Abstract / Summary
This article describes the design, training, and evaluation of a deep learning model based on aConvolutional Neural Network (CNN) for the automated classification of respiratory pathologies from chestradiographs. The objective of the research is to discriminate with high accuracy among three clinical conditions:COVID-19, pneumonia, and normal lung status. The methodology covers the preprocessing of a dataset ofmedical images in a standard format, incorporating Data Augmentation techniques to increase the robustness ofthe model against spatial variations. The architecture of the sequential network was designed with threefeature-extraction blocks (convolutional and max-pooling layers), followed by a dense classification stage withDropout regularization. To mitigate the effects of overfitting observed in the initial experimental phases, anEarly Stopping mechanism based on monitoring the validation loss was integrated. The experimental resultsdemonstrate the viability of the algorithm, reaching a validation accuracy above 94%, a metric that wascontrasted through the analysis of a confusion matrix to assess the distribution of false positives and falsenegatives among the different classes. It is concluded that optimized CNN architectures represent an effectiveand rigorous computational tool to support the differential medical diagnosis of respiratory infections.