Introduction
In the field of connectomics, researchers aim to map the intricate wiring of the brain by reconstructing 3D models from Serial Block-Face Scanning Electron Microscopy (SBF-SEM) data. This technique generates hundreds of consecutive 2D slices, creating a massive "labeling bottleneck." Deep learning models typically require vast amounts of human-annotated training data, which is time-consuming and expensive to produce. This paper introduces a semi-supervised learning (SSL) approach that achieves high-quality segmentation using as few as three manually labeled initial images.
Iterative Pseudo-Labeling Method
The core of the proposed method is an iterative pseudo-labeling strategy that leverages the spatial continuity of high-resolution EM stacks. Since visual changes between consecutive slices are incremental, a model trained on slice $N$ is likely to produce a reliable prediction for slice $N+1$.
The process works in a continuous loop:
- Initial Training: A Deep Contextual Network is trained on a small set of $M$ manually labeled slices (e.g., $M=3$).
- Prediction: The model predicts the segmentation mask for the next unlabeled slice ($M+1$).
- Pseudo-Labeling: This prediction is treated as a "pseudo-label," effectively allowing the model to trust its own output as ground truth for further training.
- Iterative Retraining: The model is retrained using a sliding window of slices that includes these new pseudo-labels.
- Propagation: This process repeats slice-by-slice through the entire volume.
Experimental Evaluation
The researchers evaluated their method on two distinct datasets: the ISBI 2012 EM Segmentation Challenge (30 slices of neuronal membranes) and a custom Ant Olfactory Sensory Unit dataset (377 slices). The latter represents a real-world "small data" scenario where sample preparation is difficult, leaving researchers with very limited initial data.
Results
The semi-supervised approach significantly outperformed standard supervised learning when labels were scarce. On the ISBI 2012 dataset, the SSL method maintained segmentation quality even as it moved further from the initial labels.
As shown in the results for the 30th slice of the ISBI stack, the SSL method achieved a Rand Score ($V_{rand}$) of 0.7310, a massive improvement over the 0.0645 achieved by the supervised baseline. Similarly, on the 100th slice of the ant dataset, the SSL method reached an Intersection over Union (IoU) of 0.5314, compared to 0.3426 for the baseline. These results confirm that iterative pseudo-labeling prevents the "performance drift" that typically occurs when a model trained on early slices encounters the morphological shifts present later in a 3D volume.
Discussion and Future Work
While promising, the authors acknowledge that absolute accuracy is not yet high enough for fully automated 3D reconstruction. They suggest several areas for future improvement, including the use of 3D Convolutional Neural Networks or Recurrent Neural Networks (RNNs) to better capture spatial correlations across the Z-axis. Additionally, implementing post-processing steps to clean up pseudo-labels could help mitigate the risk of error accumulation in the iterative feedback loop.
Summary updated