Research And Application Of 96-well Microtiter Plate Image Recognition Based On Deep Learning

The resistance of animal-originated pathogenic antimicrobial in China is becoming increasingly serious,posing a major threat to public health and food safety.The prerequisite for achieving antimicrobial resistance(AMR)control is real-time and accurate AMR surveillance.At present,the main method of AMR surveillance in China is to use96-well microtiter plate(MTP)to perform antimicrobial susceptibility testing(AST).However,due to the limited equipment conditions and the limited professional ability of the practitioners in China’s primary farms,especially small and medium-sized farms,the interpretation of the results of AST is manual interpretation or depends on expensive equipment,resulting in slow veterinary clinical AST.In recent years,deep learning technologies such as convolutional neural networks have developed rapidly in the field of image processing,and are widely used in various fields such as medical treatment and transportation.In order to save labor and time,and improve the efficiency of interpretation of AST,this paper will use deep learning technology to study and apply the problem of image recognition of MTP to automate theinterpretation of AST.In this paper,the recognition of MTP image is divided into two major problems:segmentation of microwells in MTP image and classification of microwells image.Finally,a system called ASTRAS(antimicrobial susceptibility testing result analysis system)is designed and implemented.The specific work is as follows:(1)Establish a micropore image data set.In order to conduct the research in this paper,740 MTP images were taken in the NVMRRAL(National Veterinary Microbial Resistance Risk Assessment Laboratory).These images were made into a microwells image data set,among them,there were 58728 microwell images marked with bacteria growth and microwell images marked without bacteria growth 55,032 photos,totaling 113,760 microwell images.(2)A microwell segmentation algorithm named MTPMS-MM for MTP images is proposed.This algorithm mainly combines the Canny algorithm and mathematical morphology to detect the boundary of the MTP,and then calculates the microwell coordinates according to the invariance of the ratio of MTP to extract the microwell image.The invariance of the MTP means that the ratio of the center position and radius of the microwell to the boundary of the drug sensitive plate is fixed.The experimental results show that the accuracy of using MTPMS-MM algorithm to segment microwells is as high as 98%,which can basically meet the requirements of practical applications.(3)A microporous image classification network model named DCSE-M is proposed.In order to expand the receptive field of the neural network and enhance the effective features learned by the network,this paper combines dilated conv and SE module to optimize the Mobilenet V2.The experimental results show that the improved network model DCSE-M has faster convergence speed and higher accuracy rate,reaching 93.99%,which is 1.19%higher than the original Mobilenet V2 network.At the same time,the precision,recall,and F1 score have also increased by 1.37%,1.15%,1.26%.In order to further improve the accuracy,the network training process is optimized using different parameter settings.The final microwell image classification network model has improved accuracy,precision,recall,and F1 score by 0.79%,0.67%,0.81%,and 0.74% compared with the model without training optimization.(4)Design and implement the AST results analysis system named ASTRAS.The system is implemented using java Fx technology.The system can identify the user-selected MTP image,and give the identification results of 96 microwells,the MIC and the AMR results of each drug.Functional tests were performed on the system,and the test results were as follows: the accuracy rate of single microwell image recognition,MIC accuracy,and AMR accuracy reached 93.66%,92.12%,and 95.01%,respectively.The analysis of the experimental results of an average MTP image took 32 seconds,which was about one-eighth of the manual time.The test results verify the feasibility and effectiveness of the application of deep learning technology to the image recognition of MTP.