摘要
目的在基于病理切片的脑胶质瘤临床诊断中,细胞核密度估计是最关键的任务之一。脑胶质瘤密度估计通常高度依赖细胞核的分割效果,而细胞核的形态多样性与染色差异性易导致语义分割模型的过拟合,最终带来密度估计误差而影响临床决策。鉴于此,本研究提出U-net+DropBlock分割模型驱动的细胞核密度估计方法,通过提升模型的泛化性以降低过拟合风险,获得精确的细胞核密度估计。方法首先,对脑胶质瘤的全切片图像(whole slice image,WSI)进行预处理,包括数据清洗、数据增强、颜色正则化、金标准确立等。然后,利用脑胶质瘤全切片的补丁图像,构建引入DropBlock正则化模块的U-net领域分割模型,称为U-net+DropBlock模型。特别是,DropBlock模块通过随机丢弃特征图中的区域信息,能够削弱特征间的过度空间相关性,以降低分割模型的过拟合风险。最后,对各补丁图像中的细胞核分割结果进行密度计算与可视化,以形成全切片范围的密度热力图。结果与最先进的细胞核领域分割模型相比,本文采用的U-net+DropBlock模型可获得更优的分割性能,确保提出的密度估计方法的准确性。结论本文提出的全域化细胞核密度估计方法,可助力脑胶质瘤临床诊疗的精准化和高效化发展。
Objective In the clinical diagnosis and grading of brain glioma from histopathological slides,wholeslide cell nucleus density estimation is a critical task.This metric is a key biomarker directly correlated with tumor malignancy,proliferative activity,and patient prognosis,as defined by the World Health Organization(WHO)classification system.Glioma density estimation typically relies heavily on the performance of underlying nucleus segmentation.However,segmentation accuracy is challenged by substantial heterogeneity in nucleus morphology and significant staining variations both across slides and within individual specimens.This variability often causes standard semantic segmentation models to overfit the training data,leading to considerable errors in density estimation.Such inaccuracies can compromise downstream pathological assessments,particularly the subjective and time-consuming manual selection of regions of interest(ROI)for grading.To address these limitations,this study aims to develop a precise and robust whole-slide nucleus density estimation method that enhances model generalization and mitigates overfitting,thereby providing an objective,automated tool for glioma analysis.Methods We propose a systematic three-stage pipeline.(1)Preprocessing:whole-slide images(WSIs)of glioma undergo comprehensive preprocessing,including automated data cleaning to discard blurry or artifact-contaminated patches,data augmentation through geometric transformations(e.g.,rotation,flipping)to increase dataset diversity,and color normalization.The latter,based on RGB channel ratios,remaps the color space of all patches to a standardized target,reducing domain shifts caused by staining inconsistencies and improving model robustness.A rigorous semi-automated ground-truth annotation protocol is also implemented,where initial binarization assists annotators in accurately labeling even faint or blurry nuclei,ensuring highquality training data.(2)Segmentation:using the preprocessed patches,we construct a U-net-based segmentation model that incorporates the DropBlock regularization module—here termed U-net+DropBlock.Unlike standard Dropout,which removes individual neurons,DropBlock eliminates contiguous,spatially correlated regions within feature maps.This structural regularization disrupts undesirable spatial dependencies,forcing the network to learn a more distributed and robust feature representation,thereby reducing overfitting.(3)Quantitative analysis:for each segmented patch,density is computed as the ratio of the total nucleus area to the total patch area—a more robust approach than simple nucleus counting,as it accounts for variations in nucleus size.Patch-wise density values are then assembled into a whole-slide density heatmap,offering an intuitive,global overview of tumor cellularity.Results The U-net+DropBlock model was evaluated both quantitatively and qualitatively against state-of-the-art nucleus segmentation methods,including standard U-net and Hover-net.Quantitatively,our model achieved an F1 score of 90.1%,outperforming U-net and Hover-net,which both scored 87.6%.Qualitative analysis confirmed that our method effectively balances precision and recall,substantially reducing the oversegmentation artifacts common with U-net and the under-segmentation issues observed with Hover-net.This enhanced segmentation quality directly improved the accuracy and reliability of the proposed density estimation approach.Conclusion The proposed whole-slide nucleus density estimation method provides a powerful tool for improving the precision and efficiency of glioma diagnosis.By enabling automated,rapid,and objective analysis of cellular density,it overcomes key limitations of manual pathological review.The generated heatmaps allow pathologists to rapidly identify high-density“hotspots”critical for accurate grading and prognostic evaluation,supporting a more standardized and reproducible ROI selection process.This work lays a solid foundation for developing advanced AI-assisted diagnostic systems,paving the way for more precise,efficient,and reproducible glioma assessments in clinical practice.
作者
夏瑞晨
叶臣
赵来定
刘凯
潘敏鸿
姚佳烽
XIA Rui-Chen;YE Chen;ZHAO Lai-Ding;LIU Kai;PAN Min-Hong;YAO Jia-Feng(College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China;School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Department of Pathology,The First Affiliated Hospital with Nanjing Medical University,Nanjing 210029,China;College of Physics and Optoelectronic Engineering,Jinan University,Guangzhou 510632,China)
出处
《生物化学与生物物理进展》
北大核心
2025年第11期2869-2883,共15页
Progress In Biochemistry and Biophysics
基金
国家自然科学基金(62471225)
济纶医工智能科技(南京)有限公司基金(2024外298)
南京邮电大学引进人才科研启动基金(自然科学)(NY223134)
江苏省高水平医院结对帮扶建设专项(JDBFSQ202512)资助。
