Leveraging Smartphone Imaging and Deep Transfer Learning for Bacterial Colony Classification: From Uncontrolled to Controlled Settings

The classification of bacterial colony types is crucial for microbiology and medical diagnostics because it provides essential information on pathogen identification and treatment planning. Traditional classification methods are resource intensive and require specialized equipment and expertise. This study explores the integration of smartphone imaging and transfer learning techniques for bacterial colony classification. We expanded an existing dataset of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Escherichia coli (E. coli)</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Klebsiella pneumoniae (K. pneumoniae)</i> bacteria grown on MacConkey (MAC) agar by incorporating <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Pseudomonas aeruginosa (P. aeruginosa)</i> on the same medium, as well as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Staphylococcus aureus (S. aureus), P. aeruginosa, K. pneumoniae</i>, and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E. coli</i> on brain–heart infusion (BHI) agar. The total numbers of images are 9819 and 12348 in the MAC agar and BHI agar datasets, respectively. Using state-of-the-art convolutional neural networks (CNNs) such as SqueezeNet1.0, MobileNetV3 Small, ShuffleNetV2 X1.5, EfficientNetB0, DenseNet121, and ResNet18, we evaluated classification performance across uncontrolled and controlled imaging environments. We also compared the results for the same bacterial type on different media (MAC agar and BHI agar). The results show that the lightweight models EfficientNetB0 and ShuffleNetV2 X1.5 consistently perform well across datasets, reaching 100% accuracy under controlled conditions. We believe that this work provides a foundation for cost-effective bacterial colony classification on smartphones using transfer learning techniques, ultimately enabling practical implementation. Future work will focus on enhancing model explainability and evaluating deployment feasibility on smartphone devices.