SC-MAPS: Image-based deep learning for bacterial identification and antibiotic-susceptibility testing

Rapid bacterial identification could help guide antimicrobial susceptibility testing (AST), but image-based classification must generalise across bacterial strains. We developed SC-MAPS, a multichannel microscopy dataset comprising 916,435 single cells from 13 bacterial species, with independent strains reserved for evaluation. Combining phase-contrast and fluorescence information using deep learning achieved 65.4% single-cell species-classification accuracy on held-out strains. To assess whether imperfect species predictions retained useful information for downstream testing, we defined eight antimicrobial-resistance-informed AST-routing groups. Pooling species probabilities increased routing accuracy to 77.1%, rising to 91.1% when predictions from five cells were combined and 96.2% with 20 cells. Synthetic mixtures revealed limitations in minority-species recovery, while exploratory analysis of unseen organisms demonstrated the potential for out-of-distribution detection. SC-MAPS provides a framework for investigating transferable bacterial phenotypes and developing image-based identification and AST-routing strategies.