Estratégias de classificação de imagens radiológicas utilizando redes neurais convolucionais e transformada Wavelet

Abstract

The outbreak of the COVID-19 pandemic has motivated massive worldwide efforts to tackle the problem, such as the simplification of protocols for accessing data repositories and metadata of the new virus and the disease. Searches for research suggest in several areas from biochemical, biological research, serological investigation to genetic engineering and information technology. In the artificial intelligence filed, deep learning techniques were used in search of support tools that would contribute to facing the pandemic. The possibility of reducing errors in the analysis of chest radiological images was a goal of such research, as they complemented the medical examination of the disease. Thus, the pandemic was the first motivation for the creation of the WCNN model and the second motivation was the observation of the use of techniques for resizing medical images to adapt to ready-made models in the literature, which can cause distortions or loss of information in the detection. of the dis-ease under study. WCNN was based on a Convolutional Neural Network (CNN) and wavelet transform. The model proposes a custom input layer, called Wave Layer, which processes the images without resizing them. To assess WCNN, an experiment was performed that exemplifies its behavior, using a set of chest CT images from patients diagnosed with COVID-19 and other lung infections. The result of the metrics Accuracy (ACC), Sensitivity (Sen) and Specificity (Sp) were 0.9819, 0.9783 and 0.98, respectively. Hence, it can be concluded that these expressive results indicate that the association of CNNs and wavelet transforms is promising for the creation of classification models.