U-Net Guided EfficientNet Framework for Accurate Bacterial Colony Counting and Species Recognition

Abstract

Accurate enumeration of bacterial colonies on agar plates is a critical step in microbiology, enabling reliable assessment of microbial growth, contamination, and antimicrobial activity. Traditional manual counting is labor intensive, subjective, and prone to error, particularly when colonies overlap or vary in size and contrast. To address these challenges, we propose a two stage deep learning framework that integrates U-Net based image segmentation with EfficientNet family regression and classification models for automated colony detection, counting, and species identification. In this first stage, a U-Net architecture performs pixel level segmentation of agar plate images to isolate colony regions from the background. This preprocessing step effectively handles variations in illumination and background noise, providing high quality binary masks that delineate individual colonies. In the second stage, segmented images are processed by a series of EfficientNet variants (B0 through B7) to simultaneously estimate total colony count and classify colonies by species. Training and evaluation were performed on a curate dataset of  annotated agar plate images spanning five clinically relevant species Bacillus subtilis (B.subtilis), Candida albicans (C.albicans), Escherichia coli (E.coli), and Pseudomonas aeruginosa (P.aeruginosa), Staphylococcus aureus (S.aureus) with ground truth counts and species labels provided in JSON format. Comprehensive experiments demonstrate that all EfficientNet models achieve strong predictive accuracy with EfficientNetB7 delivering the best overall performance Mean Absolute Error (MAE=2.60), Root Mean Square Error (RMSE=12.57), Coefficient of Determination (R²=0.67), accuracy=96.6%. Per-species F1-Scores range from 0.93 to 0.96 indicating consistent detection across diverse colony morphologies. The proposed U-Net + EfficientNet pipeline offers a fast, reproducible, and scalable solution for high throughput microbial analysis, significantly reducing manual effort while improving the reliability of colony enumeration and species classification in clinical and research laboratories.