Abstract
This paper presents a novel deformable registration framework, leveraging an image prior specified through a denoising function, for severely noise-corrupted placental images. Integration of powerful image denoisers into advanced registration methods provides our model with a flexibility to accommodate datasets that have low signal-to-noise ratios (SNRs). We demonstrate the performance of our method under a wide variety of denoising models in the context of diffeomorphic image registration. Experimental results show that our model substantially improves the accuracy of spatial alignment in applications of 3D in-utero diffusion-weighted MR images (DW-MRI) that suffer from low SNR and large spatial transformations.
Introduction
Imaging the placenta in-utero is notoriously difficult. Researchers use Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI) to monitor placental health, but these scans are plagued by two major issues: extreme noise at high b-values and significant motion artifacts from fetal movement and maternal respiration. Standard registration algorithms often fail in these low-SNR environments, resulting in poor alignment or geometric artifacts like tearing. While sequential denoising-then-registration approaches exist, they often smooth out critical structural details needed for accurate alignment.
Methodology: The PnP-RR Framework
The authors propose a joint framework called Plug-and-Play Registration-Reconstruction (PnP-RR). Traditionally, solving an optimization problem involving an image "prior" requires a mathematically differentiable function. PnP-RR bypasses this by replacing complex regularizers with off-the-shelf denoisers (like BM3D or Total Variation).
The framework operates via alternating minimization, looping between two primary steps:
- Registration Step: It utilizes the FLASH (Fast LDDMM) algorithm to find the optimal transformation between the source image and a reconstructed target image.
- Denoising Step: It updates the reconstructed image by passing it through a chosen denoiser.
Because these steps happen iteratively, the registration benefits from a cleaner image, and the reconstruction benefits from the spatial information provided by the registration.
Experimental Evaluation
Synthetic 2D Results
Initial validation was performed on 2D synthetic binary images corrupted with heavy white Gaussian noise ($\sigma = 0.3$). As shown in the figure below, the baseline FLASH algorithm struggles to recover the shape of the target under such noise, while PnP-RR variants (using TV, TGV, and BM3D priors) successfully reconstruct the boundary and align the images.
Real-World 3D Placental DW-MRI
The framework was evaluated on 3D placental DW-MRI scans from pregnant subjects at approximately 20 weeks gestation. Quantitative results using the Dice Similarity Coefficient (DSC) showed that PnP-RR consistently achieved higher average overlap scores with smaller variations than traditional two-step methods.
Qualitative inspection (above) confirms that the model successfully captures large deformations on the edges of the placenta, where standard methods typically struggle or create folding artifacts. The PnP approach also proved less sensitive to parameter tuning, making it more viable for clinical settings.
Conclusion
PnP-RR offers a flexible and robust solution for medical image registration in high-noise environments. By decoupling the image prior from the registration logic, the system allows researchers to "plug in" the most effective denoising algorithms for their specific data. Future work may involve integrating deep learning-based denoisers to further improve accuracy and computational speed for real-time clinical use.
Summary updated