Abstract

Objectives: AI (artificial intelligence) technologies have allowed us to automate the analysis, processing, and decision-making based on vast amounts of medical data. As a result, disease diagnosis has become more accurate and faster, saving time and enhancing treatment capabilities. Building upon this, we conducted a study on the application of machine learning in image recognition of Kirby-Bauer diffusion antibiotic susceptibility test results to lay the foundation for further research in the use of artificial intelligence to support healthcare professionals in examination, diagnosis, and decision-making in the medical field. Material and methods: We utilized 200 images of antibiotic susceptibility test results obtained through Kirby-Bauer diffusion technique, with the number of antibiotic zones ≤ 6. The images were labeled to distinguish the inhibition zone and antibiotic zones. We employed the Single-shot Detection (SSD) MobileNet method to achieve a balance between speed, ease of deployment, and accuracy, especially when working with devices with limited computational power. The results obtained from the model’s detection of antibiotic zone labels were used as input for the result recognition algorithm. Results and discussion: Out of the 200 images, 160 (80%) were successfully identified with sufficient information regarding the number of antibiotic zones. Among the 40 failed images, 30 (15%) lacked sufficient recognition of both the bacterial inhibition zones and the number of antibiotic zones, while 10 (5%) images were unrecognizable. The application demonstrated accurate identification and classification of antibiotic zone positions and bacterial inhibition circles in images with appropriate brightness and contrast. Conclusion: We have successfully developed an initial tool and algorithm based on a machine learning model to quickly and accurately determine the necessary information of a Kirby-Bauer antibiotic susceptibility test result, including the number of antibiotic zones, bacterial inhibition circles, and diameter of inhibition circles, achieving an accuracy rate of 80%.