Automated detection of Mycobacterium tuberculosis using transfer learning

Area of Science:

• Medical diagnostics
• Computer vision
• Microbiology

Background:

• Microscopic analysis of Mycobacterium tuberculosis is crucial for tuberculosis diagnosis.
• Challenges include small bacterial size, low contrast, overlapping cells, and time-consuming manual review, leading to potential misdiagnosis.

Purpose of the Study:

• To develop and evaluate an automated detection method for Mycobacterium tuberculosis using a pretrained AlexNet model.
• To address limitations in manual microscopic diagnosis and improve diagnostic accuracy.

Main Methods:

• An automated detection system was developed using a pretrained AlexNet convolutional neural network.
• The model was trained and tested across three experimental sets (A, B, and C) with adjusted protocols.
• Performance was compared against ground truth results provided by microbiologists.

Main Results:

• Experiment A: 98.15% accuracy, 96.77% sensitivity, 100% specificity.
• Experiment B: 98.09% accuracy, 98.59% sensitivity, 97.67% specificity.
• Experiment C: 98.73% accuracy, 98.59% sensitivity, 98.84% specificity.

Conclusions:

• The automated AlexNet-based detection method demonstrates robust and promising performance.
• The system achieved high accuracy, sensitivity, and specificity, indicating its potential to aid microbiologists in tuberculosis diagnosis.