Abstract
Semi-supervised learning offers a cost-effective approach for neuron segmentation in electron microscopy (EM) volumes by leveraging extensive unlabeled data. However, distribution mismatch between limited labeled data and diverse unlabeled structures often hinders performance. We propose a distribution-aware pipeline featuring an unsupervised heuristic for selective labeling and an axial-through mixing strategy for semi-supervised training. Our method achieves performance comparable to full supervision with only 20% of the labels across diverse EM datasets.
Introduction
Connectomics aims to map neural circuits at nanoscale resolution using Volume Electron Microscopy (EM). The sheer scale of these datasets—often reaching petabytes—necessitates automated neuron segmentation. While deep learning models are effective, they require vast amounts of expert-annotated data. Manual labeling is prohibitively slow; for instance, only 0.003% of the fly hemi-brain dataset was annotated for training.
Semi-supervised learning (SSL) attempts to mitigate this by using unlabeled data. However, a significant challenge is distribution mismatch: the small subset of labeled data often fails to represent the structural diversity of the entire volume. This leads to unreliable predictions on unlabeled data, limiting model generalization.
Methodology
To address distribution mismatch, we introduce a two-stage pipeline focusing on data selection and model regularization.
1. Coverage-based Greedy Selection (CGS)
Instead of random sampling, we employ an unsupervised heuristic to select representative sub-volumes for annotation. First, a self-supervised Patch Encoder (trained via contrastive learning) maps unlabeled patches into a feature space. We then use a greedy algorithm to select sub-volumes that maximize "semantic coverage" of the entire dataset's distribution. This ensures that the human expert annotates the most diverse and representative neuronal structures.
2. Mixed-view Consistency Learning (IIC-Net)
Our semi-supervised framework, IIC-Net, utilizes a novel Axial-through Spatial Mixing strategy. EM data is typically anisotropic, with lower resolution along the Z-axis. Standard random mixing can create artifacts that confuse the model. Our strategy mixes labeled and unlabeled data only along high-resolution axes, preserving structural continuity. By enforcing consistency across these mixed views, the model learns shared semantics between labeled and unlabeled distributions.
Experimental Results
We evaluated our pipeline on diverse public EM benchmarks, including rat somatosensory cortex (AC3/AC4, Kasthuri15), fruit fly brain (CREMI, Hemi-brain), and human temporal lobe (AxonEM-H).
- Label Efficiency: On the Kasthuri15 dataset, our method achieved a Variation of Information (VI) score of 0.7842 using only 20% of the labels, nearly matching the 100% supervised baseline (0.7751).
- State-of-the-Art Performance: Our approach consistently outperformed existing SSL methods like Mean Teacher and SSNS-Net across multiple metrics (VI and ARAND).
- Robustness to Anisotropy: The axial-through mixing strategy proved critical for maintaining accuracy in anisotropic datasets where standard augmentations fail.
Conclusion
Our distribution-aware pipeline significantly reduces the manual annotation burden for large-scale connectomics. By intelligently selecting representative data and employing structure-preserving mixing strategies, we bridge the gap between labeled and unlabeled distributions, enabling cost-effective and robust neuron segmentation across species and imaging modalities.
Summary updated