Segmentation Certainty Through Uncertainty: Uncertainty-Refined Binary Volumetric Segmentation Under Multifactor Domain Shift

Carianne Martinez, Kevin M. Potter, Matthew D. Smith, Emily A. Donahue, Lincoln Collins, John P. Korbin, Scott A. RobertsWorkshops @ CVPR 2019 · DOI

Introduction

Deep learning models, particularly those used in volumetric image segmentation, often exhibit "brittleness" when encountering data from a different domain than their training set. This phenomenon, known as domain shift, occurs when variations in scanner hardware, resolution, or noise levels lead a model to produce inaccurate results. In the context of CT scans for materials science, domain shift frequently manifests as "overzealous" predictions, where the model over-segments material in the new domain. Researchers at Sandia National Laboratories have proposed a novel method to address this by leveraging the model's own uncertainty to refine predictions in the output space, avoiding the need for costly retraining.

Methodology: Uncertainty Quantification (UQ)

The core of the proposed approach is the use of Monte Carlo (MC) Dropout within a 3D V-Net architecture. Unlike standard inference, which provides a single deterministic output, MC Dropout involves running multiple inference passes (five in this study) with dropout layers active. This generates a distribution of predictions for every voxel.

Figure 3: Visualization of (a) the original CT slice, (b) the initial binary segmentation mask, and (c) the corresponding uncertainty map generated via Monte Carlo Dropout.

The standard deviation across these passes is used to construct an uncertainty map. The researchers observed a critical correlation: in shifted domains, the model's over-predictions typically occur in regions where its uncertainty is significantly higher than in regions where it is correct. This uncertainty serves as a functional signal for automated correction.

Automated Refinement and Threshold Optimization

The refinement process involves flipping "Label 1" (material) predictions back to "Label 0" (background) if the uncertainty at a given voxel exceeds a specific threshold $t$. To determine the optimal threshold without manual intervention, the authors developed an optimization metric that maximizes the statistical separation between the intensities of the two resulting classes. This allows the image statistics themselves to guide the correction process, making the method robust across varying scan qualities.

Figure 4: Refinement process visualization: (a) input slice, (b) initial binary mask, and (c) the mask overlaid with uncertainty-based boundary refinements (red regions) targeted for correction.

Key Results and Insights

  • High In-Domain Accuracy: On data matching the training domain, the model achieved a baseline accuracy of 99.1%.
  • Robustness to Shift: In domains where the original model produced "unusable" results due to extreme over-segmentation, the uncertainty-refined method successfully recovered the underlying structure of the woven composites.
  • Model-Agnostic Utility: By focusing on output-space refinement (the segmentation and uncertainty maps) rather than internal feature adaptation, the method remains computationally efficient and does not require a secondary deep learning model for post-processing.
  • Expert Validation: Materials science experts confirmed that the uncertainty-driven approach produced cleaner segmentations with less noise than traditional intensity-based thresholding methods.

Conclusion and Future Work

This research demonstrates that uncertainty is not just a metric for human oversight but a powerful tool for automated model self-correction. While currently validated qualitatively on material scans, future work will focus on quantifying these improvements against ground-truth labels in shifted domains and exploring applications in medical imaging, where domain shift between different hospital scanners remains a significant challenge. The method's ability to refine segmentations without retraining offers a scalable path toward more reliable AI in production environments.

Summary updated

BibTeX
@inproceedings{martinez2019segmentation,
  author = {Carianne Martinez and Kevin M. Potter and Matthew D. Smith and Emily A. Donahue and Lincoln Collins and John P. Korbin and Scott A. Roberts},
  title = {{Segmentation Certainty Through Uncertainty: Uncertainty-Refined Binary Volumetric Segmentation Under Multifactor Domain Shift}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
  doi = {10.1109/cvprw.2019.00066},
}
Figure