Application of a Convolutional Neural Network using transfer learning for tuberculosis detection

Application of neural networks for image classification has significantly increased prediction accuracy rates. One area where neural networks could perform well includes classification of medical image datasets. Several convolutional neural networks (CNNs) such as VGG work by building a pretrained model that is easy to set up with minimal preprocessing. It uses libraries with weights containing millions of images to train the model prior to application on the actual data. This process is also called transfer learning. In this study we evaluate how CNNs could be a suitable alternative to decision tree based system for medical image classification. We apply CNNs on chest X-ray (CXR) dataset to identify if the patient has Tuberculosis (TB). Our goal is to develop a generalized model which takes care of all the complex preprocessing steps done by a conventional decision tree approach. Since CNNs have many hidden layers with filters, the model is able to achieve high accuracy of 80% without applying augmentation and 81.25 % with application of augmentation. Accuracies obtained are comparable to previous work done on the dataset. However, applying CNN bypasses the requirement of building sophisticated segmentation algorithms which could be time consuming, require professional expertise and are mostly specialized making them inadmissible for application to other similar problems.