Avaliação de Desempenho de Redes Neurais Convolucionais no Diagnóstico de Amiloidose Cardíaca

Abstract

Rare diseases, affecting up to 65 people per 100,000, are largely caused by genetic mutations, often chronic and typically incurable, though man ageable. Cardiac amyloidosis (CA), one of these diseases, is marked by the buildup of proteins in heart tissues, leading to stiffness and impaired heart func tion. Due to its severity and high mortality rate, early diagnosis is crucial. This study aims to evaluate the use of convolutional neural networks (CNNs) to identify this rare disease in a data-scarce context. Five CNN architectures were employed — ResNet-50, VGG16, Xception, MobileNet, and InceptionV3 —trained with a dataset of 138 transthoracic echocardiograms (TTEs). To in terpret the network performance, Grad-CAM was applied to generate heatmaps highlighting the most relevant TTE areas for classification. Among the results, MobileNet achieved 75.14% accuracy with a ROC curve of 0.87, while ResNet 50 reached 81.43% accuracy and a ROC curve of 0.93, demonstrating the best performance. In conclusion, despite data limitations, ResNet-50 showed the highest efficacy and generalization capability among the networks evaluated.