Avaliação dos impactos da virtópsia associada a tecnologias de redes neurais convolucionais sobre métodos de autópsias convencionais

Abstract

Forensic sciences have always been present in a significant amount of historical and relevant facts in humanity. From its postmortem anatomical study prelude to the most sophisticated technologies, these have always sought to elucidate crimes and build a strong and consolidated chain of custody. However, there are still limitations in some of the forensic investigation methods, such as, for example, the autopsy technique, where there is human limitation with regard to the analysis of small-scale lesions. As a complement to traditional methods, it uses imaging techniques by ionizing radiation in its modality of computed tomography, for analysis of fractures in bones of the skull and face, in victims of violent deaths. This complementary technique can be optimized with the help of convolutional neural networks for deep image analysis. This research sought to develop a technique for analyzing the images produced in virtopsy to aid forensic diagnoses. The methodological outline of this work, initially, carried out the construction of a mass of data from PACS servers, with DICOM and JPEG images totaling 820 images. During data processing, artifacts and secondary noise were removed. The Orange data mining software Ⓡ developed by the University of Ljubljana located in Slovenia was used for training and testing neural networks. Neural networks with architecture of 16, 19 layers and also with the intercept architecture were used. After analyzing the results, it was identified that the intercept network obtained the highest accuracies, being 0.993 and 0.973 KNN and TREE respectively, demonstrating a promising result and wide potential for other applicability and new segments for this research and society