Colony Binary Classification Based on Persistent Homology Feature Extraction and Improved EfficientNet.

Area of Science:

• Microbiology
• Computer Vision
• Computational Biology

Background:

• Accurate classification of microbial colonies is crucial for infection source identification and precision medicine.
• Conventional computer vision methods struggle with ambiguous features in early-stage colony images.
• Existing algorithms lack efficiency and precision in classifying diverse microbial colonies.

Purpose of the Study:

• To develop a high-precision and efficient method for classifying microbial colonies.
• To overcome limitations of traditional computer vision techniques in analyzing early-stage colony images.
• To improve the accuracy of identifying bacterial species like Candida albicans and Staphylococcus epidermidis.

Main Methods:

• Application of Persistent Homology (PH) for topological feature extraction from microbial colonies.
• Modification of the EfficientNet architecture, specifically the MBConv module, to enhance attention mechanisms for small targets.
• Introduction of a novel Spatial and Contextual Transformer (SCoT) for multi-scale feature processing and improved aggregation.

Main Results:

• The proposed method achieved a classification accuracy of 98.64%.
• Demonstrated a 10.29% improvement in accuracy compared to the original classification model.
• Successfully captured topological information from Candida albicans and Staphylococcus epidermidis colonies.

Conclusions:

• The combined Persistent Homology and improved EfficientNet approach offers a highly accurate and efficient solution for microbial colony classification.
• This method effectively addresses challenges posed by ambiguous early-stage colony images.
• The findings support the clinical value of advanced computational methods in microbiology and precision medicine.