Abstract
Unsupervised medical image registration typically employs a Mean Squared Error (MSE) loss function, which implicitly assumes uniform, or homoscedastic, noise across the image domain. However, medical imaging modalities such as MRI and ultrasound exhibit spatially varying, heteroscedastic noise influenced by anatomical structures and acquisition physics. This paper introduces a collaborative training framework that estimates input-dependent noise variance to improve registration robustness. By utilizing a novel $\gamma$-exponentiated relative signal-to-noise ratio (SNR) weighting scheme, the proposed method adaptively discounts noise-induced outliers. Experimental results on cardiac MRI and ultrasound datasets demonstrate that this approach significantly enhances registration accuracy and anatomical consistency across both CNN and Transformer-based architectures.
Introduction
Image registration—the spatial alignment of two or more images—is a fundamental task in medical image analysis, enabling longitudinal studies and multi-modal fusion. Modern unsupervised deep learning methods often rely on the Mean Squared Error (MSE) as a similarity metric. A critical limitation of MSE is its underlying assumption of homoscedasticity: the belief that noise is uniformly distributed.
In clinical practice, medical images are inherently heteroscedastic. Noise levels fluctuate due to patient motion, anatomical complexity, and scanner-specific artifacts (e.g., speckle in ultrasound). When a registration model assumes uniform noise, it may erroneously attempt to align noise-induced intensity variations as if they were true anatomical features, leading to warped and physically implausible deformations.
Methodology: Collaborative Heteroscedastic Modeling
To address these challenges, the researchers developed a collaborative architecture consisting of two primary components: a Displacement Estimator and a Variance Estimator.
The Displacement Estimator (e.g., Voxelmorph or Transmorph) predicts the deformation field required to align a moving image $I_m$ to a fixed image $I_f$. Simultaneously, the Variance Estimator—a U-Net-based module—predicts a spatial noise map $\hat{\sigma}_f^2$ for the image pair.
The core innovation is the $\gamma$-exponentiated relative SNR weighting scheme. Standard heteroscedastic modeling (like Negative Log-Likelihood) often penalizes high-intensity regions where absolute noise is higher, even if the signal is strong. The proposed scheme weights the registration loss based on the relative SNR, ensuring the model trusts high-signal regions while discounting areas where noise dominates the anatomy. The two modules are trained using an alternating strategy, creating a feedback loop where the variance estimator identifies alignment failures due to noise, refining the displacement model's focus.
Experimental Results
The framework was validated on three cardiac datasets: ACDC (human cardiac MRI), CAMUS (2D human echocardiography), and a private 3D echocardiography dataset. The proposed method consistently outperformed baseline models across multiple metrics, including Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Average Surface Distance (ASD).
As shown in the results table, the method achieved a DSC of 80.73% on the ACDC dataset, surpassing standard Voxelmorph (80.20%) and other uncertainty-aware baselines like $\beta$-NLL. Notably, the framework is "plug-and-play," showing improvements when integrated into both CNN-based and Transformer-based (Transmorph) registration models. Statistical analysis via paired t-tests confirmed the significance of these improvements, with p-values reaching as low as $1.02 \times 10^{-41}$.
Discussion and Conclusion
Beyond quantitative gains, the model produces sensible uncertainty maps that highlight image artifacts and ambiguous edges, providing valuable interpretability for clinicians. However, some limitations remain. The SNR-based weighting relies on a manually tuned hyperparameter $\gamma$, and the model can still struggle with extremely thin anatomical structures where the signal is sparse.
Future research will focus on developing fully data-driven objectives to eliminate manual hyperparameter tuning and expanding validation to a broader range of clinical applications beyond cardiac imaging. By moving away from simplistic noise assumptions, this work paves the way for more reliable and anatomically accurate medical image registration in the presence of real-world noise.
Summary updated