Processamento e classificação de imagens estereoscópicas cerebrais usando machine learning

Abstract

Stereoscopy is a technique which allows the representation of three-dimensional space through two-dimensional images. This effect is achieved by superimposing two images of the same object, known as a stereoscopic pair, taken with a subtle horizontal camera movement while maintaining the same focal point. The depth effect produced by stereoscopy can be useful in teaching neuroanatomy and in training neurosurgeons, because the three-dimensional understanding of body structures is important in these contexts. Improperly produced stereoscopic images can cause side effects in the viewer, such as dizziness, pain, and eyestrain. In order to evaluate the production quality of a stereoscopic image, a computational algorithm can be employed, avoiding the occurrence of these symptoms in the observer. To elaborate such an algorithm, feature extraction processes are required, followed by the implementation of Machine Learning techniques. In this work, Pearson’s Correlation Coeffcient, the angle between vectors, and the ORB algorithm, Oriented FAST and Rotated BRIEF, were employed in extracting class discriminant parameters from stereoscopic pairs. Six parameters were chosen according to Fisher’s Discriminant Ratio. From these parameters, three computational algorithms were implemented based on machine learning techniques: Single Variable Classifer, Least Squares Classifer and Naive-Bayes Classifer. After the analysis of accuracy, precision, recall, specifcity and runtimes, the Naive-Bayes Classifer was indicated as the most suitable for the classifcation of the stereoscopic images.