Anatolution_demo

Introduction: Supervised statistical learning for cell-level segmentation and morphometry in optical microscopy is limited less by algorithmic capacity than by the scarcity of reliable, expert-validated ground truth. In comparative neuroscience and quantitative histology, where classical stains such as Nissl's method remain the primary means to study cellular morphology, this bottleneck is acute: manual annotation is expensive, subject to individual bias, and rarely performed at the scale or consistency that computational approaches demand. No existing platform integrates a stain-specific bioimage segmentation protocol, a structured multi-annotator workflow, and consensus-based quality control into a single pipeline from image ingestion to machine-readable training data. Methods: We present Anatolution, an open-source, web-based platform designed to address the gap of quality annotations at https://anatolution.herokuapp.com/public-tool/. Anatolution organizes microscopy images, including 2D arrays or 3D volumes, into project workspaces where multiple annotators independently label cellular structures against a shared computer vision catalogue. This design enables systematic inter-rater and intra-rater reliability assessment, with consensus derived from agreement across annotators rather than from any single expert's judgment. The platform enables the export of aggregated labels or annotation datasets for downstream statistical learning methods. We describe the system's architecture, its Nissl-specific segmentation pipeline, the consensus annotation workflow, and validation of inter-rater reliability. Conclusion: Across 20+ histological annotation containers annotated by up to 15 independent raters, consensus boundary agreement increased monotonically with annotator count, reaching a median Dice of 0.79 against the full-rater reference at seven annotators, with top-tier containers achieving leave-one-out ceiling values of 0.621–0.769 for cell-body segmentation. The segmentation pipeline provided effective spatial anchoring, with 88% of consensus-annotated polygons containing at least one algorithmically detected seed. Anatolution provides open-source infrastructure for producing consensus-validated training data from classical histological preparations, addressing the primary bottleneck limiting supervised learning for cell-level morphometry.引言：在光学显微成像领域，用于细胞级分割与形态计量的监督式统计学习方法，其发展瓶颈更多源于可靠且经专家验证的标注真值（ground truth）的匮乏，而非算法性能的不足。在比较神经科学与定量组织学领域，诸如尼氏染色（Nissl's method）这类经典染色技术仍是研究细胞形态的核心手段，此时上述瓶颈尤为突出：人工标注成本高昂、易受个体偏倚影响，且难以达到计算方法所需的标注规模与一致性。目前尚无平台能够将染色特异性生物图像分割协议、结构化多标注者工作流以及基于共识的质量控制，整合为从图像导入到可用于机器学习的训练数据的完整流程。 方法：我们研发了Anatolution——一款开源的网页式平台，旨在解决高质量标注的缺口问题，其部署地址为https://anatolution.herokuapp.com/public-tool/。Anatolution将各类显微图像（包括二维阵列与三维体数据）整合至项目工作区中，多名标注者可基于共享的计算机视觉目录独立标注细胞结构。该设计支持开展系统性的标注者间与标注者内信度评估，其共识结果基于多位标注者的一致性判断，而非依赖单一专家的主观定论。该平台支持导出聚合后的标注结果或标注数据集，以供下游的统计学习方法使用。本文详细阐述了该系统的架构、适配尼氏染色的专用分割流程、共识标注工作流以及标注者间信度验证方案。 结论：在由最多15名独立标注者完成标注的20余个组织学标注数据集中，共识边界一致性随标注者数量单调提升，当标注者数量为7时，与全标注者参考结果相比的戴斯系数（Dice）中位数达到0.79；优质数据集的细胞体分割留一法上限值可达0.621~0.769。该分割流程可实现有效的空间锚定，经共识标注的多边形中，有88%至少包含一个算法检测得到的种子点。Anatolution为从经典组织学样本中生成经共识验证的训练数据提供了开源基础设施，解决了制约细胞级形态计量监督学习发展的核心瓶颈。