Machine learning-guided screening of urine cultures for urinary tract infection diagnosis according to AMCLI guidelines

Urinary tract infections (UTIs) are among the most prevalent bacterial infections, and their diagnosis relies on the accurate interpretation of urine cultures. However, manual evaluation is time-consuming, operator-dependent, and susceptible to diagnostic variability. In this study, we propose a machine learning–guided pipeline for the automatic classification and segmentation of urine culture images, aligned with the most recent diagnostic guidelines. A dataset of 115 original chromogenic agar plate images, each corresponding to a unique patient, was acquired under standardized conditions and annotated by clinical microbiologists into three diagnostic categories: negative, positive (monomicrobial), and polymicrobial/contaminated. Data augmentation was subsequently applied after dataset partitioning to increase image variability for model development. Five convolutional neural networks (ResNet-18, -50, -152; VGG-11, -19) were trained and evaluated for image-level classification. VGG-19 achieved the best performance, with an accuracy of 99.49% on the test set. For colony-level detection and segmentation, we employed a state-of-the-art YOLOv11 instance segmentation model, enabling simultaneous localization, classification, and mask generation for individual bacterial colonies. This model showed strong results (mAP@50 = 0.82), particularly in detecting Enterobacter aerogenes and Klebsiella spp ., although segmentation quality decreased at higher IoU thresholds for less-represented species. This study suggests that deep learning models can support the automated screening of urine cultures under standardized acquisition conditions. The proposed pipeline achieved high classification performance and may represent a useful approach for assisting routine laboratory workflows. Further validation on larger multicenter datasets will be required before clinical implementation.