Deep Learning-Based Detection of Tuberculosis Using a Gaussian Chest X-Ray Image Filter as a Software Lens

Tuberculosis remains one of the most prevalent and lethal infectious diseases, with millions of cases reported each year. Convolutional neural networks have proven effective in detecting such diseases from medical images, achieving high accuracy in identifying tuberculosis from chest X-rays. However, many models are limited by small datasets, lack of cross-validation or have not achieved an optimal level of detection performance. In the context of diagnosing diseases, it is crucial to continually strive for increasingly accurate and robust solutions. This study focuses on the distinct characteristics of tuberculosis lesions, such as their large structures and gradual transitions between healthy and infected tissue. We propose that optimal detection performance may not rely on more complex architectures but instead on optimizing preprocessing techniques to highlight these features. Specifically, a ResNet50-based architecture with Gaussian filtering was evaluated on a dataset of 7,000 images using stratified 5-fold cross-validation. The results show an average accuracy of 99.2%, outperforming the unfiltered model (97.7%) and literature (99.05%), thus setting a new benchmark. The findings demonstrate that leveraging tuberculosis-specific features through Gaussian filtering provides an effective approach to enhance diagnostic performance.