基于Mask R-CNN的柑橘主叶脉显微图像实例分割模型

Instance Segmentation Model for Microscopic Image of Citrus Main Leaf Vein Based on Mask R-CNN

针对目前植物解剖表型的测量与分析过程自动化低,难以应对复杂解剖表型的提取和识别的问题,以柑橘主叶脉为研究对象,提出了一种基于掩膜区域卷积神经网络(Mask region convolutional neural network,Mask R-CNN)的主叶脉显微图像实例分割模型,以残差网络ResNet50和特征金字塔(Feature pyramid network,FPN)为主干特征提取网络,在掩膜(Mask)分支上添加一个新的感兴趣区域对齐层(Region of interest Align,RoI-Align),提升Mask分支的分割精度。结果表明,该网络架构能够精准地对柑橘主叶脉横切面中的髓部、木质部、韧皮部和皮层细胞进行识别分割。Mask R-CNN模型对髓部、木质部、韧皮部和皮层细胞的分割平均精确率(交并比(IoU)为0.50)分别为98.9%、89.8%、95.7%和97.2%,对4个组织区域的分割平均精确率均值(IoU为0.50)为95.4%。与未在Mask分支添加RoI-Align的Mask R-CNN相比,精度提升1.6个百分点。研究结果表明,Mask R-CNN模型对柑橘主叶脉各类组织区域具有良好的识别分割效果,可为柑橘微观表型研究提供技术支持与研究基础。

There is a low efficiency of automatically measuring and analyzing plant anatomic phenotypes currently,which makes it difficult to well deal with the issue of extracting and recognizing the complex anatomical phenotypes.In order to solve this problem,a mask region convolutional neural network(Mask R-CNN)based instance segmentation model for microscopic images of the citrus main leaf veins was proposed.In this model,the deep residual network(ResNet50)and the feature pyramid network(FPN)were used as the backbone feature extraction network.In addition,a new region of interest Align(RoI-Align)layer was added to the Mask branch to improve the segmentation accuracy.The results showed that the network can accurately identify and segment pith,xylem,phloem and cortical cells,respectively,in the citrus main leaf veins.The average precision(IoU was 0.50)of the model for segmentation of pith,xylem,phloem and cortical cells was 98.9%,89.8%,95.7%and 97.2%,respectively,and the overall average precision(IoU was 0.50)for segmentation of the four tissue regions was 95.4%.The mean average precision of Mask R-CNN with adding RoI-Align to the Mask branch was improved by 1.6 percentage points compared with that without.The results showed that Mask R-CNN model presented good performance of recognition and segmentation of various tissue regions of citrus main leaf veins,which can provide technical support for citrus microscopic phenotyping.