A Novel Feature Extraction and Selection Technique for Chest X-ray Image View Classification

Chest X-ray (CXR) image view information can play an important role to make a computer-aided diagnosis (CAD) system more superior and robust. Moreover, if the image header file provides the view label, we can apply specialized techniques for different positional views. Since this type of information is not always available with the image, a view classifier is essential. In this paper, we present a method to classify the two most common CXR projection views - posteroanterior (PA) and anteroposterior (AP). We developed the method specifically with four key components - pre-processing, feature extraction, feature selection, and classification. The features we extracted are histogram of oriented gradients, image projection profile and our developed feature named 'CXF30'. We generated this new feature from the signal properties of vertical image profile feature. For obtaining an optimal subset of features, we applied a genetic algorithm-based feature selection technique. Finally, for classification purpose, we trained a random forest machine learning model with the selected features. We evaluated the method on a large dataset consists of 24,000 chest X-ray images, and consequently, it achieved a high 10-fold cross-validation accuracy (above 96%). This result indicates the effectiveness of the presented method. In addition, through different experiments, we justified the quality of our extracted features and also demonstrated how feature selection process improved the performance of the model.