TUBERCULOSIS PROGNOSIS THROUGH RADIOGRAPHIC PREDICTIVE MODELING

Abstract

In the field of lung disease classification, the fusion of diverse datasets shows significant challenges and opportunities for improving predictive accuracy. This study showcases the efficacy of various deep-learning models in classifying pulmonary tuberculosis using a combined dataset of CX-Rays sourced from diverse resources. Leveraging datasets from multiple origins and combining them, we performed classification on state-of-the-art ConvNets, including but not limited to, ResNet50, DenseNet121, EfficientNetB0 to B7. Through rigorous experimentation and analysis, we identified the model with the highest performance on the combined dataset. We identify the model with the best performance on the pre-processed dataset with the help of precise experimentation and analysis. Additionally, to strengthen the classification results, additional layers and optimization techniques were applied into the best model. These enhancements ranged from feature extraction with transfer learning to fine tuning, aimed at improving the model's predictive capabilities. By thoroughly comparing the performance of different convolutional models and techniques, we emphasize insights into the most effective approaches for pulmonary tuberculosis classification. Our research illustrated the addition of traditional ConvNet Classification models with extra layers and techniques yields notable improvements in model evaluations. The outcomes of this research have shown how different models work on chest X-ray images and how they can be used in decision making in healthcare and for the diagnosis of lung diseases.