Learning Transformations for Automated Classification of Manifestation of Tuberculosis using Convolutional Neural Network

automated classification of tuberculosis in x-ray images is of an increasing interest to all researchers and physicians. Due to the high level of intensity inhomogeneity and variations, statistical machine-learning approaches usually fail to offer a generic solution to image classification. Convolution neural networks (CNNs) have demonstrated superior effectiveness in computer-aided diagnosis systems. Transfer learning can provide a powerful deep learning solutions to the limited availability of labelled images. In this paper we study the effect of knowledge transferred from a pre-trained ImageNet, in different ways via a pre-trained CNN model, to classify chest x-ray images as having manifestations of tuberculosis or as healthy. We evaluated and compared various models using the learning curve between training and validation set, and receiver operating characteristic (ROC) curve. Our experiments revealed that using fine-tuning technique outperformed both shallow-tuning and deep-tuning techniques and achieved 0.998 for the AUC, 0.999 for specificity, and 0.997 for sensitivity rate.