Investigação de redes convolucionais para classificação de imagens de radiografia do tórax

Abstract

The identification of diseases in the human body is often challenging, as many of them remain invisible to the naked eye. In such situations, the disease may remain latent in the organism, making it difficult to identify the problem. In this context, X-ray exams play a fundamental role in identifying and diagnosing these issues, allowing for the internal visualization of the body and the assessment of organ integrity. Chest X-rays, in particular, are one of the most common and effective exams in this regard. Through this exam, it is possible to identify anomalies in the size or shape of the heart, disor ders in the veins and arteries, as well as problems in the lungs, such as pneumonia and other respiratory diseases. The work in question investigates the relevance of classifying images of chest X-rays amidst the growing use of convolutional neural networks (CNNs), while also explaining the changes that have been made over the years that have allowed CNNs to become famous not only for image classification but also for object detection and segmentation. In this context, transformer architectures, particularly vision transfor mers (ViTs), emerge as dominant models. The main goal is to understand and evaluate the performance innovations of a hierarchical visual transformer called swin transformer. To do this, we introduce some of these modifications to classic CNNs, such as VGG16 and ResNet-50, and train these models using the CheXpert database, evaluating their accuracy. Through a series of sequential adjustments, keeping only the modifications that demonstrated a positive impact on accuracy, we achieved a precision of 86.15% for VGG16 and 80.05% for ResNet-50, while the swin transformer obtained 78.75%. These results suggest that, despite the robustness of the swin transformer, certain adaptations in CNNs can result in superior performance, highlighting the importance of exploring different architectural approaches in the context of medical image classification.