Deep learning-based survival monitoring for antimicrobial efficacy evaluation in the silkworm infection model

Silkworm infection models offer a low-cost in vivo platform for antimicrobial screening, but repeated survival assessment is labor-intensive. We developed a time-lapse Faster R-CNN workflow for a silkworm Pseudomonas aeruginosa (strain PAO1) infection model. In a primary benchmark, 508 of 570 expert-annotated larval instances (89.1%) were detected at prespecified thresholds. Among these instances, classification accuracy was 94.7%, Cohen’s kappa was 0.743, infection-killed sensitivity was 70.8%, and precision was 85.2%. Because individual larvae were not tracked, outputs were analyzed as count-based survival trajectories. The pool-adjacent-violators algorithm (PAVA) had a negligible effect on time-integrated trajectories (< 0.01%) and did not change group ordering. Across two independent PAO1 experiments, meropenem-treated groups consistently showed higher time-integrated survival proportions than untreated infected groups. The workflow supports automated group-level antimicrobial efficacy screening under defined imaging conditions.