Design And Implementation Of Feature Recognition System For Embryo Cleavage Stage Based On Time Series Images

In the field of reproductive medicine,the analysis of the developmental status of embryos produced by artificial insemination is mainly based on the images collected by time-lapse photography equipment.Embryo development needs to go through several periods.The level of characteristic development in each period has an important impact on the final embryo transfer.At present,doctors mainly rely on naked eyes and paper records from a large number of embryo development images,which is very costly.Doctor’s human resources,and there are problems such as untimely observation and omission.Therefore,there is an urgent need for an automated feature recognition software to reduce the burden on doctors.The main characteristics of embryo development identified in this thesis are the number of blastomeres,fragment ratio grades in the cleavage stage.Sample data in the cleavage stage of the embryo has problems such as small data volume,redundancy,uneven sample distribution between classes,and small differences in characteristics between classes.These are the important factors that restrict the accuracy of computer vision-related algorithm model recognition,and the existing model methods cannot achieve good results.This thesis designs and implements an embryo cleavage stage feature recognition method based on time series images,which effectively solves the above problems and improves the usability of the system.The main research contents of this thesis are as follows:1.Research the preprocessing method of embryo image data.This thesis first extracts the main feature area of the embryo based on the UNet3+ semantic segmentation model,so that the embryo feature that needs to be recognized occupies the main area of the image,avoids impurity interference,and improves the usability of the data;then the sample data enhancement is used to increase the number of training samples to relieve the embryo The problem of the small amount of medical sample data;finally,the balanced sampling of samples is used to alleviate the problem of imbalanced feature distribution among feature classes during embryonic cleavage.2.Research on the feature recognition algorithm of embryo cleavage stage based on time series images.This thesis first compares the recognition effects of various traditional classification models on the number of embryonic blastomeres and fragment ratio levels,and unearths the potential connection between medical features and model concerns.The SE-Res Net model is selected to determine the number of embryonic blastomeres and eggs.Fragment ratio grades in the cleavage stage were identified,and a good benchmark effect was achieved;then,the T3 D time series network model was designed and improved,and the number of embryonic blastomeres and fragment ratio grade categories were used as "behavior" categories to explore the embryonic development process The existing time sequence effectively solves the problem of a large number of sample data redundancy and unavailability and small sample data volume;finally,based on the idea of integrated learning,SE-Res Net and T3 D are integrated to supplement the definition of "chronological behavior" and the improvement model Overall recognition accuracy.3.Design and implement an embryo cleavage stage feature recognition system based on time series images.Combining the specific needs of users,this thesis uses the framework django to implement the system,split the business,reduce the coupling while ensuring efficiency,and also provide convenience for the addition and modification of future functions.