Abstract:
Tuberculosis (TB) remains a significant global public health issue, particularly
affecting resource-constrained countries like the Philippines. Early and accurate
diagnosis of TB is crucial for effective patient management and control of its
spread. However, conventional diagnostic processes relying on human interpretation
of chest X-rays are prone to delays, variability, and errors. This study
proposes an automated diagnostic solution using state-of-the-art convolutional
neural network (CNN) architectures—ResNet50, EfficientNetB0, VGG19, and InceptionV3—
to classify chest X-ray images as either TB-positive or TB-negative.
Among the evaluated models, InceptionV3 achieved superior performance. The
system integrated preprocessing techniques such as Contrast Limited Adaptive
Histogram Equalization (CLAHE) to improve image quality, enhancing prediction
accuracy. Moreover, Gradient-weighted Class Activation Mapping (Grad-CAM)
was implemented as an Explainable Artificial Intelligence (XAI) technique, significantly
enhancing the interpretability of model predictions by visually indicating
regions relevant to TB pathology. A user-friendly Django-based web application
was developed, enabling healthcare professionals to interact seamlessly with the
diagnostic system. Despite high performance, rare instances of false positives and
false negatives were observed, emphasizing the necessity for clinical validation of
AI-driven diagnostics. Overall, this research contributes towards improving TB
diagnostic accuracy, reducing healthcare worker burden, and facilitating interpretability
in clinical practice.
