Visual Correspondences for Unsupervised Domain Adaptation on Electron Microscopy Images.

Area of Science:

• Electron Microscopy
• Computational Biology
• Machine Learning

Background:

• Domain shifts in electron microscopy data hinder the performance of pretrained models.
• Manual annotation of electron microscopy volumes is time-consuming and labor-intensive.
• Unsupervised domain adaptation aims to bridge the gap between different data domains without labeled data.

Purpose of the Study:

• To develop an unsupervised domain adaptation strategy for electron microscopy volumes.
• To enable pretrained models to operate effectively on new, unannotated data by compensating for domain shifts.
• To achieve high-quality segmentation of cellular structures without requiring new annotation efforts.

Main Methods:

• Aggregating visual correspondences (motifs) across different acquisitions to infer parameter changes in pretrained models.
• Identifying pivot locations in reference segmentations and using patch matching to find candidate correspondences in new volumes.
• Constructing a consensus heatmap based on aggregated correspondences to guide model adaptation.
• Employing Multiple Instance Learning or using heatmap regions as soft labels for model adaptation.

Main Results:

• Demonstrated high-quality segmentations on unannotated electron microscopy volumes.
• Achieved results qualitatively consistent with full supervision for mitochondria and synapse segmentation.
• Validated the effectiveness of unsupervised domain adaptation techniques in compensating for domain shifts.
• Eliminated the need for new annotation efforts for target volumes.

Conclusions:

• The proposed unsupervised domain adaptation strategy effectively addresses domain shifts in electron microscopy.
• The method successfully enables accurate segmentation of cellular structures without manual annotation.
• This approach offers a significant advancement for analyzing large-scale electron microscopy datasets.