Deep ensemble approach for pathogen classification in large-scale images using patch-based training and...

Area of Science:

• Microbiology
• Computer Science
• Artificial Intelligence

Background:

• Pathogenic bacteria pose significant health risks, necessitating accurate identification.
• Bacterial identification can be challenging due to species similarities.
• Automated classification offers a standardized and accurate solution.

Purpose of the Study:

• To develop and evaluate a robust automated classification system for pathogenic bacteria using convolutional neural network (CNN) models.
• To improve bacterial identification accuracy and efficiency in diagnostic settings.

Main Methods:

• Dataset augmentation via image patching, random rotation, reflection, and translation.
• Application of various CNN models: training from scratch, fine-tuning, and weight adjustment.
• Modification of existing architectures (InceptionV3, MobileNetV2) and development of an ensemble model.
• Evaluation of model robustness using 7:2:1 and 6:2:2 data splits.

Main Results:

• Augmentation and fine-tuning of deep CNN models yielded optimal results.
• The ensemble model demonstrated exceptional performance across data splits.
• Achieved high accuracy (up to 99.94%) and F-Score (up to 99.28%) in bacterial classification.
• Demonstrated robustness with increased training data.

Conclusions:

• Automated classification using ensemble CNN models is a valuable tool for accurate pathogenic bacteria identification.
• This technology can assist diagnostic staff and microbiologists, improving disease control.
• Effective bacterial identification can mitigate the social and economic impact of infections.