Efficient Deep Network Architectures for Fast Chest X-Ray Tuberculosis Screening and Visualization.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Automated tuberculosis (TB) diagnosis from chest X-rays (CXR) often uses complex deep learning models.
• Existing models, adapted from natural image classification, have high parameter counts and hardware needs, limiting their use in mobile settings.

Purpose of the Study:

• To propose a simple, optimized convolutional neural network (CNN) for accurate and efficient TB diagnosis from CXR.
• To investigate the visualization capabilities of CNNs for TB diagnosis using saliency maps and grad-CAMs.

Main Methods:

• Developed a streamlined convolutional neural network (CNN) architecture specifically for tuberculosis detection in chest X-rays.
• Evaluated model performance against existing methods, focusing on accuracy, speed, and efficiency.
• Applied and analyzed saliency maps and gradient-weighted class activation mapping (grad-CAM) for visualizing diagnostic features.

Main Results:

• The proposed simple CNN achieves comparable accuracy to more complex models.
• The optimized model demonstrates increased speed and efficiency, reducing hardware requirements.
• Visualization techniques provide radiological insights into the CNN's decision-making process.

Conclusions:

• A simple, efficient CNN offers a viable alternative for automated TB diagnosis from CXR.
• The model's reduced complexity facilitates deployment in resource-constrained environments.
• CNN visualization methods enhance interpretability and can aid radiological assessment.