Luiz Henrique Cartaxo Fernandes et al. José Lucas Formiga Dantas Marilia Graziela Vieira de Macena Lima Glaudir Donato Marcelo Dantas Tavares de Melo
Abstract
Neural networks are computer models that mimic the workings of the human brain, learning from large volumes of data to perform increasingly complex tasks. This is done by means of interconnected artificial neurons that propagate information from raw layers to refined layers, sometimes of a weighted nature, weaving in the classification or identification of non-linear patterns. The Convolutional Neural Network (CNN) is a specialized deep learning architecture designed to extract and analyze spatial patterns within images, enabling applications in diagnostic decision-making and longitudinal clinical monitoring. Its use in Computed Tomography and Magnetic Resonance Imaging through cardiac image segmentation makes it possible to quickly and efficiently define cardiac structures, playing an important role in the reconstruction of three-dimensional images and in supporting preoperative procedures. In addition, compared to other forms of conventional analysis, it guarantees objective interpretations and superior image quality, based on the removal of artifacts from the studied segment. Echocardiograms have shown good results in identifying pathologies such as pulmonary hypertension, amyloidosis, hypertrophic cardiomyopathy and congenital diseases, and studies on the automation of these exams have shown an opening for what could be a dissemination framework in more remote areas where manual operators are not available. Consequently, CNN-based approaches have the potential to streamline cardiovascular imaging workflows, minimize inter-observer variability, and expand diagnostic capabilities to underserved and remote regions with limited access to specialized care.