The Forensic Application Of Automation Diatom Test Based On Artificial Intelligence Deep Learning

Object Diatom test is one of main method to help forensic diagnose death cause of body in water.And forensic experts can also detect the site of drowning according to diatom species distribution.However,present diatom tests have disadvantages in difficult universal experiment or time-consuming.Artificial intelligence(AI)is a technology that use computer to study and simulate human behavior for completing the task more efficiently.Deep learning,an algorithm that realize AI,can simulate human brain to process data.Because of its superior performance in image processing and other aspects,it has been widely used in medical auxiliary inspection,and there are related researches in forensic field.In this study,we trained artificial intelligence deep learning model to automatically identify and classify diatoms based on traditional diatom chemical test,and verified its ability of automatic identification in actual cases.The purpose is to test the feasibility of an automated diatom test system based on AI technology.Methods We developed and tested a diatom automation test system based on AI technology.The research was divided into three parts.In the first part,we used diatom-none database to train convolutional neural networks(CNNs)and developed a diatom automation identification model.Subsequently,we built an artificial intelligence-based diatom automatic recognition system,and compared the recognition capabilities of the system with human experts.In the second part,the AI diatom automatic test system was used to predict 10 actual cases of forensic science(9 drowning and 1 non-drowning),and the prediction results were analyzed.In the third part,diatom classification database was established by collecting water samples from five sites of Huangpu River and Suzhou River in Shanghai.And we used the database to train CNNs model for classifying diatom species.Then we established an experimental model of drowning animals.Moreover,we used the identification and classification model to identify and classify diatoms in rat lungs.Finally,we combined statistical methods to calculate the similarity between the diatom distribution in the rat lung and in water samples for detecting the site of drowning.Results In the first stage,we developed a diatom automation test system based on AI technology.The optimal CNNs model was obtained from augmented database(including 6585 training sets and 1,545 validation sets)and transfer learning.The accuracy rate of best model reached 97.67% and its AUC reached 99.51%.The Grad-CAM images shown the proportion of images that the model can successfully recognize is as high as 89.6%.The AI diatom automation test system was faster than the speed of human experts to identify diatoms.The AI system spent only for nearly 4 hours,while the human experts took an average of 6-7 hours.In the second stage,we optimized the CNNs model in previous stage by expanding database,so that accuracy rate of the best model was 99.16%.The overall precision of diatoms identification was as high as 92.45% in the application of 10 actual cases.In the third stage,we developed the diatom automatic classification technology based on AI.The accuracy rate of the best six-classes model reached 96.4%,and the overall AUC value reached 97.1%.In the animal experiment,the result of statistical method using detecting the site of drowning showed that two animal models failed to predict the correct drowning site,with an accuracy of 80%.Conclusions In this paper,we combined traditional diatom test with AI deep learning technology to develop the diatom automatic identification and classification system to perform diatom automatic test and detection of drowning sites.1.The AI diatom automatic identification system had shown superior performance over human experts in the task of identifying diatoms.That has initially confirmed the feasibility of the diatom automatic identification model.2.In the actual cases,the AI diatom automatic identification system could assist forensic experts to diagnose the cause of drowning,while it could improve the efficiency of diagnosis.3.The method for detection of drowning site using model prediction result with statistics was verified on the animal experiment.