Improving Tuberculosis Detection in Chest X-Ray Images Through Transfer Learning and Deep Learning: Comparative Study...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Tuberculosis (TB) poses a significant global health challenge, with current diagnostics often being resource-intensive and inaccessible.
• There is a critical need for efficient and accurate diagnostic methods for early TB detection and improved treatment outcomes.

Purpose of the Study:

• To evaluate six convolutional neural network (CNN) architectures for classifying chest X-ray (CXR) images as normal or TB-positive.
• To assess the impact of data augmentation on model performance, training times, and parameter counts.

Main Methods:

• Trained and tested six CNNs (VGG16, VGG19, ResNet50, ResNet101, ResNet152, Inception-ResNet-V2) on a dataset of 4200 CXR images (700 TB-positive, 3500 normal).
• Evaluated models using accuracy, precision, recall, F1-score, and AUC.
• Analyzed computational efficiency by comparing training times and parameter counts.

Main Results:

• VGG16 achieved the highest performance with 99.4% accuracy, 97.9% precision, 98.6% recall, 98.3% F1-score, and 98.25% AUC.
• Simpler models like VGG16 demonstrated superior diagnostic accuracy with fewer computational resources.
• Data augmentation did not improve performance, and larger models (ResNet152, Inception-ResNet-V2) required longer training times without proportional gains.

Conclusions:

• VGG16 offers an optimal balance of diagnostic accuracy and computational efficiency for TB detection in CXR images.
• Model selection should be tailored to specific task requirements for effective medical image classification.
• Findings provide valuable insights for future research and clinical implementation of AI in TB diagnostics.