Automatic detection and quantification of antimicrobial inhibition zones using YOLO11n with post-hoc interpretability validation

Abstract

Introduction The escalating prevalence of antimicrobial resistance (AMR) constitutes a global healthcare crisis, necessitating rapid and standardized diagnostic solutions for antimicrobial susceptibility testing (AST). This study introduces an advanced, end-to-end artificial intelligence framework designed for the fully automated detection, quantification, and clinical interpretation of inhibition zones from disk diffusion assays using the state-of-the-art You Only Look Once (YOLO11n) object detection model.Methods A high-resolution dataset of Petri dish images, featuring antibiotic discs tested against Escherichia coli, Salmonella, and Staphylococcus aureus, was curated to train and validate the system under standardized conditions. The proposed pipeline localizes inhibition zones, quantifies their diameters with sub-millimeter precision, and automates resistance classification by integrating dynamic Clinical and Laboratory Standards Institute (CLSI) breakpoint criteria. Model interpretability was further ensured through Grad-CAM visualizations.Results Evaluation results demonstrate that the YOLO11n-based system achieved a Categorical Agreement (CA) of 94.2%, with a Very Major Error (VME) rate of 1.2% and a Major Error (ME) rate of 1.8%, performing well within clinical safety thresholds. High spatial accuracy was confirmed by a correlation coefficient of R2 = 0.98 and a Mean Absolute Error (MAE) of 0.42 mm in zone diameter prediction.Discussion Grad-CAM analysis confirmed that the architecture's attention is consistently aligned with biologically relevant inhibition boundaries rather than background artifacts. By providing an objective, reproducible, and transparent method for digital antibiogram analysis, this framework offers significant potential for seamless integration into clinical microbiology workflows and large-scale AMR monitoring programs. Grouped bar graph with two panels compares baseline and augmented models. Panel A shows mAP@50 and categorical agreement scores, where augmented outperforms baseline (mAP50: 95.6% vs 88.4%, categorical agreement: 94.2% vs 89.1%). Panel B shows the very major error rate, where augmented is lower (1.2%) than baseline (3.4%).