Deep Learning-Based Classification and Semantic Segmentation of Lung Tuberculosis Lesions in Chest X-ray Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Tuberculosis (TB) diagnosis relies heavily on chest X-ray (CXR) interpretation.
• Accurate detection and segmentation of TB lesions in CXRs are crucial for effective patient management.
• Existing methods may lack the precision required for detailed lesion analysis.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) network for detecting and segmenting specific TB lesions in CXR images.
• To compare the performance of various U-Net based architectures, including Attention U-Net and PSP Attention U-Net++.
• To enhance diagnostic accuracy through ensemble methods combining multiple DL models.

Main Methods:

• Utilized U-Net, Attention U-Net, U-Net++, Attention U-Net++, and PSP Attention U-Net++ architectures.
• Optimized model parameters through rigorous testing and comparison.
• Employed data augmentation and preprocessing techniques to strengthen lesion features.
• Developed four ensemble approaches combining the top five performing models.

Main Results:

• The proposed ensemble model achieved a maximum mean intersection-over-union (MIoU) of 0.70.
• Achieved a mean precision of 0.88, mean recall of 0.75, and mean F1-score of 0.81.
• Demonstrated superior performance over single-network models, with an accuracy of 1.0.
• Significantly improved lesion classification and segmentation results.

Conclusions:

• Deep learning ensemble models offer enhanced accuracy for TB lesion detection and segmentation in CXRs.
• The developed method shows potential as a valuable diagnostic tool for clinicians.
• Further research can explore larger datasets and diverse lesion types for broader clinical application.