Detection of Tuberculosis Using Convolutional Neural Network

EasyChair Preprint 13500

Abstract

Tuberculosis (TB) remains a major public health challenge globally, and its burden
is particularly pronounced in the Kyrgyz Republic, where the prevalence of
multi-drug resistant (MDR) TB is high. This study aims to enhance early detection
of TB by developing a Convolutional Neural Network (CNN) model trained
on chest X-ray (CXR) images. Due to the lack of well-labeled CXR datasets in
Kyrgyz hospitals, our research utilized an open dataset of TB and normal CXR
images to train and validate the model. One of the challenges was the imbalance
in the target class. To tackle this problem, we computed the class weights.
We developed two models from scratch: the first one without class weights, and
the second one implemented with class weights. Our class weights improved the
performance of the model, which achieved 97% accuracy, 94% sensitivity, 98%
specificity, 88% precision and 91% F1 score. Our results demonstrate the potential
of CNN-based approaches in TB diagnosis and highlight the importance of
data infrastructure enhancement for advancing TB care in the Kyrgyz Republic.

Keyphrases: Chest X-ray, Convolutional Neural Network, Kyrgyz Republic, Tuberculosis, binary classification

@booklet{EasyChair:13500,
  author    = {Iskender Tologonov and Remudin Reshid Mekuria and Kylychbek Istamov and Mekia Shigute Gaso},
  title     = {Detection of Tuberculosis Using Convolutional Neural Network},
  howpublished = {EasyChair Preprint 13500},
  year      = {EasyChair, 2024}}