Three-dimensional Phenotype Identification And GWAS Analysis Of Maize Kernels Based On CT Images And Deep Learning

Corn is one of the highest yield food crops in the world,and it is also one of the three major food varieties in our country.The corn kernel contains internal structures such as embryo,endosperm and cavity,among which endosperm can be divided into silty endosperm and horny endosperm according to the hardness.Quantitative analysis of corn kernel structure components is crucial for precise breeding of corn varieties and improving the edible value and market value of corn varieties.At present,conventional corn seed testing can be only based on machine vision or manual measurement and obtain grain length,grain width,grain thickness,and 100-grain weight,but cannot accurately obtain the three-dimensional and internal structural components of the grain,such as grain,embryo,endosperm and cavity volume.Micro-CT is a non-destructive 3D imaging technology that can clearly,accurately and intuitively display the internal structure,composition,material and defect status of corn kernels in the form of two-dimensional or three-dimensional images.However,automated,high-precision quantitative phenotyping methods and software based on CT images of maize kernels are still lacking.In this study,a corn kernel CT image annotation dataset was constructed,a deep learning 2D/3D semantic segmentation network was developed,and an automated CT image kernel phenotype extraction pipeline was constructed to accurately extract and evaluate the traits of the kernel and its internal structural components.Three-dimensional phenotyping software for maize kernels,and applied to phenotypic evaluation and GWAS analysis of a large number of maize kernel varieties.The specific research work is as follows:1)In the aspect of maize grain identification,a high-throughput grain scanning and phenotype extraction method was constructed,and 23 phenotypic indicators that could characterize the differences in grain varieties were proposed.Firstly,low-resolution scanning(1000×1000)of densely placed grains was carried out by Micro-CT,and single grain was split by 3D watershed algorithm;then,the grain embryos were marked with ITK-SNAP to construct a 3D dataset of grain embryos for use in the process of training and testing of the 3D segmentation network RAUNet;and 3D image processing technology was used to segment semantic objects such as grain embryo,endosperm,and cavity,and a total of 23 related phenotypic parameters were extracted.On the test set,the DICE accuracy of RAUNet was 93.4%,and the R~2of grain length,grain width and grain thickness are 0.902,0.926 and 0.904,respectively,compared with the artificially measured values.Statistical analysis among varieties and identification of varieties were performed based on the extracted phenotypic indicators,and the identification accuracy of SVM was the highest,which was 90.4%.2)In terms of GWAS analysis of maize grains,an extraction method of grain fine structure was constructed,and the genetic research of grain structure traits was carried out by combining CT phenotype technology and GWAS technology.First,Micro-CT was used to scan the maize association analysis group grain with high resolution(2000×2000),and Vgg UNet was used to segment multiple embryos and grains in parallel;based on the high-resolution images,the endosperm was further decomposed into farinaceous endosperm and keratinous endosperm,and the cavities were subdivided into three categories:subcutaneous cavities,embryonic cavities and endosperm cavities;based on the extracted phenotypic indicators,the grain structure of the association analysis population was grouped into three categories,and further GWAS analysis was performed on embryo volume,embryo ratio,endosperm volume,endosperm ratio,cavity volume and cavity ratio,and the genes that control the development of grain structure were mined.Among them,many genes encoded proteins,such as b HLH,whose functions have been reported by predecessors.3)In terms of software design,a three-dimensional phenotype analysis software based on CT images has been designed and developed,which supports batch processing,result browsing,phenotype quality control and model training functions,and supports rendering segmentation results,which greatly improves the efficiency of grain three-dimensional phenotype acquisition based on CT images.