MitoEM

电子显微镜 (EM) 可以鉴定细胞内细胞器，例如线粒体，为临床和科学研究提供见解。但是，公共线粒体分割数据集仅包含数百个具有简单形状的实例。目前尚不清楚在这些小数据集上实现人为准确性的现有方法在实践中是否可靠。为此，我们介绍了MitoEM数据集，这是一个3D线粒体实例分割数据集，分别具有来自人和大鼠皮质的两个 (30 µ m)3个体积，比以前的基准大3个，600个。在大约40k的实例中，我们发现线粒体在形状和密度方面有很大的多样性。为了进行评估，我们以45的加速速度定制了3D数据的平均精度 (AP) 度量的实现。在MitoEM上，我们发现现有的实例分割方法通常无法正确分割形状复杂或与其他实例紧密接触的线粒体。因此，我们的MitoEM数据集对该领域提出了新的挑战。

Electron microscopy (EM) enables identification of intracellular organelles such as mitochondria, providing insights for clinical and scientific research. However, public mitochondrial segmentation datasets only contain hundreds of instances with simple shapes. Currently, it remains unclear whether existing methods that achieve human-level accuracy on these small datasets are reliable in practical scenarios. To address this gap, we introduce the MitoEM dataset, a 3D mitochondrial instance segmentation dataset comprising two 30 μm³ volumes derived from human and rat cerebral cortex respectively, which is 3,600 times larger than previous benchmarks. With approximately 40,000 instances in total, we observe significant diversity in mitochondrial shape and density. For evaluation purposes, we developed a customized implementation of the 3D average precision (AP) metric, which achieves a 45× speedup. Evaluated on MitoEM, we find that existing instance segmentation methods often fail to correctly segment mitochondria with complex shapes or those in close contact with other instances. Thus, our MitoEM dataset poses novel challenges for the field.

搜集汇总

背景与挑战

背景概述

MitoEM是一个3D线粒体实例分割数据集，包含来自人和大鼠皮质的两个体积，规模比以前的基准大3600倍，具有约40k个实例，展示了线粒体在形状和密度上的多样性。该数据集旨在评估现有实例分割方法在复杂形状或紧密接触的线粒体上的性能。