Detection and Measurement of Inhibition Zones in AST Using Deep Learning
Jadrná, Veronika (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025060621134
https://urn.fi/URN:NBN:fi:amk-2025060621134
Tiivistelmä
The antimicrobial resistance is one of the current global health treats in the 21st century. One of the common methods for classifying bacteria as resistant or susceptible is antimicrobial susceptibility testing. This method involves measuring the inhibition zones created by the reaction of bacteria to different antimicrobials and measuring them based on the interpretation guidelines and breakpoints. One of the disadvantages of this method is a manual and time-consuming measurement process.
This study presents an alternative solution for the detection and measurement of inhibition zones using deep learning and aims to contribute to more efficient and faster laboratory diagnostics. During this study several object detection models were implemented for the identification of inhibition zones from agar plate images. After the evaluation of different models including YOLO, SSD and Faster R-CNN, it was found out that the YOLO models achieved the best performance results. Specifically, YOLOv8m
achieved the best recall of 100% and mean average precision of 85.3% at IoU threshold between 0.5 and 0.95.
Subsequently, three best models were used for bounding box measurement analysis. The measurement task included width calculation of the bounding boxes in millimetres and their comparison to ground truth annotations. The results suggest that YOLOv8m model achieved best localization accuracy with 88.8% predictions falling within 1 mm deviation and 98.9% predictions falling within 2 mm deviation. Furthermore, the results showed that the maximum difference between predicted bounding boxes and their annotations was 3 mm.
These findings support a conclusion that the deep learning models can achieve high detection and measurement results and have a potential for practical use in laboratory environment. However, it is important to point out, that while these results are promising, further improvements in terms of localization accuracy of bounding boxes are inevitable to ensure more precise and accurate laboratory diagnostics.
This study presents an alternative solution for the detection and measurement of inhibition zones using deep learning and aims to contribute to more efficient and faster laboratory diagnostics. During this study several object detection models were implemented for the identification of inhibition zones from agar plate images. After the evaluation of different models including YOLO, SSD and Faster R-CNN, it was found out that the YOLO models achieved the best performance results. Specifically, YOLOv8m
achieved the best recall of 100% and mean average precision of 85.3% at IoU threshold between 0.5 and 0.95.
Subsequently, three best models were used for bounding box measurement analysis. The measurement task included width calculation of the bounding boxes in millimetres and their comparison to ground truth annotations. The results suggest that YOLOv8m model achieved best localization accuracy with 88.8% predictions falling within 1 mm deviation and 98.9% predictions falling within 2 mm deviation. Furthermore, the results showed that the maximum difference between predicted bounding boxes and their annotations was 3 mm.
These findings support a conclusion that the deep learning models can achieve high detection and measurement results and have a potential for practical use in laboratory environment. However, it is important to point out, that while these results are promising, further improvements in terms of localization accuracy of bounding boxes are inevitable to ensure more precise and accurate laboratory diagnostics.