基于自然图像模型微调的小鼠脑部电镜图像实例分割  

Instance segmentation of mouse brain scanning electron microscopy images based on fine-tuning nature image model

作　　者：承骜 赵国强[2] 张若冰[2] 王丽荣[1] CHENG Ao;ZHAO Guoqiang;ZHANG Ruobing;WANG Lirong(Soochow University,School of Electronic and information,Suzhou 215000,China;Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Sciences,Suzhou 215000,China)

机构地区：[1]苏州大学电子信息学院,江苏苏州215000 [2]中国科学院苏州生物医学工程研究所,江苏苏州215000

出　　处：《光学精密工程》2024年第18期2836-2845,共10页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.32271430);国家重点研发计划资助项目(No.2023YFF0715904)。

摘　　要：分割模型的准确性和鲁棒性是小鼠脑电镜图像处理中的主要考虑因素。针对电镜图像的技术特点,提出了高度稳健的二维分割算法,准确识别每张切片中各物体的形态结构。本文提出了基于大型自然图像模型的主干网络微调的大尺度二维电镜图像分割模型EM-SAM,用于脑部电镜图像中的实例分割。模型主干网络采用大型自然图像模型SAM中的已训练完成的图像编码器,在电镜图像处理任务中最大化模型提取图像特征的能力。此外,模型采用了U型的解码器设计,并通过小鼠脑电镜图像分割任务进行微调。实验结果表明:在公开数据集SNEMI3D中A-Rand可达到0.054;在公开数据集MitoEM中AP-50和AP-75分别可达到0.883,0.604。EM-SAM在电镜图像神经分割任务中准确性高、鲁棒性强,并且可针对不同任务进行微调。The accuracy and robustness of segmentation models are critical considerations in the processing of mouse brain electron microscopy images.We proposed a highly robust two-dimensional segmentation algorithm tailored to the technical characteristics of electron microscopy images,aiming to accurately delineate the morphological structure of cells in each slice.Aiming at accurately delineate the morphological structure of cells in each section,a highly robust two-dimensional segmentation algorithm based on natural image model tailored to the technical characteristics of electron microscopy images was proposed.EMSAM was based on fine-tuning the backbone network of pre-trained large natural image model SAM for maximizing the capability of features extraction.The model employed the image encoder from the SAM architecture,augmented with a U-shaped decoder,and was fine-tuned specifically for the segmentation of mouse brain electron microscopy images.Experimental results demonstrate that A-Rand achieves 0.054 on public dataset SNEMI3D.Additionally,AP-50 and AP-75 reach 0.883 and 0.604,respectively,on public dataset MitoEM.EM-SAM exhibits high accuracy and robustness in neural segmentation tasks of electron microscopy images,and it can be fine-tuned for different tasks.

关 键 词：深度学习 分割 大模型 电镜图像 小鼠脑部 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]