基于稠密块和注意力机制的肺部病理图像异常细胞分割  

Abnormal cell segmentation for lung pathological image based on denseblock and attention mechanism

作　　者：崔文成[1] 王可丽 邵虹[1] CUI Wencheng;WANG Keli;SHAO Hong(School of Information Science and Engineering,Shenyang University of Technology,Shenyang 110870,China)

机构地区：[1]沈阳工业大学信息科学与工程学院,辽宁沈阳110870

出　　处：《智能科学与技术学报》2023年第4期525-534,共10页Chinese Journal of Intelligent Science and Technology

摘　　要：针对肺部细胞病理图像亮度不均衡、异常细胞轮廓精准分割难以实现的问题,提出一种以U-Net为基本框架,结合稠密块以及注意力机制的异常细胞分割模型。首先,利用具有编码器-解码器结构的U-Net对异常细胞进行分割;然后,在U-Net中引入稠密块,以提高特征之间的传播能力,提取更多异常细胞的特征信息;最后,利用注意力机制提高异常细胞区域的权重,降低亮度不均衡对模型的干扰。实验结果表明,该方法的IoU和Dice相似系数值分别为0.6928和0.8060,与其他模型相比,提出的方法能够分割出低对比度区域和形状多样的异常细胞。Aiming at the problems of unbalanced brightness of lung cell images and achieving accurate segmentation of abnormal cell contour difficultly,an abnormal cell segmentation model based on U-Net was proposed,which combined the dense connection mechanism and attention mechanism.Firstly,U-Net with encoder-decoder structure was used to segment abnormal cells.Secondly,the dense block was introduced into U-Net to improve the propagation ability between features and extract more characteristic information of abnormal cells.Finally,the attention mechanism was used to increase the weight of abnormal cell regions and reduce the interference of the imbalance of brightness to the model.The experimental results show that the IoU value and Dice similarity coefficient achieved by this method are 0.6928 and 0.8060,respectively.Compared with other models,this proposed method is able to segment low-contrast regions and abnormal cells with diverse shapes.

关 键 词：肺部细胞病理图像 细胞分割 U-Net 稠密块 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]