Research On Corneal Image Acquisition And Turbidity Classification Algorithm

Objective: Objective and accurate inference of postmortem interval is the focus and difficulty of forensic research,having great significance on solving major criminal cases.Corneal turbidity is an important change in the early stage of death and has always been applied to infer postmortem interval in the field of forensic medicine.The corneal turbidity changes regularly with the prolongation of postmortem interval,which is reflected in the digital images as variations of image color and texture.The purpose of this study is to design an image acquisition device dedcicated to cornea,and based on variation patterns of corneal turbidity,a multi-classification model is established to objectively quantify it,and then the model is solidified into the acquisition device for real-time corneal turbidity analysis and postmortem interval inference of the deceased.Methods: This study constructed classification model based on image processing,feature engineering,machine learning and other technologies,achieving objective quantification of corneal turbidity and the algorithm was cured into hardware equipment for real-time analysis.First,a corneal image acquisition device was created,meanwhile,for handy operation and external light source isolation,we also designed optical lens barrel and manipulation software for the device.Then we segmented corneal regions from the original images by exploiting the auto-extraction algorithm designed by our study,and extracted their color and texture features based on prior experiences,and then removed the unrelated and redundant features by feature selection.Based on the Support Vector Machine,Naive Bayes,K-nearest Neighbor,Decision Tree,Random Forest and Adaboost 6 algorithms,multi-classification models were established by using optimized feature set.We used Precision,Sensitivity and F1 score to evaluate performance of those models,selecting the model with best property,and finally solidified it into image acquisition device to analyze corneal turbid image in real time.Results: After four pigs died,corneal images of their left eyes were collected by the device designed by our study,and the images were clear with no reflective points.For each acquired picture,a total of 1097 features were extracted,then by using feature selection 971 unrelated and redundant features were removed while 126 relevant features were retained,so the original feature set was optimized.Among the six multi-classification models,the Adaboost model performed best with an average accuracy of 0.981 in classifying no turbidity,mild turbidity,moderate turbidity and serious turbidity,achieving objective quantification of corneal turbidity.Conclusions: The classification algorithm designed in this study can classify corneal turbidity based on image features of cornea after death,and then infer the postmortem interval.The image acquisition device devised in our study can isolate external light sources,avoiding reflective points formation,and it can invoke the classification algorithm through manipulation software to analyze corneal turbidity,integrating the functions of images collection and real-time analysis,which is of great significance to speeding up the inference of the postmortem interval.