基于改进的SAM免疫组化细胞分割算法研究  

Research on Cell Segmentation Algorithm Basedon Improved SAM Immunohistochemistry

作　　者：余一聪 何领朝 林凯鑫 洪家军 YU Yi-cong;HE Ling-chao;LIN Kai-xin;HONG Jia-jun(College of New Engineering Industry,Putian University,Putian 351100,Fujian,China;Engineering Research Center of Big Data Application in Private Health Medicineof Fujian Province Universities,Putian 351100,Fujian,China;Fuzhou Institute of Data Technology,Fuzhou 350200,China)

机构地区：[1]莆田学院新工科产业学院,福建莆田351100 [2]民营健康医疗大数据应用福建省高校工程研究中心,福建莆田351100 [3]福州数据技术研究有限公司,福建福州350200

出　　处：《兰州文理学院学报(自然科学版)》2025年第2期48-53,共6页Journal of Lanzhou University of Arts and Science(Natural Sciences)

基　　金：福建省本科高校教育教学研究项目(FBJY20230216);莆田学院科研项目(2023043);福建省中青年教师教育科研项目(JAT242008)。

摘　　要：医学领域的分割任务常需要使用专门的深度学习模型,Segment Anything Model(SAM)可实现目标对象的有效分割,而SAM在医学领域的适用性尚未得到充分探索.为进一步挖掘SAM模型在医学领域的应用潜力,本研究将免疫组化数据的点标注转为掩码标签,提出一种实用的损失函数加权组合策略,并应用LoRA技术改进SAM模型,在掩码解码器旁增加旁路,采用先降维后升维的操作方式模拟内在秩的实现机制,冻结部分预训练参数,使其以最小交互生成稳健的分割结果.实验结果表明,该方法能有效提升SAM模型在细胞图像分割领域的性能,准确率提升了近6.3%,F1分数提升了近4.4%.Segmentation tasks in the medical field often require the use of specialized deep-learning models,Segment Anything Model(SAM)can achieve effective segmentation of target objects,but the applicability of SAM in the medical field has not been fully explored.In order to further explore the application potential of SAM model in the medical field,this study converts the point labeling of immunohistochemical data into mask labels,proposes a practical loss-function weighted combination strategy,applies LoRA technology to improve the SAM model,adds bypass beside the mask decoder,simulates the implementation mechanism of internal rank by the operation mode of first reducing dimension and then increasing dimension,and freezes some pre-training parameters to generate robust segmentation results with minimal interaction.The experimental results show that this method can effectively improve the performance of SAM model in the field of cell image segmentation,the accuracy rate is improved by nearly 6.3%,and the F1 score is improved by nearly 4.4%.

关 键 词：改进SAM 细胞图像分割 损失函数设计 掩码解码器 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]