Automated AI-driven platform for comprehensive Kirby-Bauer antibiogram testing analysis = 커비-바우어 항생제 감수성 검사 분석을 위한 AI 기반 자동화 플랫폼

다국어 초록 (Multilingual Abstract)

Antimicrobial resistance （AMR） is a global health emergency responsible for over 1.27 million direct deaths annually, with projections reaching ten million by 2050. Al- though the Kirby-Bauer disk diffusion test remains the most widely adopted antibiotic susceptibility testing （AST） method in clinical microbiology, manual plate reading is time-consuming, suffers from inter-reader variability, and is prone to measurement errors when estimating zone diameters. This thesis presents an automated platform that takes a single photograph of an inoculated Petri dish and returns Susceptible/Intermediate/Resistant （SIR） interpretations through a five-stage deep learning pipeline. A Faster R-CNN de- tector with a ResNet-101 FPN backbone localizes each antibiotic disk and its zone of inhibition; disk-zone pairs are matched geometrically; a fine-tuned ResNet-18 classi- fies the printed antibiotic label across 108 classes; zone diameters are then computed in millimeters via a calibration-free pixel-ratio formula and interpreted against CLSI M100 breakpoints. The pipeline is served through a FastAPI backend and a Next.js web inter- face that supports human-in-the-loop verification. On a test set of 59 images containing 348 annotated zones, the system achieves 81.9% zone-measurement similarity, the high- est among five benchmarked detector architectures, and the ResNet-18 label classifier reaches 99.39% Top-1 and 99.85% Top-5 accuracy across 108 antibiotic classes. These re- sults suggest that an end-to-end open-source pipeline can meaningfully assist manual Kirby-Bauer interpretation, though broader clinical validation remains necessary before deployment.

목차 (Table of Contents)

• 1 Introduction 1
• 1.1 Motivation 2
• 1.2 Contributions 5
• 1.3 Thesis Organization 6
• 2 Background and Related Work 7
• 2.1 Antibiotic Susceptibility Testing 7
• 2.2 Deep Learning for Object Detection 9
• 2.3 Deep Learning for Image Classification 11
• 2.4 Automated Antibiotic Susceptibility Testing 12
• 2.5 Positioning of This Thesis 14
• 3 Proposed System Design 16
• 3.1 System Overview 16
• 3.2 Object Detection Module 18
• 3.3 Antibiotic Label Classification Module 20
• 3.4 Geometric Zone Measurement 21
• 3.5 CLSI Breakpoint Classification 24
• 4 System Implementation 28
• 4.1 Dataset and Annotation 28
• 4.2 Training Pipeline 29
• 4.3 Backend API 34
• 4.4 Web Application 38
• 4.5 Human-in-the-loop Workflow 40
• 5 Experiments and Results 42
• 5.1 Experimental Setup 42
• 5.2 Object Detection Performance 44
• 5.3 Label Classification Performance 48
• 5.4 Zone Measurement Performance 50
• 5.5 Threshold Robustness Across Architectures 52
• 5.6 Inference Cost and Deployability 53
• 5.7 Qualitative Observations from the Test Set 54
• 6 Limitations and Future Works 56
• 6.1 Limitations 56
• 6.2 Future Works 58
• 7 Conclusion 63
• Bibliography 65