Impact of Transfer Learning on Chest X-Ray (CXR) Images for Tuberculosis Classification

Tuberculosis (TB) is an infectious lung disease and one of the leading causes of death worldwide. Various computer-aided diagnostic (CAD) systems have been proposed for the early detection of TB using different tools and techniques varying from conventional to deep learning-based approaches. The deep neural network (DNN) architectures are evolving and transfer learning models such as Inception-v3, ResNet50, and VGGNet that are pre-trained on the ImageNet dataset are available. Medical images are structurally different from generic images and a model trained on generic images can sometimes perform unexpectedly when utilized for disease diagnosis using medical images. In this paper, the impacts of multiple transfer learning models on the CXR datasets for TB detection are explored and compared with the proposed simple DNN architecture. In contrast with the complex transfer learning models, the proposed model has fewer layers, limited parameters, and simpler architecture with dropout and batch normalization to deal with model overfitting. The experimental results showed that the proposed model performs reasonably well when compared with the complex transfer learning models for TB detection and classification tasks. This is because the transfer learning models are over-parameterized and are trained for predicting a large number of class labels as compared to few possible labels in medical images. It is concluded that a simple model trained from the scratch on a medical image dataset can outperform the highly complex transfer learning models for TB detection. For the experimentation of the proposed model, the publicly available TB-specific CXR datasets including Shenzhen and Montgomery County CXR (MC) datasets are used.