Experiments with Dimensionality Reduction on Chest X-Ray Images

Abstract

In this work, we study the problem of chest X-ray image classification. We use well-known convolutional neural network architectures (ResNet18, ResNet50, VGG19, MobileNet, ShuffleNet, and EfficientNetB0) trained on chest X-ray images, to extract feature vectors for the images in the dataset. The classification process is mainly performed with Support Vector Machines (SVM) and Random Forest (RF) methods. The central focus of this work is to study the influence of dimensionality reduction on the classification results. First, we used Principal Component Analysis (PCA) and reduced the length of the feature vectors. Secondly, we perform a selection process of the images in each class, from the training set. This selection procedure aims to preserve in the training set the images that provide better characteristics of the class they belong to. The selection process uses clustering methods (k-means, hierarchical clustering, Self-Organizing Maps). For the numerical experiments we used RSNA (Radiological Society of North America) dataset with images divided into three classes. The best results were obtained using Random Forest classification method applied on features extracted with ResNet50 and by using a reduction of the dataset based on Ward agglomerative clustering technique.