Bacterial ID stops chasing exact species: pooling 13-species single-cell images into AST-routing groups lifts accuracy from 65.4% to 77.1%

Bacterial identification now draws on several microscopy views of the same cell rather than a single imaging channel. Among individual channels, phase contrast generalised best, reaching 54.8% species accuracy with DenseNet121 on the independent holdout strain; DAPI was next, WGA-AF647 and FM1-43 varied by species, and bright field was consistently weakest. Averaging class probabilities from independently trained DAPI, WGA-AF647, FM1-43 and phase-contrast models raised DenseNet121 holdout accuracy to 61.2%, and feature-level fusion raised it further to 65.4%.

Beyond exact species names, the image information supports coarser routing groups tied to antimicrobial-susceptibility testing. The 13 species were assigned a priori to eight AST-routing groups defined by initial testing requirements, with single-cell species accuracy at 65.4%. Summing probabilities of species in the same group gave 77.1% single-cell routing accuracy and 74.8% macro-F1, above a group-size-matched randomised mean of 68.1% (empirical P = 0.0115); 34.5% of species-level errors still landed in the correct routing group.

Combining predictions from several cells of one sample is more reliable than judging cells one at a time. Across 120 holdout samples with at least 200 segmented cells, single-cell routing accuracy was 77.9%. Routing accuracy rose to 91.1% with five cells, 94.2% with ten, 96.2% with twenty and 99.1% with 200, while species accuracy went from 66.3% with one cell to 80.5% with five and 86.5% with twenty.

The model can also flag organisms it never saw as unfamiliar instead of forcing them into a known species. The classifier is closed-set, so cells from five unseen species (Streptococcus agalactiae, Enterococcus faecalis, Enterococcus faecium, Listeria monocytogenes and Haemophilus influenzae) must be assigned to one of the 13 known classes. k-nearest-neighbour distance in feature space separated known from unseen cells with an AUROC of 0.944, rejecting 77.8% of unseen cells while retaining 94.3% of known cells; averaging five cells raised unseen detection from 79.7% to 99.8%.

The evaluation covers fixed microscopy data from 13 species with one independent holdout strain per species; routing accuracy refers to the eight study-defined groups and does not predict susceptibility, MIC, resistance mechanism or treatment response. The authors propose broader isolate collections and external evaluation across operators, laboratories, culture conditions and imaging systems as next steps.

The authors note that strains per species remain limited and that the open-set analysis included only five unseen organisms, so it is proof of principle rather than validation of a general open-set classifier; synthetic mixtures were computed from independently imaged monocultures and do not capture physical co-culture, staining interactions or sample-matrix effects; confidence intervals for direct eight-class training and species-derived pooling overlapped, so the two were not shown to differ significantly.