Uncertainty Learning towards Unsupervised Deformable Medical Image Registration

Xuan Gong, Luckyson Khaidem, Wentao Zhu, Baochang Zhang, David S. Doermann

Abstract

Unsupervised deep learning models for medical image registration often lack reliability metrics, hindering their adoption in critical clinical settings. This paper introduces a novel framework for simultaneous registration flow and uncertainty estimation. By integrating a predictive uncertainty module with efficient ensemble techniques, such as snapshot ensembles, the proposed method achieves state-of-the-art accuracy while providing reliable confidence maps for clinical decision-making. Evaluation on benchmark datasets demonstrates that "knowing what the model doesn't know" significantly enhances registration performance.

Introduction

Deformable image registration (DIR) is the process of aligning two images—such as a pre-operative MRI and an intra-operative scan—by accounting for non-linear tissue changes. While traditional methods are computationally expensive, modern unsupervised deep learning models like VoxelMorph offer near real-time performance. However, these models typically operate as "black boxes," providing no measure of confidence in their predictions. In high-stakes environments like neurosurgery, a small registration error can have catastrophic consequences. This work addresses the challenge of quantifying uncertainty in an unsupervised setting where no ground-truth transformation is available during training.

Methods

The proposed framework extends the VoxelMorph architecture by introducing several key innovations aimed at capturing both aleatoric (data-dependent) and epistemic (model-dependent) uncertainty.

The proposed framework architecture, featuring a predictive uncertainty module and spatial transform block.

Predictive Uncertainty and Ensembles

  1. Predictive Uncertainty Module: A dedicated branch learns to encode reconstruction error. By modeling registration flow as a Laplace distribution, the network outputs both the mean (registration) and variance (uncertainty).
  2. Empirical Ensembles: To capture epistemic uncertainty, the authors investigated three strategies: Bootstrap (multiple independent models), MC Dropout (inference-time dropout), and Snapshot Ensembles. The latter uses a cyclic learning rate to save multiple model states during a single training run, providing ensemble benefits at the cost of a single training cycle.
  3. Bayesian Integration: Predictive and empirical uncertainties are unified into a single metric, which is then used in a residual refinement block to improve warped segmentation masks.

Key Results and Insights

The framework was evaluated on the Hippocampus and Prostate datasets from the Medical Segmentation Decathlon. A striking finding was that adding an uncertainty prediction branch actually improves registration accuracy. For example, on the Hippocampus dataset, the "Ours+Snap" method achieved a 95.12% Dice score, significantly outperforming the standard VoxelMorph (91.20%).

Sparsification as Validation

To validate the uncertainty metric without ground-truth flow, the authors utilized sparsification plots. These plots measure how quickly the registration error drops as pixels with the highest predicted uncertainty are removed.

Sparsification plots showing that removing high-uncertainty pixels leads to a rapid decrease in registration error compared to baseline methods.

As shown in the figure above, the proposed method's uncertainty metric serves as a highly accurate proxy for actual error. The Area Under the Sparsification Error (AUSE) dropped from 0.0077 in the baseline to 0.0003 in the "Ours+Snap" configuration, indicating superior uncertainty calibration.

Conclusion

This work demonstrates that uncertainty learning is not just a safety feature but a driver of registration accuracy. By leveraging snapshot ensembles, the framework provides a computationally efficient way to generate reliable confidence maps. Future research will focus on reducing inference latency for real-time applications and exploring direct flow uncertainty validation techniques that do not rely on segmentation labels.

Summary updated

Abstract

Uncertainty estimation in medical image registration enables surgeons to evaluate the operative risk based on the trustworthiness of the registered image data thus of paramount importance for practical clinical applications. Despite the recent promising results obtained with deep unsupervised learning-based registration methods, reasoning about uncertainty of unsupervised registration models remains largely unexplored. In this work, we propose a predictive module to learn the registration and uncertainty in correspondence simultaneously. Our framework introduces empirical randomness and registration error based uncertainty prediction. We systematically assess the performances on two MRI datasets with different ensemble paradigms. Experimental results highlight that our proposed framework significantly improves the registration accuracy and uncertainty compared with the baseline.
BibTeX
@inproceedings{gong2022uncertainty,
  author = {Xuan Gong and Luckyson Khaidem and Wentao Zhu and Baochang Zhang and David S. Doermann},
  title = {{Uncertainty Learning towards Unsupervised Deformable Medical Image Registration}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
  doi = {10.1109/wacv51458.2022.00162},
}
Figure