Deep Learning for Tuberculosis Detection From Chest X-Rays

Abstract

This thesis explores detection of tuberculosis using deep learning models. A comparison is made between a shallow convolutional neural network (CNN) with four pre-trained models. The comparison is done across a small and large dataset to observe the differences in classification performance and computational efficiency. Segmentation, data augmentation, dataset imbalance handling are investigated. Finally, the performance of the best models is validated using Grad-CAM and further tested on a dataset with a different distribution.