An ensemble transfer learning approach for automated tuberculosis diagnosis from chest x-rays using optimized feature selection and XGBoost classification

Tuberculosis (TB) continues to pose a major global health challenge, which demands reliable and efficient diagnostic methods. The chest X-rays (CXR) are clinically used for TB detection; however, the visual similarity of TB with other lung diseases makes it challenging for radiologists to identify TB at early stages. Using medical imaging techniques, such as chest X-ray (CXR) images, and computer-aided deep learning (DL) can help automate the early diagnosis of TB. This paper presents a deep ensemble transfer learning framework that integrates ResNet101 and Xception for feature extraction (ERX-Net), followed by iterative ReliefF-based feature selection to preserve the most pertinent attributes, aimed at detecting TB using DL algorithms. Deep features are extracted using two pre-trained DL networks: ResNet101 and Xception. To leverage features for identifying TB CXR images, the iterative ReliefF method is used to select the most relevant attributes. The refined features are classified using an optimized Extreme Gradient Boosting (XGBoost) ML classifier algorithm to identify normal and TB CXR images. The proposed model, evaluated on publicly available TB CXR datasets, achieves 98.07% accuracy, 98% specificity, and 98.14% sensitivity, outperforming state-of-the-art DL and ML approaches. The proposed framework delivers high diagnostic precision while maintaining computational efficiency. This framework has the potential to serve as a reliable computer-aided screening tool for TB, particularly valuable in resource-constrained healthcare environments.