This work aims to use convolutional neural networks and class activation mapping to perform tasks of detection and localization of infections resulting from pneumonia in chest radiographs of the RSNA Pneumonia Detection Challenge dataset. In this way, following the two-step training methodology (FRID-ADAR, M. et al. 2021), two convolutional neural network architectures, ResNet50 and EfficientNet B2, were implemented using the TensorFlow and Keras frameworks in the Python programming language. The results obtained by the two architectures were compared using the metrics Accuracy (for the detection task), Dice Similarity Coefficient and Intersection over the Union (for the localization task). The results obtained by ResNet50 in the test set were an accuracy, Dice and IoU of 0.86079, 0.7860 and 0.6475, respectively; and the results obtained by EfficientNet B2 in accuracy, Dice and IoU were 0.94744, 0.5748 and 0.4033, respectively.
|
