Comparative Study of Lung Image Representations for Automated Pneumonia Recognition

This paper presents a method for automatic pneumonia recognition by localizing and normalizing the position, rotation, and scale of lung regions in chest X-ray images. Likewise, we propose a classifier composed by: "Eigenfaces" method for feature reduction, Fisher’s discriminant criterion for feature selection, and a Multilayer Perceptron (MLP). We showed that a proper lung region alignment improves accuracy in several classifiers. Several CNN classifiers (MobileNetV2, ResNet-50, ResNet-18, Compact, and AlexNet), and our feature-discrimination-based classifier were tested with non-aligned images from a semantic segmentation stage, and on the other hand, with aligned images from our proposed normalization stage. Results show that the proposed normalization method improves MobileNetV2's accuracy to 97.1%, compared to 95.3% when using semantic segmentation. Furthermore, our proposed classifier based on feature discrimination increases its accuracy to 97.3% when using the normalization stage. Our classifier and normalization method achieve comparable or superior results to CNN classifiers.