Introduction
Segmenting mitochondria in Electron Microscopy (EM) is a critical task in neuroscience, yet it remains challenging due to the high variability in mitochondrial morphology and the presence of noise and look-alike sub-cellular structures. Traditional training methods often rely on cross-entropy loss, which treats all pixels equally. This approach is inefficient as "easy" pixels (e.g., large mitochondrial interiors or background) dominate the loss, while "hard" pixels (e.g., boundaries and ambiguous regions) are neglected. This paper proposes a contrastive learning framework to address these challenges by focusing on hard-to-segment regions and ensuring spatial consistency.
Methodology
The proposed framework enhances a standard 3D U-Net backbone with a contrastive learning objective. The architecture consists of three main components:
- Point Selection Block: Instead of dense pixel-wise loss, the model samples representative "uncertain" and "certain" points. This strategy prioritizes learning from high-error regions.
- Hybrid Feature Mapping: High-level features from the U-Net are combined with raw mask probability maps to create a robust "hybrid" representation for the sampled points.
- Dual-Term Contrastive Loss:
- Similarity Loss: Encourages features of the same class to cluster together while pushing different classes apart in feature space.
- Consistency Loss: Enforces temporal and spatial smoothness across the z-axis, leveraging the 3D continuity of biological structures.
A marker-controlled watershed algorithm is applied as a post-processing step to refine instance segmentation.
Datasets and Evaluation
The method was evaluated on two major benchmarks:
- MitoEM (R and H): Large-scale datasets from rat and human tissue, significantly larger than previous benchmarks.
- FIB-SEM: A standard mouse hippocampus dataset for historical comparison.
Results and Discussion
The contrastive learning approach yielded significant improvements over the U3D-BC baseline. On MitoEM-R, the AP-75 score increased from 0.844 to 0.870, while on MitoEM-H, it improved from 0.773 to 0.787. For the FIB-SEM dataset, the model achieved a Jaccard index of 0.895.
Qualitative analysis demonstrates that the proposed method is more effective at distinguishing mitochondria from similar-looking organelles, reducing false positives that typically plague standard models. By performing contrastive learning in a high-dimensional feature space, the model develops a more robust representation of mitochondrial structures.
Conclusion
This work demonstrates that targeted sampling and contrastive loss can significantly improve 3D instance segmentation in EM imagery. By focusing on "hard" pixels and cross-slice consistency, the model achieves state-of-the-art performance. Future research will explore semi-supervised learning to leverage the vast amounts of unlabeled EM data available in the field.
Summary updated