Abstract
Medical image registration is one of the key processing steps for biomedical image analysis such as cancer diagnosis. Recently, deep learning based supervised and unsupervised image registration methods have been extensively studied due to its excellent performance in spite of ultra-fast computational time compared to the classical approaches. In this paper, we present a novel unsupervised medical image registration method that trains deep neural network for deformable registration of 3D volumes using a cycle-consistency. Thanks to the cycle consistency, the proposed deep neural networks can take diverse pair of image data with severe deformation for accurate registration. Experimental results using multiphase liver CT images demonstrate that our method provides very precise 3D image registration within a few seconds, resulting in more accurate cancer size estimation.
Introduction
Accurately aligning 3D scans taken at different times or during different physiological phases is a fundamental challenge in medical imaging. This process, known as image registration, is essential for tracking tumor growth or organ changes. However, patient movement and internal organ deformation make this task difficult for traditional algorithms.
Currently, there are two main approaches: classical variational methods, which are accurate but computationally expensive (often taking many minutes), and supervised deep learning, which is fast but requires nearly impossible-to-obtain "ground truth" deformation fields. Existing unsupervised deep learning methods solve the data problem but often suffer from "folding" (non-physical mappings) and lack guarantees of reversibility.
Methods
The proposed framework builds upon the VoxelMorph architecture but introduces two critical components to improve performance: cycle-consistency and identity loss.
Cycle-Consistent Learning
Instead of training one network to map Image A to Image B, the researchers train two networks ($G_{AB}$ and $G_{BA}$) simultaneously. The cycle constraint dictates that if you warp Image A to match Image B, and then warp that result back using the second network, you should recover the original Image A. This forces the networks to learn a one-to-one, reversible mapping.
Loss Functions
The network is optimized using a combination of three primary loss terms:
- Registration Loss ($L_{regist}$): Uses cross-correlation as a similarity metric, which is robust to contrast changes in multiphase CT scans.
- Cycle-Consistency Loss ($L_{cycle}$): Ensures the transformation is reversible, acting as a powerful topological regularizer to avoid "folds."
- Identity Loss ($L_{identity}$): Penalizes the network if it attempts to deform an image when the source and target are identical, ensuring stability in stationary regions.
Results
The method was evaluated using a clinical dataset of multiphase liver CT scans from 605 patients.
- Speed: The proposed method takes only 10 seconds to register a 3D volume, compared to 19.6 minutes for the classical Elastix software.
- Topology Preservation: The cycle-consistent model significantly reduced folding problems, showing a lower percentage of voxels with non-positive Jacobian determinants (0.0175%) compared to the baseline VoxelMorph (0.0327%).
- Accuracy: The method achieved a Target Registration Error (TRE) of 3.76–4.91 mm. While slightly higher than the best classical methods (~3 mm), it was a marked improvement over standard unsupervised baselines (~5.3–6.6 mm) and provided more accurate clinical estimations of cancer size.
Conclusion
This cycle-consistent unsupervised framework offers a compelling balance between speed and accuracy. By ensuring reversibility, it avoids common pitfalls of deep learning registration like non-physical folding. While a small accuracy gap remains compared to state-of-the-art classical methods, the 100x speedup makes it highly suitable for real-time clinical applications in cancer diagnosis and monitoring.
Summary updated