Detection of Tuberculosis in Radiographs using Deep Learning-based Ensemble Methods

Diagnosis of Tuberculosis has witnessed a substantial improvement owing to the infusion of technology in the form of deep learning classification. The transfer learning techniques implemented to date have delivered acceptable results only for academic research. For real-time implementation, the model needs to be generalized and trained on a larger dataset with diverse features. The results can be improved by using combinations of pre-trained models. This paper presents the implementation of Ensemble methods using Stacked Generalization and Weighted Ensemble (Sum of probabilities and Voting Majority) methods using an amalgamation of VGGNet, DenseNet and EfficientNet models, which yield higher accuracies of 95.19% and a significant reduction in Type I and Type II errors. The data-centric approach has been followed to enhance the data quality and volume while the hyperparameters were deemed consistent in the models. A verified and validated approach of model development and data improvisation has resulted in improved performance of the TB classification.