Aprimoramento de um modelo iterativo de segmentação de lesões de esclerose múltipla em imagens de ressonância magnética

Abstract

Multiple Sclerosis (MS) is a cronical, inflammatory and demyelinating disease that impairs the central nervous system (CNS). MS is possibly autoimmune and mainly affects the young adult population. Although its diagnosis is primarily clinical, magnetic resonance imaging (MRI) has proved to be a very important subsidiary tool in confirming the diagnosis and assessing the disease’s progression and treatment, since it allows highlighting the neurological changes in space and time. As manual (or semi-automatic) delineation of MS lesions in MR images is time-consuming and prone to intra and interobserver variabilities, segmentation techniques have been proposed to assist in the segmentation and volumetric measurement of MS lesions using tridimensional (3D) MR images, among which can be mentioned the iterative Student’s t Mixture Model (iStMM), developed inside the Biomedical Image Processing (BIP) Group at the Computing Department of Federal University of São Carlos - FAPESP project number 2016/15661-0. In spite of producing results comparable to those in current scientific literature, the method has two limitations that were the subject of study of this work: (a) the lack of a stopping criterion for the model, which currently uses a fixed number of iterations; (b) the lack of a criterion for automatically selecting the number of clusters, which currently considers the same number of possible clusters at all iterations. Thus, this work improved the aforementioned limitations by investigating the use of texture patterns of lesion masks, obtained at each iteration of the method, to define a stopping criterion for the algorithm. In addition, the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) were analyzed as a way to automatically determine the number of clusters for the mixture model, also at each iteration. At last, a microservices-based web platform was developed as means to allow the execution of this algorithm (and also any other technique that may be included) by authenticated users.