Benchmarking Image-Based Motion-Correction Methods for Ultrasound Localization Microscopy

Ultrasound localization microscopy (ULM) achieves sub-diffraction resolution imaging in vivo through localizing and tracking microbubbles. However, the need to accumulate microbubble signals over time makes ULM highly sensitive to tissue motion, necessitating accurate motion correction. The accuracy of motion-correction techniques poses a limit to the attainable resolution, and there is currently no gold standard algorithm or approach.

This study benchmarked seven publicly available implementations of non-rigid image registration algorithms using two simulated datasets illustrating soft tissue and cardiac images, as well as in vivo acquisitions of a rabbit kidney and human breast tumor. Five benchmarks were used to evaluate the seven implementations using image-based similarity metrics, errors against ground truth deformation fields, robustness to hyperparameter choice and image quality, including data with varying contrast-to-noise ratios. Using Bayesian optimization and Sobol sensitivity analysis, optimal parameters for each algorithm were identified, with guidelines for data-adaptive algorithm selection proposed.

Parameter sensitivity analysis was reported for all implementations, which can be used to prioritize parameters when performing optimization. Motion characteristics and image spatial heterogeneity were found to be important factors for implementation accuracy. Spline-based algorithms, such as free-form deformations implemented in Elastix, performed optimally with small deformations and low spatial heterogeneity. In contrast, methods designed for large deformations, such as large deformation metric matching, implemented by Ceritoglu et al., or free-form deformations with diffeomorphic constraints, such as Niftyreg, were effective at correcting larger data displacements with high heterogeneity, but struggled to identify accurate correspondences when deformation magnitudes were small. Invertibility was beneficial when correcting larger deformations as well as when images are more sparse. All implementations were robust to varying image quality and hyperparameter choice, which was reflected in our demonstration using different implementations to obtain ULM in a rabbit kidney and human breast tumor. Elastix was found to achieve more vessel alignment and structure in the kidney data, where motion was smaller. In the breast dataset, Niftyreg obtained better alignment. These findings highlight the need for data-driven approaches and provide a foundation for identifying appropriate and reliable motion-correction methods in ULM. Our code is available on GitHub.

Algorithm choice should be performed in a data-adaptive manner that considers the spatial heterogeneity of the data and its motion characteristics to enable ULM. This study benchmarks readily available image registration algorithms and provides recommendations as to when they may be most appropriate.