CEM500K, a large-scale heterogeneous unlabeled cellular electron microscopy image dataset for deep learning

Area of Science:

• Cellular Biology
• Microscopy
• Artificial Intelligence

Background:

• Automated segmentation of cellular electron microscopy (EM) data is difficult.
• Supervised deep learning (DL) models require extensive region-of-interest (ROI) annotations and do not generalize well to new datasets.
• Unsupervised DL methods need pre-training data, but existing EM datasets are large, homogeneous, computationally expensive to train on, and offer limited value for diverse biological contexts.

Purpose of the Study:

• To develop a more effective pre-training strategy for deep learning (DL) models in cellular electron microscopy (EM) segmentation.
• To create a versatile and computationally efficient dataset for pre-training EM segmentation models.
• To enhance the generalization capabilities and performance of DL models across various EM datasets.

Main Methods:

• Curated a novel dataset, CEM500K, comprising 0.5 million unique 2D cellular EM images from diverse sources.
• Utilized CEM500K for pre-training DL models, focusing on feature learning resilient to image augmentations.
• Evaluated transfer learning performance of CEM500K pre-trained models on multiple benchmark EM segmentation tasks.

Main Results:

• Models pre-trained on CEM500K learned biologically relevant features.
• Pre-trained models demonstrated robustness against image augmentations.
• State-of-the-art results were achieved on six public and one new benchmark EM segmentation tasks using transfer learning from CEM500K.

Conclusions:

• CEM500K provides a computationally efficient and effective resource for pre-training DL models in EM.
• Pre-training on CEM500K significantly improves model generalization and segmentation performance.
• The CEM500K dataset, pre-trained models, and curation pipeline are released to benefit the EM research community.