Inteligência artificial aplicada na detecção de pneumonia em imagens de raio-x do tórax

Abstract

INTRODUCTION: There are fewer specialized professionals than the available demand in the market; this is the case in many fields of work, and it is no exception when this refers to radiologists. Due to the large volume of exams to be analyzed, radiologists are considered as a bottleneck in the flow of imaging exams released. When it comes to diagnostic imaging in a hospital emergency setting; we can point out two major problems. The first is the non-radiologist physician who does not have the enough expertise to provide a diagnosis from an image. The problem is the order in which these images are analyzed; often more serious problems, which should be analyzed with priority, are analyzed later due to the order of arrival and the order in which the exam was performed. OBJECTIVES: Development and application of an Artificial Intelligence system capable of identifying pneumonia in chest X-ray images, as a proposal to assist physicians in screening and imaging diagnosis, contributing to the speed and assertiveness of the diagnosis. METHODS: A cross-sectional technological assessment study with a descriptive design; applied nature through experimental procedures, with a quantitative approach of analytical type. The artificial intelligence system in this research was developed by the researcher himself, following proper software development practices. The evaluation of the system performance in the diagnosis of pneumonia in X-ray images was performed using the most widely used performance evaluation models in AI and Machine Learning systems. Among the performance metrics used in the system evaluation we can mention the ROC curve, F1 score, precision, recall and accuracy. The system performance evaluation was carried out using the set of images available for the research, separated into training set and validation set. RESULTS: The final architecture of the algorithm had a total of 27 layers. The Algorithm was trained using 2,290 images and 1,430 images were used to evaluate its performance. Initially 40 training rounds were set, but it was stopped in round 37 for stabilizing learning and not achieving any further improvement in accuracy. The Algorithm achieved in the validation set an area under the curve of 0.68. The algorithm was made available in a web system capable of receiving a file in DICOM format, processing the image and returning the probability of the image being positive for pneumonia. CONCLUSIONS: The artificial intelligence system for identifying pneumonia in chest X-ray images has the ability to improve the efficiency of imaging diagnoses; being a tool to assist the physician at the time of screening and prioritization of exams to be analyzed, as well as a second opinion for the diagnosis of of these. In this way, the system can improve the speed in releasing important results and at the same time contributes to increase assertiveness in diagnoses.