Revisiting Transfer Learning Method for Tuberculosis Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Transfer learning (TL) is effective for deep learning (DL) in specialized domains.
• Traditional TL uses pre-trained models as feature extractors for new classifiers.
• Existing methods face challenges in maximizing diagnostic accuracy for diseases like tuberculosis.

Purpose of the Study:

• To propose and evaluate a novel ensemble transfer learning approach for improved classification accuracy.
• To compare the ensemble method against direct transfer learning using popular DL models.
• To investigate the impact of feature extraction layers on classification performance.

Main Methods:

• An ensemble transfer learning approach was developed, integrating multiple neural network classifiers.
• The ensemble method was tested on VGG-16, ResNet-50, and MobileNet models.
• Performance was evaluated on Montgomery County (MC) and Shenzhen (SZ) tuberculosis datasets using accuracy, sensitivity, specificity, precision, and F1-score.

Main Results:

• The proposed ensemble transfer learning approach demonstrated superior performance over the direct approach.
• ResNet-50, with features extracted from a middle layer, achieved the highest accuracy (91.27%) on the MC dataset.
• The ensemble approach yielded accuracy improvements of 7-8% on the MC dataset and 4-5% on the SZ dataset.

Conclusions:

• The ensemble transfer learning strategy offers a significant improvement in tuberculosis detection accuracy.
• Feature extraction from middle layers of pre-trained models enhances performance.
• This approach holds clinical relevance for improving diagnostic capabilities in medical imaging.