Evaluation of Depth-wise Separable Convolution and Channel Attention Mechanism to Bacilli Segmentation

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Tuberculosis (TB) remains a significant global health threat, causing millions of deaths annually.
• Automating the identification of Mycobacterium tuberculosis bacilli in sputum smear microscopy is crucial for rapid diagnosis.
• Deep learning has emerged as a powerful tool for image analysis in medical diagnostics.

Purpose of the Study:

• To evaluate deep learning architectures for automated bacilli segmentation in microscopy images.
• To investigate the impact of channel attention mechanisms and depth-wise separable convolutions on segmentation performance.
• To identify the optimal deep learning model for accurate tuberculosis bacilli detection.

Main Methods:

• Convolutional neural network (CNN) architectures of varying depths were implemented.
• Channel attention mechanisms were integrated to enhance feature representation.
• Depth-wise separable convolutions were employed to improve computational efficiency.
• Performance was evaluated using the F1-score metric for bacilli segmentation.

Main Results:

• Deeper CNN architectures generally outperformed shallower ones.
• The integration of channel attention and depth-wise separable convolutions significantly boosted performance.
• The best-performing model, a deeper architecture with these enhancements, achieved an F1-score of 99.38%.
• This indicates highly accurate segmentation of bacilli in microscopy images.

Conclusions:

• Deep learning, particularly with channel attention and depth-wise separable convolutions, offers a highly effective solution for automated bacilli segmentation.
• The optimized deep learning model demonstrates potential for improving the speed and accuracy of tuberculosis diagnosis.
• Further research can explore deployment in clinical settings for real-time TB detection.