Explorando técnicas de aprendizado profundo no diagnóstico de doenças infecciosas

Abstract

The World Health Organization (WHO) defines infectious diseases as those caused by pathogenic microorganisms such as bacteria, viruses, parasites, and fungi. These diseases can have high mortality rates and serious global repercussions due to the rapid and unexpected nature of their spread. One example is malaria, a potentially fatal infectious disease caused by protozoa transmitted to people through the bite of female mosquitoes of the genus Anopheles. Another example is COVID-19, caused by the SARS-CoV-2 coronavirus, which also spread rapidly and unexpectedly, creating a global health crisis. This study proposes automated methods for detecting malaria parasites in optical microscopy images and segmenting lung areas affected by COVID-19 in medical tomography images. For malaria, a method was developed for microscopic images of thick blood smear, utilizing pixel classifiers and a 34-layer deep neural network. The methodology also uses HSV components of the image as input variables for multilayer perceptron and decision tree classifiers. The dataset was segmented into different groups based on parasite sizes, allowing the evaluation of performance metrics such as accuracy, sensitivity, specificity, precision, and F1-score. The best results were obtained in images with larger parasites, with accuracy rates of 91.71% and 93.14%. For images with smaller parasites, accuracy rates were lower, reaching 76.58% and 71.58%. The results are comparable to those of previously published studies. Regarding COVID-19, a method was developed to be used with computed tomography (CT) images, and three automatic segmentation approaches were compared: 1) segmentation of affected areas using two data partitioning strategies (examination and slice strategies); 2) automatic segmentation and interobserver agreement between two groups of radiologists; and 3) segmentation performance using both manual and automatic lung segmentation methods. Two datasets and two convolutional neural network architectures (U-Net and CNN2) were used. The Dice metrics for automatic COVID-19 segmentation were 73.01±9.47% and 84.66±5.41% for the examination and slice strategies, respectively. For manual lung segmentation, the Dice metrics were 74.47±9.94% and 85.35±5.41%, respectively. The slice strategy showed slightly better performance than the examination strategy. No statistically significant differences were observed between automatic segmentation and interobserver agreement in a group of seven CT scans. The results highlight the effectiveness of the proposed methods, showing good performance in evaluation metrics.