EM-stellar

EM-stellar benchmarks state-of-the-art deep learning methods for segmentation of cellular ultrastructures in electron microscopy (EM) datasets, including low-contrast, high-resolution data from electron tomography and serial block-face imaging.

Key Features:

• Benchmarking of deep learning methods: Benchmarks a variety of state-of-the-art deep learning (DL) methods on electron microscopy datasets.
• Segmentation of cellular ultrastructures: Focuses on segmentation of cellular ultrastructures in EM data.
• Low-contrast image handling: Addresses challenges associated with low-contrast EM images.
• Support for high-resolution EM modalities: Applies to high-resolution big-data acquired by electron tomography and serial block-face imaging.
• Comparative performance evaluation: Evaluates and compares DL model performance across quantitative performance metrics, noting that no single DL approach consistently excels across all metrics.
• Reproducibility and methodological comparison: Facilitates methodological comparisons and supports reproducibility and transparency in DL evaluation.
• Implementation: Implemented in Python.

Scientific Applications:

• Benchmarking segmentation algorithms: Evaluates and compares the accuracy of DL-based segmentation algorithms on EM datasets.
• Model selection and optimization: Informs selection and tailoring of DL models to specific EM image properties and datasets.
• Assessment for electron tomography and serial block-face imaging: Assesses DL performance on data from electron tomography and serial block-face imaging modalities.
• Study of performance variability: Provides insight into how image characteristics, such as low contrast, affect DL segmentation outcomes.

Methodology:

Benchmarks multiple state-of-the-art deep learning models on electron microscopy datasets and evaluates model performance using quantitative performance metrics; implemented in Python.