Research On Zebrafish Eggs And Larvae Microscopic Image Analysis Algorithm Based On Deep Learning

Zebrafish is a frequently used animal model in biological experiments.Currently,the analysis of zebrafish microscopic image still extensively relies on human labor,which suffers from low efficiency,imperfect accuracy and poor objectivity.In this study,we use deep learning methods to realize automated detection and classification of zebrafish eggs and larvae from microscopic images.This thesis focuses on solving the problems of scare training images,imbalanced training data number of different classes,fuzzy inter-classes differences of image appearance and multi-class property of the target objects.This thesis also proposes a solution for integrating deep learning device into the conventional optical microscopic imaging system to realize onsite real-time data analysis.This thesis has the following contributions:(1)The algorithm of zebrafish egg detection and classification from optical microscopic images.This study proposes an algorithm for zebrafish egg fertilization status classification based on the combination of statistical shape model and convolutional neural network.To tackle the problem of insufficient training image,we designed a dedicated data augmentation strategy which avoids the overfitting issue of training deep network using small dataset.We also used the global pooling layers instead of conventional softmax classifiers to copy with the fuzzy inter-class differences.Experiment results proved that the average classification accuracy has been improved to 95%.(2)The algorithm of zebrafish larvae development phenotype classification.This algorithm is developed to assist the study of drug effect on zebrafish larvae development.To solve the problems of fuzzy inter-class difference,we adopted the separable convolution structure to improve the network's multi-class recognition ability.To cope with the multi-label nature of the zebrafish larvae,we also designed a two-tier classification strategy and use different tiers of networks to classify the phenotypes with different recognition difficulties.Besides,specialized classification networks are trained for each phenotype which is difficult to classify.Our method effectively improves the classification accuracy and outperformed other state-of-the-arts methods based on the same open source test dataset.(3)The method for integrating deep learning algorithms into conventional optical microscopic devices.To facilitate the usage of deep learning algorithms in daily biological experiments,it is necessary to integrate the deep learning algorithms into the conventional optical microscopic imaging workflow.We used a portable computing device to upgrade the conventional optical microscopy system and adopted a combined-detection-classification network architecture which fits the portable computation device.Our system performed online zebrafish microscopic image analysis with a processing rate of 3 frames per second,a sensitivity of 0.88 and a specificity of 0.94.This method provides a low-cost and convenient solution for the AI-upgrading of conventional microscopic devices.This thesis solves a series of technical problems for applying deep learning methods in zebrafish microscopic image analysis and proposes a hardware upgrading method for popularizing deep learning methods in daily biological experiments.Although focused on zebrafish,the methods of this study may also be extended for the analysis of other types of microscopic images.