[Paper Review] Morphological Error Detection in 3D Segmentations

David Rolnick, Yaron Meirovitch+6 more|arXiv (Cornell University)|May 30, 2017

Cell Image Analysis TechniquesBiochemistry, Genetics and Molecular Biology29 references20 citations

TL;DR

This paper proposes MergeNet, a 3D ConvNet that detects merge errors in neuronal segmentations by learning high-level morphological patterns from binary masks, enabling unsupervised, generalizable error localization across datasets without manual annotation. The method achieves high accuracy in identifying complex merge configurations and can even correct errors in its own training data.

ABSTRACT

Deep learning algorithms for connectomics rely upon localized classification, rather than overall morphology. This leads to a high incidence of erroneously merged objects. Humans, by contrast, can easily detect such errors by acquiring intuition for the correct morphology of objects. Biological neurons have complicated and variable shapes, which are challenging to learn, and merge errors take a multitude of different forms. We present an algorithm, MergeNet, that shows 3D ConvNets can, in fact, detect merge errors from high-level neuronal morphology. MergeNet follows unsupervised training and operates across datasets. We demonstrate the performance of MergeNet both on a variety of connectomics data and on a dataset created from merged MNIST images.

Motivation & Objective

• To address the persistent challenge of merge errors in automated 3D neuronal segmentation, which are common yet difficult to detect with existing methods.
• To develop a method that leverages morphological intuition—learned from human-like perception of neuronal shape—without relying on membrane predictions or image data.
• To enable unsupervised training on any reasonably accurate segmentation, allowing error detection and correction without additional manual annotation.
• To create a generalizable, scalable solution that works across diverse datasets and segmentation algorithms, independent of imaging modality or network architecture.

Proposed method

• MergeNet uses 3D ConvNets to analyze binary segmentation masks, focusing on global morphological context rather than voxel-level membrane predictions.
• The network is trained in an unsupervised manner using only the segmentation masks, learning to distinguish plausible neuronal shapes from merged ones.
• It localizes merge errors by generating probability maps where high values indicate regions likely to be erroneous fusions of distinct neurons.
• The model is capable of detecting errors within its own training data, even when those errors were not explicitly labeled during training.
• It generalizes across datasets by learning invariant morphological features, enabling deployment on new data without retraining.
• The architecture is optimized for speed, processing up to three million voxels per second on a single GPU, making it suitable for large-scale connectomics pipelines.

Experimental results

Research questions

• RQ1 Can a deep learning model detect merge errors in 3D neuronal segmentations by learning high-level morphological patterns rather than low-level image features?
• RQ2 To what extent can a 3D ConvNet detect diverse and complex merge configurations—such as parallel, perpendicular, or multi-object fusions—without explicit supervision?
• RQ3 Can the model detect errors in its own training data, even when those errors were not annotated, demonstrating self-correcting capability?
• RQ4 How well does the model generalize across different datasets and segmentation algorithms, and can it be deployed without retraining?
• RQ5 Can the model detect non-neuronal objects like glia as anomalies based solely on morphological plausibility, despite no training on such classes?

Key findings

• MergeNet successfully detects a wide range of merge error configurations, including parallel, perpendicular, and multi-object fusions, as demonstrated in both connectomics and MNIST-based datasets.
• The model achieves high accuracy in localizing merge errors, with visualizations showing precise identification of error-prone regions such as X-junctions and complex fusions.
• It can detect and localize errors in its own training data without any manual annotations, demonstrating self-correcting capability through unsupervised learning.
• The algorithm generalizes across datasets and segmentation methods, maintaining high performance even when applied to data from different sources and imaging modalities.
• MergeNet also detects non-neuronal objects like glia as morphologically implausible, despite no training on such classes, suggesting emergent generalization to shape anomalies.
• The system processes data at a rate of three million voxels per second on a single GPU, outperforming many membrane-prediction systems in speed and enabling integration into scalable connectomics pipelines.