Research On Feature Extraction Method Based On Mathematical Representation Of C.elegans Shape

The phenotype of C.elegans is the result of the joint action of internal genes and environment,and the accurate extraction of phenotypic features is of great significance for the study of nematode movement behaviour,physiological changes,and lifespan prediction.This article proposes an efficient,low-cost,and automated method for extracting nematode phenotypic features based on shape mathematical representation and image processing technology,and proposed a pose prediction and nematode image segmentation model based on mathematical representation of C.elegans shape.The segmentation of nematode microscopic images is a prerequisite for the dynamic tracking and behavior analysis of C.elegans.However,due to the problems of noise,similarity of edge pixels and surrounding environment in nematode microscopic images,traditional image segmentation algorithms cannot meet the requirements of highprecision segmentation.Therefore,we propose an improved deeplabv3+ algorithm.The specific improvement method is to use a resnet50 network with low model parameter requirements in the backbone network,and in conv2＿x and conv3＿x,use the 1x1 convolution to extract underlying features.Then the structure of the encoder-decoder was modified.In the encoder stage,deep separable convolutions were used instead of ordinary convolutional kernels for multi-scale feature extraction of nematodes.In the decoder,the IAFF model was used to fuse the conv2＿x and conv3＿x of the Resnet50 network.Then low-level features extracted are concatenated with high-level features after two double upsampling operations,paying more attention to the fine details of nematodes(edge features).The experimental results show that the segmentation model proposed in this paper can improve the accuracy of nematode segmentation and demonstrate sufficient stability and robustness in nematode microscopic image segmentation tasks.In this thesis,we investigated the design of a mathematical representation method for nematode shape and its application in phenotype extraction.Firstly,smooth the image to remove noise and interference from other pixels.Identify the head of the nematode in each preprocessed image frame,extract the skeleton of the nematode,and use it to describe the shape of the nematode.Divide the main curve into 60 equal parts,calculate the angle between each arc segment and the horizontal plane,obtain a digital vector of the nematode shape,and standardize it to reduce the impact of image rotation and other factors on the results.This shape vector can reduce the dimensionality of nematode behavior analysis,be used for nematode behavior analysis,activity mode analysis,behavior comparison between different nematodes,and can also use shape reconstruction algorithms to restore the shape of C.elegans.Then,with the help of the parametric vector of mathematical representation of the shape of model organism,the body bending degree,the number of head swings,angular velocity,motion frequency and other phenotypes of the nematode are extracted,which are used to quantify and analyze the behavior and motion differences of the nematode.BP classification algorithm is used to predict the nematode under different strains and motion states.According to the experimental results,it can be concluded that this method is feasible and effective.The behavior of nematodes is composed of a series of basic motion modules,which have the characteristics of periodicity and repeatability.Therefore,in this article,we propose in detail a pose prediction method based on mathematical representation of nematode shapes,Time2vec-LSTM-Attention.Firstly,the shape of the nematode is mathematically represented,and the shape vectors obtained from each frame(including60 normalized angles)are composed into sequence data in chronological order.Secondly,in terms of network model design,we use the Time2 vec model to embed time features for each time step in the sequence to capture the periodic and non-periodic posture changes of nematodes.Finally,the embedded time feature sequence data is used to predict the future short-term pose of nematodes,and the predicted results are analyzed using root mean square error and shape reconstruction algorithms.According to the conclusions drawn from the experiment,it can be seen that the model performs well in the pose prediction task of the online worm,and the prediction speed is faster.