Optimized CNN-Based Approach for Tuberculosis Detection from Chest X-rays Under Limited Data Constraints

In the majority of undeveloped and underdeveloped nations in the globe, tuberculosis (TB) is one of the most serious diseases and a major public health problem. Chest radiography is one common way to identify tuberculosis. The expertise and interpretive abilities of radiologists who examine chest radiographs (referred to as chest X-rays), are essential to the effectiveness of this kind of screening. The advancement of computer assistance technology has contributed to the early identification of tuberculosis. Convolution Neural Networks (CNNs) have demonstrated a significant contribution to the fields of segmentation of images and feature extraction for abnormality detection. In order to classify chest X-rays (CXR) as either TB positive or negative, this work describes the design of CNN architecture with adequate pre-processing and optimized hyper-parameter tuning. Here we use five set of convolution layers and two set of dropout layers to design our proposed model. Three publicly accessible datasets: the Montgomery County chest X-ray set, the Shenzhen chest X-ray set, and the India chest X-ray set were utilized to evaluate our system’s viability. For combined dataset, we have achieved highest accuracy of 93.52%, and our method performs better and has a quicker detection rate, indicating an encouraging trend. Our proposed model outperforms other state-of-the-art CNN models in every assessment metrics. Another advantage of our method is that it is easy to adjust the weights when the system is updated with more datasets. It may also be easily installed on mobile devices and utilized in web-based applications because to its lightweight architecture.