Implementation of convolutional neural networks for microbial colony recognition.

Area of Science:

• Microbiology
• Artificial Intelligence
• Computer Vision

Background:

• Microorganism classification relies heavily on subjective visual identification by skilled microbiologists.
• This subjectivity can lead to inconsistencies in clinical classification due to varying expertise and interpretation.

Purpose of the Study:

• To apply deep learning for automated microbial colony recognition and classification.
• To develop an objective, data-driven tool to assist microbiologists in clinical identification tasks.

Main Methods:

• Photographs of clinically isolated microbial colonies were used to create a 48x48 pixel dataset.
• Eight convolutional neural networks (CNNs) were trained and evaluated for colony classification.
• Performance metrics included accuracy, precision, recall, and F1 score for categories like gram-negative bacilli, gram-positive cocci, Candida, and Aspergillus.

Main Results:

• GoogLeNet demonstrated the highest classification accuracy at 98.80%, with other models like MobileNet and ShuffleNet also exceeding 98% accuracy.
• The models showed high generalization performance, correctly identifying strains not present in the training data.
• GoogLeNet misclassified only 6 out of 500 test images.

Conclusions:

• Deep learning, particularly CNNs, offers a robust and accurate method for microbial colony recognition.
• These AI-driven tools can serve as valuable auxiliary decision-making aids for microbiologists, enhancing objectivity and efficiency.
• The developed method is practical, requiring no specialized equipment and showing strong performance on diverse microbial strains.