Efficient Tuning of Hyper-parameters in Convolutional Neural Network for Classification of Tuberculosis Images

Deep Learning (DL) algorithms, especially Convolutional Neural Network (CNN) have outperformed in medical image classification tasks and have achieved human-competitive performance. This has become possible because CNN learns image features through backpropagation. However, the strategy for designing a CNN model with the highest accuracy for a specific application is often unclear. Because finding an appropriate network structure with the best combination of hyperparameters for different datasets is always a challenging task.  To address this, we propose an optimized CNN framework that automatically and efficiently tune its hyper-parameters using a hyperband search optimization approach. In this paper, an efficient CNN with optimized hyperparameters for the classification of tuberculosis disease in Chest X-Ray (CXR) images is trained and tested on a publicly available NLM-China dataset. The experimental results illustrate that the hyperparameters optimize the CNN framework and achieve 91.42% accuracy for the classification of tuberculosis disease in CXR images.