Harnessing semi-supervised graph-based learning to advance automated bacilli detection in digital tuberculosis...

Area of Science:

• Medical Diagnostics
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Automated acid-fast bacilli (AFB) detection via digital microscopy is crucial for tuberculosis diagnosis, especially where expert annotations are scarce.
• Digital pathology offers potential for scalable and accessible diagnostic tools in resource-limited settings.

Purpose of the Study:

• To develop and evaluate a semi-supervised graph-based learning approach for automated AFB detection using digital microscopy.
• To leverage both limited labeled and extensive unlabeled high-resolution smear microscopy images for robust diagnostic model training.

Main Methods:

• A semi-supervised graph-based learning framework utilizing label spreading was implemented.
• Smooth diffusion-over-graph algorithms with soft constraints were employed to propagate label evidence and minimize error propagation.
• The approach processed high-resolution smear microscopy images for bacilli localization.

Main Results:

• The proposed framework demonstrated robust and precise localization of bacilli.
• The method achieved highly competitive overall accuracy and F1-score, even with limited expert annotations.
• Experimental results validated the framework's effectiveness in handling imbalanced and limited labeled data.

Conclusions:

• The developed semi-supervised learning method is applicable for rapid, scalable, AI-powered tuberculosis diagnosis in low-resource settings.
• This approach effectively reduces the need for extensive manual annotation in digital pathology workflows.
• Semi-supervised learning shows significant promise for enhancing tuberculosis diagnostic capabilities globally.