Enhanced lung image segmentation using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Pulmonary Disease Diagnosis

Background:

• Deep learning models show promise in diagnosing diseases like tuberculosis (TB) from X-rays.
• Accurate lung segmentation is crucial for improving classification accuracy in pulmonary diagnostics.
• Existing methods often overlook segmentation or use less advanced techniques, potentially leading to data leakage.

Purpose of the Study:

• To evaluate the efficacy of U-Net++ for precise lung segmentation in X-ray images.
• To compare U-Net++ performance against other segmentation architectures for pulmonary disease diagnosis.
• To demonstrate the benefits of segmentation-assisted classification in minimizing data leakage and enhancing TB detection.

Main Methods:

• Implementation and analysis of the U-Net++ architecture for lung segmentation on X-ray data.
• Comparative evaluation of U-Net++ against three benchmark segmentation models.
• Integration of segmented lung data into deep learning classification models for TB diagnosis.

Main Results:

• U-Net++ achieved over 98% accuracy and a 0.95 mean Intersection over Union (mIoU) in lung segmentation.
• The study validated the comparative analysis, demonstrating U-Net++'s superior performance.
• Segmentation using U-Net++ significantly improved diagnostic accuracy compared to using whole X-ray images.

Conclusions:

• U-Net++ offers a novel and highly effective approach for lung segmentation in medical imaging.
• Implementing advanced segmentation techniques like U-Net++ is vital for accurate and reliable AI-driven pulmonary disease diagnosis.
• This research highlights the potential of U-Net++ to advance assistive medical systems and reduce diagnostic errors.