An adaptive convolution neural network model for tuberculosis detection and diagnosis using semantic segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Tuberculosis (TB) remains a global health threat, often misdiagnosed or untreated in resource-limited settings.
• Chest X-rays are crucial for TB diagnosis but face challenges due to varied presentations and radiologist shortages.
• Deep learning offers a promising solution for computer-aided TB detection in medical imaging.

Purpose of the Study:

• To develop and evaluate deep learning models for enhanced detection of Tuberculosis (TB) in chest X-ray images.
• To improve the accuracy and precision of TB diagnosis using AI-driven image analysis.
• To address diagnostic challenges in areas with limited access to trained radiologists.

Main Methods:

• A Res-UNet model was trained for lung segmentation on 704 chest X-rays and applied to 1400 scans.
• A novel deep learning network was developed for classifying segmented lung regions as TB or normal.
• Preprocessing involved gamma correction and a gradient-based technique for contrast enhancement.

Main Results:

• The Res-UNet segmentation model achieved high performance (e.g., 98.18% accuracy, 97.97% F1-score).
• The classification model demonstrated outstanding results (e.g., 99.45% accuracy, 99.29% F1-score, 99.9% AUC).
• The gradient-based enhancement method yielded satisfactory image quality metrics.

Conclusions:

• The developed system shows high efficiency in diagnosing TB from chest X-rays, potentially exceeding clinician-level precision.
• The AI tool is particularly valuable for resource-limited settings lacking radiological expertise.
• The modified Res-UNet model outperformed standard U-Net, indicating potential for improved diagnostic accuracy.