Detecção automática do coração em tomografia computadorizada utilizando técnicas de aprendizagem profunda

• Autor(a) principal: Azevedo, Jefferson da Silva Fernandes de
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/363
• topic: Coração; Faster R-CNN 3D; Detecção de objetos; Tomografia computadorizada; Redes neurais artificiais; CNPQ::CIENCIAS DA SAUDE::MEDICINA::CLINICA MEDICA::CARDIOLOGIA

Artificial intelligence algorithms have been a strong topic of study in the field of computer- aided medicine. The goal of such a technique is to reproduce tasks that, until now, were only achievable by human beings. The main advantage of these algorithms is to eliminate tedious hours of work performed exclusively by humans. In addition, in some approaches, it mitigates the bias that occurs in the marking between experts in the same database. More specifically, extracting only the region containing the heart in a CT scan facilitates many techniques that diagnose heart diseases. To do this, a Convolutional Neural Network algorithm dedicated to object detection can be used, which, in turn, has been manifested in several current scientific papers on two-dimensional images, the Faster R-CNN. This algorithm had an adaptation for 3 dimensions in the 2017 Multi-Modality Whole Heart Segmentation challenge. Therefore, in this dissertation, improvements and analysis were incremented in this presented algorithm. More precisely, in the steps responsible for detecting the 3D space region in which the heart is included. After the appropriate modifications and analysis of the hyperparameters, an accuracy of 74% was observed in the metric Intersection Over Union in the test of this model, as well as there are, on average, cuts in the region of the heart around 6% of its volume. Therefore, we conclude that the algorithm implemented and studied is able to detect the spatial region of the CT image containing the heart with acceptable accuracy, considering that the intention is to reduce the size of the image, highlighting only the essential, in order to facilitate the processing of diagnostic techniques.