A Semantic Segmentation-Based Adaptive Convolution Neural Network Model for Tuberculosis Detection and Diagnosis

Abstract

Objective: Tuberculosis (TB) continues to be a major cause of death from infectious diseases globally. Tuberculosis (TB) is a treatable condition with antibiotics, yet it is often misdiagnosed or left untreated, particularly in rural and resource-constrained regions. While chest X-rays are a key tool in TB diagnosis, their effectiveness is hindered by the variability in radiological presentations and the lack of trained radiologists in high-prevalence areas. Deep learning-based imaging techniques offer a promising approach to computer-aided diagnosis (CAD) for tuberculosis (TB), enabling precise and timely detection while alleviating the burden on healthcare professionals. This study aims to enhance tuberculosis detection in chest X-ray images by developing deep learning models. We leverage the Res-UNet architecture for image segmentation and introduce a novel deep learning network for classification, targeting improved accuracy and precision in diagnostic performance.

Methods: A Res-UNet segmentation model was trained using 704 chest X-ray images sourced from the Montgomery County and Shenzhen Hospital datasets. Following training, the model was applied to segment lung regions in 1,400 chest X-ray scans, encompassing both tuberculosis cases and normal controls, obtained from the National Institute of Allergy and Infectious Diseases (NIAID) TB Portal program dataset. The segmented lung regions were subsequently classified as either tuberculosis or normal using a deep learning model. This integrated approach of segmentation and classification aims to enhance the accuracy and precision of TB detection in chest X-ray images. Classification of segmented images was done using customized CNN and visualization was done using Grad-CAM.

Results: Res-UNet model demonstrated excellent performance for segmentation, achieving an accuracy of 98.17%, recall of 98.39%, precision of 97.45%, F1-score of 97.97%, Dice coefficient of 96.33%, and Jaccard index of 96.05%. Similarly, the classification model exhibited outstanding results, with a classification accuracy of 99.42%, precision of 99.00%, recall of 99.29%, F1-score of 99.29%, and an AUC of 99.9%.

Conclusion: The findings demonstrate the efficiency of our system in diagnosing tuberculosis from chest X-rays, potentially surpassing clinician-level precision. This underscores its effectiveness as a diagnostic tool, particularly in resource-limited settings with restricted access to radiological expertise. Additionally, the modified Res-UNet model demonstrated superior performance compared to the standard U-Net, highlighting its potential for achieving greater diagnostic accuracy.