Automated Segmentation of Intracellular Substructures in Electron Microscopy (ASEM) on AWS

The Automated Segmentation of intracellular substructures in Electron Microscopy (ASEM) project provides deep learning models trained to segment structures in 3D images of cells acquired by Focused Ion Beam Scanning Electron Microscopy (FIB-SEM). Each model is trained to detect a single type of structure (mitochondria, endoplasmic reticulum, golgi apparatus, nuclear pores, clathrin-coated pits) in cells prepared via chemically-fixation (CF) or high-pressure freezing and freeze substitution (HPFS). You can use our open source pipeline to load a model and predict a class of sub-cellular structures in naive FIB-SEM cells images. If required, a fine-tuning procedure allows a model to be trained on a small amount of additional ground truth annotations to improve a prediction on a naive dataset. Together with the trained models, we also provide the training, validation and test datasets.

自动分割电子显微镜细胞内亚结构的自动化分割（ASEM）项目提供了一系列深度学习模型，这些模型经过训练，能够对通过聚焦离子束扫描电子显微镜（FIB-SEM）获取的细胞三维图像中的结构进行分割。每个模型均针对检测单一类型结构（如线粒体、内质网、高尔基体、核孔、网格蛋白包被的凹陷）进行训练，这些细胞通过化学固定（CF）或高压冷冻与冷冻替代（HPFS）方法制备。您可利用我们提供的开源流程加载模型，并在原始聚焦离子束扫描电子显微镜细胞图像中预测亚细胞结构类别。如需进一步优化，可通过微调程序对模型进行训练，使其基于少量额外的真实标注数据进行训练，从而提升在原始数据集上的预测准确性。此外，我们还提供了训练、验证和测试数据集。