基于改进Mask-Scoring R-CNN的肌纤维自动分割与表型计算方法研究  

Automatic segmentation of muscle fiber and methods of calculating phenotype based on improved Mask-Scoring R-CNN

作　　者：沃靖杰 田绪红[1] 尹令[1,2,3] 杨杰 姚泽锴 蔡更元[2,3,4] WO Jingjie;TIAN Xuhong;YIN Ling;YANG Jie;YAO Zekai;CAI Gengyuan(College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642,China;National Engineering Research Center for Swine Breeding Industry,Guangzhou 510642,China;State Key Laboratory of Swine and Poultry Breeding Industry,Guangzhou 510640,China;College of Animal Science,South China Agricultural University,Guangzhou 510642,China)

机构地区：[1]华南农业大学数学与信息学院,广州510642 [2]国家生猪种业工程技术研究中心,广州510642 [3]猪禽种业全国重点实验室,广州510640 [4]华南农业大学动物科学学院,广州510642

出　　处：《华中农业大学学报》2025年第2期134-144,共11页Journal of Huazhong Agricultural University

基　　金：国家自然科学基金项目(32172780);国家重点研发项目(2023YFD1300202)。

摘　　要：为解决人工手动分割与半自动分割的精度及效率问题以及通用分割模型在面对各种噪声干扰时的表现不足,提出改进Mask-Scoring R-CNN的实例分割模型,实现对肌纤维细胞的高效分割。在Mask-Scoring R-CNN模型中引入CBAM(convolutional block attention module)注意力机制,并对其进行改进,强化模型对特征信息的提取与表达,从而提升分割效果与模型在肌纤维分割任务中的泛化能力。改进Mask-Scoring RCNN模型在103张测试集的测试结果显示,表型数据测定值的均方根误差均比原模型更小,肌纤维总数均方根误差从2.08降至1.26,面积均方根误差从212.21μm^(2)降低至181.36μm^(2),平均直径均方根误差从2.87μm降低至1.47μm。试验结果表明改进后的模型能有效应对含噪声的肌纤维图像,在常见的噪声环境下依然能够准确分割出每个肌纤维。A model for instance segmentation based on improved Mask-Scoring R-CNN was proposed and the efficient segmentation of myofibroblast cells was realized to solve the problems of manual and semi-automatic segmentation with accuracy and efficiency and the inadequate performance of general models for segmentation in encountering various interferences of noisy images.The Convolutional Block Attention Module(CBAM) attention mechanism was introduced into the Mask-Scoring R-CNN model to improve the model.The extraction and expression of feature information by the improved model was enhanced to improve the performance of segmentation and the generalization capability of the model in tasks of segmentation.The results of testing the improved Mask-Scoring R-CNN model on a dataset of 103 test images showed that the root mean square error(RMSE) of phenotype measurement value was smaller than that of the original model,with the RMSE of the total number of myofibers decreased from 2.08 to 1.26,the RMSE of area reduced from 212.21 μm^(2) to 181.36 μm^(2),and the RMSE of average diameter decreased from 2.87 μm to 1.47 μm.It is indicated that the improved model can effectively deal with noisy images of myofiber and accurately segment each myofiber even in common noisy environments.

关 键 词：实例分割 Mask-Scoring R-CNN 猪肉肌纤维表型 细胞分割 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]