AI-Based Tuberculosis Detection Using Chest Radiographs and Transfer Learning

Abstract

Early and accurate diagnosis of tuberculosis (TB) is essential to stop its transmission, especially in underdeveloped countries. A limitation of traditional diagnostic tests is that they typically require hours of assessment; they are also prone to human error in interpretation and cannot be scaled up in limited resource settings. This research proposes a machine learning-based (ML) system for automatic and efficient diagnosis of TB in CXRs, addressing these limitations. The proposed technique enhances the image quality when applied to the TBX11K database and the pre-processing consists of conversion to greyscale, picture scaling, Contrast Limited Adaptive Histogram Equalization (CLAHE) and normalization. Supplementing data with synthetic minority In order to rectify class imbalance and enhance dataset variety, SMOTE is utilized. After processing, images were categorized by Extreme Gradient Boosting (XGBoost) method. The data is divided into three sets: training, validation, and testing sets, so as to develop a complete model and test the effectiveness of the model. The suggested framework is proven effective in the experiments, which achieved an F1-score of 99.60%, an ACC of 99.71%, a PRE of 99.70%, and a REC of 99.74%. Analyses comparing the suggested method to preexisting ML and DL models have proven its higher classification performance. Automatic TB screening and early detection utilizing chest radiographs is now possible with the help of the created framework, according to the results. It's dependable, efficient, and accurate.