MULTI-SOURCE CHEST X-RAY DATASETS FOR ACCURATE AND EXPLAINABLE TUBERCULOSIS (TB) DIAGNOSIS

Abstract

Tuberculosis (TB) is another cause of death, which relates to infectious diseases in the world. Chest X-ray (CXR) is an inexpensive and convenient screening tool for TB. Nevertheless, in the majority of cases, X-rays need to be interpreted manually, making them slow, subjective, and subject to inter- and intra-observer variability. In addition, most existing automated TB detection approaches are trained on a single data source, which limits their ability to generalise to other populations. In response to this, the current research paper proposes the multi- source CXR fusion methodology that will be more reliable and open the TB diagnosis. The publicly available CXR datasets are merged with various clinical settings and geographical areas in order to maximize the data in the datasets and reduce bias in the datasets. The preprocessing steps are followed by the merging of datasets, all of which are intensive preprocessing steps, including image normalization, lung part segmentation, contrast intensification, and class balancing, and all the sources must be consistent and compatible to facilitate fusion in the end. The feature extraction and classification problem is solved using a deep-based model using a convolutional neural network (CNN) as a pre-trained model, where it is possible to learn the discriminative TB-related patterns on the joint data automatically. Gradient-weighted Class Activation Mapping (Grad- CAM ) is an explainable Artificial Intelligence (XAI) technique that is applied to improve interpretability and facilitate clinical decision making. Such methods present graphically where the contribution of each part of the lungs in predicting the model is maximum, such that the clinicians may have a more detailed view of the diagnostic outcomes and may verify them. This approach is highly likely to be applied in real-life medical practice, e.g., the community with limited resources is where a TB screening is highly needed in a real-time context.