文献类型：期刊文献

中文题名：基于多尺度注意力机制的荧光图像分割

英文题名：Fluorescent image segmentation based on multi-scale attention

作者：汤珺[1];曹志兴[1];堵威[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2025

卷号：46

期号：1

起止页码：142

中文期刊名：激光杂志

外文期刊名：Laser Journal

收录：;北大核心:【北大核心2023】；

基金：国家重点研发计划课题(No.2021YFB3301303);国家自然科学基金面上项目(No.62073137、62073144);上海市“科技创新行动计划”自然科学基金(No.22ZR1415300、No.22511104000、No.23S41900500);上海人工智能实验项目。

语种：中文

中文关键词：荧光成像技术;深度学习;图像分割;残差神经网络;注意力机制

外文关键词：fluorescence imaging technology;image segmentation;deep learning;residual convolutional neural networks;attention mechanism

摘要：针对荧光细胞图像分割中细胞轮廓重叠、形态多样等问题,本研究提出了一种结合自适应多尺度注意力机制与边界敏感损失函数的分割算法。首先,为了提升模型对多尺度细胞形态的适应能力,提出了自适应多尺度通道注意力机制,并与特征金字塔结合构建多尺度注意力金字塔结构,提高网络对复杂细胞形状特征的提取能力;其次,设计了一种边界敏感的交叉熵损失函数,通过对细胞边界区域的预测给予更高的权重,增强了网络对细胞边缘的识别精度。实验结果表明,所提方法在荧光细胞图像数据集上的平均Dice系数和平均IoU系数分别高于现有先进模型,证明了本研究方法在荧光图像分割任务中的有效性。
Addressing the challenges of overlapping cell contours and morphological diversity in fluorescence cell image segmentation,this study proposes a segmentation algorithm that integrates an adaptive multi-scale attention mechanism with a boundary-sensitive loss function.Initially,to enhance the model's adaptability to the morphology of cells at various scales,an adaptive multi-scale channel attention mechanism was proposed,which,in conjunction with a feature pyramid,constructs a multi-scale attention pyramid structure,thereby improving the network's ability to extract features of complex cell shapes.Subsequently,a boundary-sensitive cross-entropy loss function was designed,which by assigning greater weights to the prediction of cell boundary regions,enhanced the network's precision in recognizing cell edges.Experimental outcomes indicate that the proposed method surpasses existing advanced models in terms of the average Dice coefficient(mDice)and the average Intersection over Union(mIoU)scores on fluorescence cell image datasets,substantiating the efficacy of the proposed method in the task of fluorescence image segmentation.