Automated and robust nonrigid registration of serial section microscopic images using PiCNoR

Accurate registration of serial-section microscopic images is essential for maintaining the spatial integrity of structural and functional information in biology and histology datasets, enabling critical advancements in 3D reconstruction and analysis. This paper introduces a new pixel-wise cluster-driven non-rigid registration (PiCNoR) method addressing the challenges in 3D microscopic imaging. PiCNoR utilizes feature-based local rigid registration as a foundational process, followed by clustering regions using Gaussian mixture models (GMM). Local rigid transforms are computed for these regions, validated through graph-based methods, and blended to achieve non-rigid transformations at the pixel level. This method is evaluated on three distinct datasets: the Kyoto embryo collection, a Drosophila brain stack, and a rat brain stack, demonstrating superior performance in preserving tissue continuity and reducing alignment errors compared to existing rigid transformations and established non-rigid approaches. A key advantage of this approach lies in its automated determination of the best number of clusters using the Bayesian information criterion (BIC), ensuring a balance between model complexity and the ability to capture meaningful spatial variations within the feature sets, all without the necessity for manual intervention or the need to verify the registered results. Consequently, this leads to a substantial reduction in computational costs. The validation and blending steps enhanced robustness against outliers, ensuring the method's reliability in handling complex deformations and high-resolution datasets. The results highlight the versatility and potential of the PiCNoR approach for applications in microscopic imaging, enabling accurate and reliable 3D reconstructions across diverse research contexts.