Segmentação semântica do hipocampo em imagens de ressonância magnética estrutural utilizando rede neural convolucional

Abstract

The main purpose of this paper is to develop an artificial intelligence algorithm capable of segmenting the hippocampus in medical brain images. The hippocampus is a brain structure connected to the memories obtention, therefore its characteristics are connected to the presence or not of Alzheimer’s Disease. The region of interest can be identified on Magnetic Resonance Imaging by specialized medical professionals. Using an open database with 263 files with masks was possible to separate 6627 images that had the hippocampus, of which 90% were used to teach a convolutional neural network to identify and segment the hippocampus in each of the images. The outcome of this CNN was satisfactory with an accuracy higher than 99%, showing that the use of this segmentation for medical applications to aid in the diagnosis of AD is very promising. The main difficulty encountered during this project was obtaining and organizing the database to be used, different types of images were considered and even the most appropriate required intervention for processes such as zero-padding and binarization.