Sputum smears quality inspection using an ensemble feature extraction approach.

Area of Science:

• Medical Diagnostics
• Computer Vision
• Image Processing

Background:

• Tuberculosis (TB) diagnosis is critical, especially in resource-constrained nations.
• Sputum smear microscopy is accessible but manual, posing quality control challenges.
• Existing diagnostic methods lack consistent quality assessment for sputum smears.

Purpose of the Study:

• To evaluate sputum smear quality for tuberculosis diagnosis in Ethiopian regions.
• To propose and validate an automated model for sputum smear quality inspection.
• To enhance the accuracy and reliability of tuberculosis diagnosis through image analysis.

Main Methods:

• Collected and expert-labeled sputum smear images from multiple Ethiopian hospitals using a smartphone.
• Preprocessed images using bicubic resizing, ROI extraction, noise reduction (Gaussian, Gabor filters), and enhancement (CLAHE).
• Employed ensemble feature extraction (GLCM for grayscale, CNN for color) and tested KNN, SVM, and CNN classifiers.

Main Results:

• The hybrid approach combining CNN and GLCM features with a KNN classifier achieved the highest accuracy of 94%.
• This automated method demonstrated superior performance compared to using GLCM or CNN features alone.
• The proposed model showed a significant improvement over existing methods, increasing accuracy by 1.48%.

Conclusions:

• An automated sputum smear quality inspection model using ensemble feature extraction (CNN and GLCM) is effective for tuberculosis diagnosis.
• The hybrid CNN-GLCM approach with KNN classification offers a promising, accurate, and accessible tool for resource-limited settings.
• This technology can significantly improve the reliability and efficiency of tuberculosis diagnosis, aiding global health initiatives.