameters, and lesions with varying characteristics, observed in patients with different diagnosis, such as Multiple Sclerosis, Cerebrovascular Accident, Systemic Lupus Erythematosus and Systemic Sclerosis. The developed method present requirements that make possible their usage in a clinical environment: minimal or no user interaction and competitive results, not limited to a specific disease as the other methods presented in the literature. The proposed classification method achieved accuracy rates higher than 90% to distinguish lesions according to their etiology, and the proposed method to perform automatic lesion segmentation presented an average Dice coefficient of 0.7 in the analysis of the multicenter dataset (AU)