Abstract:
This study presents the development and evaluation of ResisTrack, a deep learningbased
web diagnostic system for classifying drug-resistant tuberculosis (TB) using
chest X-ray (CXR) images. The classification targets three clinically relevant TB
categories: drug-sensitive (DS-TB), multi-drug-resistant (MDR non-XDR), and
extensively drug-resistant (XDR-TB). A U-Net model was first trained on Montgomery
and Shenzhen datasets for lung segmentation, followed by preprocessing
steps to normalize and resize the images. Using the TB Portals dataset, two classification
tasks were formulated: binary (DS-TB vs. MDR non-XDR) and multiclass
(DS-TB, MDR non-XDR, XDR-TB). To address class imbalance, random oversampling
and geometric data augmentation were applied. Transfer learning was
employed with three pre-trained convolutional neural networks—DenseNet201,
DenseNet121, and InceptionV3—evaluated individually and as an ensemble using
hard voting.
The models were tested across four experimental configurations (with and without
augmentation in both binary and multiclass setups) and assessed using accuracy,
AUC, precision, recall, F1, and F2 scores. The ensemble model for binary
classification without augmentation achieved the best overall performance: 76.70%
accuracy, 83.40% AUC, 83.98% recall, and 82.19% F2 score, highlighting its capability
to minimize false negatives—critical in TB triage. This best-performing
model was deployed in the ResisTrack web system, enabling real-time CXR classification
with PDF reporting via a user-friendly interface. This work demonstrates
the potential of deep learning to support TB resistance detection in clinical and
resource-constrained settings.
