分析化学2026，Vol.54Issue(6)：1024-1033,中插27-中插29,13.DOI:10.19756/j.issn.0253-3820.251350
结合频域增强与几何向量场的高密度液滴和微孔阵列芯片图像分割算法
An Image Segmentation Algorithm for High-Density Droplet and Microwell Array Chip Combining Frequency-Domain Enhancement and Geometric Vector Fields
摘要
Abstract
Sample discretization is the core principle of single-molecule absolute quantification technologies based on droplets and micro-well array chips,such as digital polymerase chain reaction(dPCR).Consequently,high-precision image segmentation and counting of microreactors are critical prerequisites for achieving accurate concentration measurements.However,existing image processing algorithms still face dual challenges in precision and versatility,particularly under conditions of high-density microreactors,complex fluorescent backgrounds,and noise interference,where accurate segmentation remains elusive.To address these technical bottlenecks,this study proposes a U-Net model integrating frequency-domain enhancement and geometric vector fields(GVF).By mapping images to the frequency domain via Fourier transform,high-pass filters is utilized to suppress background noise while simultaneously strengthening the contour features of the microreactors.Furthermore,a geometric vector field mechanism is innovatively introduced that predicts the gradient flow field of pixels pointing toward the centroid of each microreactor.This approach enables instance-level precision in separating adhered microreactors,effectively resolving the core failure of existing algorithms under high-density conditions.Experimental results demonstrate that the proposed model achieves an average segmentation precision exceeding 99.9%across various scenarios,including droplet-based and microwell-based array chips,multiple fluorescence channels,and noisy environments.Compared with the industry-leading commercial software Crystal Miner,this model significantly improves the segmentation recall rate for high-density microreactors from below 90%to above 99.9%.This method exhibits exceptional generalization capabilities and robustness,effectively overcoming the challenges of adhesion and noise interference in high-density microreactor imaging.By breaking through the precision bottlenecks of current ultra-high-throughput sample processing,this research provides powerful technical support for high-reliability absolute quantitative detection in complex biological samples.关键词
液滴和微孔阵列芯片/图像分割/样品离散化/绝对定量/深度学习Key words
Droplet and microwell array chip/Image segmentation/Sample dispersion/Absolute quantification/Deep learning引用本文复制引用
杨伟岽,钟少龙,恢嘉楠,冀江毓,李岩,毛红菊..结合频域增强与几何向量场的高密度液滴和微孔阵列芯片图像分割算法[J].分析化学,2026,54(6):1024-1033,中插27-中插29,13.基金项目
国家重点研发计划项目(No.2023YFB3210300)、国家自然科学基金项目(No.62231025)、上海市科技重大专项项目(No.ZD2021CY001)和上海市科学技术委员会项目(No.25JC3201100,24XTCX00700)资助. Supported by the National Key R&D Program of China(No.2023YFB3210300),the National Natural Science Foundation of China(No.62231025),the Shanghai Municipal Science and Technology Major Project(No.ZD2021CY001),and the Program of Science and Technology Commission of Shanghai Municipality(No.25JC3201100,24XTCX00700). （No.2023YFB3210300）
