Identification of Microbial Species Using an Instance Segmentation Model

Accurate identification of microbial species in urinary tract infections (UTIs) is crucial for effective treatment and prevention of severe complications, such as kidney damage and mortality. Diagnostic methods used in high-resource countries are often not feasible in low-resource countries, making proper diagnosis and treatment difficult in those settings. This study addresses this issue by developing a cost-efficient and computationally inexpensive alternative using a customized YOLOv8 (You Only Look Once) deep learning model for microbial identification. We implemented the Fast Segment Anything Model (FastSAM), a powerful segmentation model, to create masks serving for the training of the YOLOv8 model. Our approach was evaluated on a dataset of nine common microbial species: Candida albicans, Enterococcus faecalis, Escherichia coli, Klebsiella pneumoniae, Pseudomonas aeruginosa, Staphylococcus aureus, Staphylococcus epidermidis, Staphylococcus saprophyticus, and Streptococcus agalactiae. The model achieved a notable mean average precision of 0.79, which demonstrates its ability to detect and identify microbial species. The results are indicative of the model’s potential as a reliable and accessible diagnostic aid. This approach can make microbial identification more efficient and widely available, thus improving patient treatment with accurate and timely diagnosis.