Deep learning based domain adaptation for mitochondria segmentation on EM volumes.

Area of Science:

• Neuroscience
• Computational Biology
• Biomedical Imaging

Background:

• Accurate segmentation of electron microscopy (EM) volumes is crucial for understanding neuronal structures.
• Supervised deep learning methods for EM segmentation require extensive annotated data and struggle with domain adaptation.
• Domain adaptation is challenging for EM datasets due to variations in tissues and species, impacting model performance on new data.

Purpose of the Study:

• To address the challenge of domain adaptation in deep learning-based mitochondria segmentation for EM datasets.
• To develop and evaluate unsupervised domain adaptation strategies for improved cross-dataset segmentation performance.
• To introduce a novel training stopping criterion for enhanced model generalization without validation labels.

Main Methods:

• Three unsupervised domain adaptation strategies were developed: style transfer, self-supervised learning with fine-tuning, and multi-task neural networks.
• Models were trained end-to-end using both labeled source and unlabeled target images.
• A morphology-based stopping criterion was proposed for model generalization, derived from source domain priors.

Main Results:

• Cross-dataset experiments were conducted using three public EM datasets to evaluate proposed strategies.
• The proposed unsupervised domain adaptation methods demonstrated superior performance compared to baseline approaches.
• The morphology-based stopping criterion effectively selected optimal models on average.

Conclusions:

• The introduced unsupervised domain adaptation strategies significantly improve mitochondria segmentation across diverse EM datasets.
• The proposed methods achieve state-of-the-art performance in unsupervised domain adaptation for EM segmentation.
• The morphology-based metric provides an intuitive and effective approach for training model selection in the absence of validation data.