Development of Deep Learning based Antimicrobial Susceptibility Test Smart System

Abstract

Antimicrobial Susceptibility Test (AST) is a highly effective diagnostic method for determining the most suitable antibiotic and its optimal dosage. Conventional AST methods, such as the Disc diffusion test, are cost-effective but require 48–72 h to obtain results. On the other hand, advanced technological methods like PCR provide results in a much shorter time, approximately 6 h, but are more expensive and require a laboratory environment. This work presents the development of a cost-effective Automated AST smart system. The system incorporates automation through the utilization of IoT, image processing, and Deep Learning algorithms to effectively categorize the results. A semantic segmentation model was developed using CNN dilated convolutions for the automatic identification of inhibition zones on agar plates. The typical computation time on the model is 45 s. When evaluating the accuracy of segmentation performance, we take into account metrics such as F1 score, IoU, Mean accuracy, Weighted IoU, and BF score. The obtained results for Mean accuracy is 96.79, Weighted IoU is 0.9331, and BF score is 0.6944. It is possible to remotely monitor inhibition zones using the suggested approach. Periodic observations are made of the zones, and their susceptibility categorization is calculated every 15 min. Using this advanced technique, it is possible to test three samples simultaneously, with the ability to test each sample against a maximum of 12 antibiotics. With the suggested automation setup, test time is significantly reduced to 4–6 h, resulting in a more cost-effective solution.