A Transfer Learning Approach to Leveraging Pre-trained CNN Models to Detect Tuberculosis using Chest X-Rays

Abstract

Pre-Trained Deep Learning models¹ are Convolutional Neural Networks that has been developed and bench marked for performing image classificationon a very large dataset such as ImageNet². These pre-trained networks have demonstrated their capability to work well even with images that are not part of the ImageNet² dataset through the transfer learning approach³. Pre-trained models help us with eliminating the cost and time involved in training a CNN model from scratch. In this paper we evaluate the performance of some ILSVRC²(The ImageNet Large Scale Visual Recognition Challenge) award winning pre-trained CNN models over the publically available Tuberculosis CXR datasets, we also appraise the ability of these pre-trained models to generalize for tuberculosis detection through CXR images. We customize the pre-trained model as per our requirements and fine tune its performance. The pre-trained CNN models that we have chosen to evaluate on the tuberculosis dataset as a part of this studyare:VGG-16⁴ ,VGG-19⁴,AlexNet⁵, ResNet-50⁶, Inception⁷ and DenseNet⁸.