Multiclass Classification Application using SVM Kernel to Classify Chest X-ray Images Based on Nodule Location in Lung Zones

Abstract

Support Vector Machine (SVM) has long been known as an excellent approach for image classification. While many studies have reported on its achievement, yet it still weak to handle multiclass classification problem because it is originally designed as a binary classification technique. It is challenging task to transform SVM to solve multiclass problems like classifying chest X-ray images based on the lung zone location. Classified X-ray images improved image retrieval hence reducing time taken to assessed back the images. Realizing this difficulties, therefore, we proposed an application method for multiclass classification using SVM kernel to classify chest X-ray images based on nodule location in lung zones. The multiclass classification experiment is performed using four popular SVM kernels namely linear, polynomial, radial based function (RBF) and sigmoid. Overall, we obtained high classification accuracy (>90%) for three classifiers that are RBF, polynomial and linear kernel while sigmoid kernel classifier is only moderately good at 82.7% accuracy. Besides, values in the confusion matrices revealed that the RBF and polynomial classifiers managed to classify test data into all classification classes. Conversely, classifiers based on linear and sigmoid kernel have missed at least one classification class. Since each classifier work differently based on their kernel types, we noticed that it is better to view them as a complimentary rather than treating them as competing options. This condition also revealed that we can modify the original SVM classification method to handle multiclass classification problem.