Artificial Intelligence-based Deep Learning Architecture for Tuberculosis Detection

Artificial Intelligence-based system for Tuberculosis (TB) detection has been proposed in this work. Manual diagnostic through radiologists can be misdiagnosis due to human error. Designing an artificial intelligence-based decision support system can help in the accurate prediction of TB through lung Chest X-ray (CXR) image. In the proposed work, data centric and model centric approaches are applied. In data centric approach, images play an important role in feature extraction. In this context, experimentation on six varied kinds of image enhancement applied to the publicly available TB image dataset along with data pre-processing and data augmentation. In model centric approach, three pre-trained Convolutional Neural Network (CNN) with modification with specialized layers are proposed. Comparative Study of six image enhancement techniques along with original dataset in behavior response with network architectures is evaluated for best performance. Evaluation of the networks is based on accuracy, precision, recall and AUC. Sharpening of images applied on Modified ResNet50 is the best performer with accuracy 99.05%, precision 98.87%, recall 100% and AUC of 99.29%. In comparability with original dataset, sharpened images performed 0.8% better in metric evaluation, which gives better classification approach and predictability.