Automatic registration of 2D histological sections to 3D microCT volumes: Trabecular bone

摘要

Abstract Histomorphometry and microCT are the two dominant imaging techniques to study bone structure and quality to evaluate repair, regeneration, and disease. These two methods are complementary; where histology provides highly resolved tissue properties on a cellular level in 2D, microCT provides spatial information of bone micro-structure in 3D. For this reason, both of these modalities are commonly used in bone studies. As it is not trivial to combine the images of these two modalities, the two methods are typically applied to different specimens within a study. However, we believe that applying both imaging modalities to the same specimen with a suitable fusion strategy may further strengthen the value of each modality. Therefore, we propose a registration method to align 2D histology slices with a 3D microCT volume, without any prior knowledge of the sectioning direction. In a preprocessing step, bone is extracted from both images. Then, we use a strategy for initializing potential locations, and an iterative approach for searching for an ideal fitting plane using Radon-based rigid transforms and feature-based affine alignments. The algorithm was tested and validated with simulated and real data. For the latter, microCT images of trabecular bone with 76 corresponding histological sections acquired from decalcified and calcified specimens were used. The registration resulted in 94.7% acceptable solutions as defined by a registration orientation error of less than 3°. Average registration accuracy of the acceptable results was 0.6°, leading to a target registration error for our method of 106.3μm, computed based on landmarks annotated by an observer. This corresponds roughly to 10pixels in the images; although, the relation to actual visible structures that provide the features to register, is arguably more relevant. Highlights ? A new method to register 2D histological images to 3D microCT images is proposed. ? No initial localization is required and the method is independent of sample geometry. ? Results are presented for real histology images and simulated data. ? 95% of real images registered well, outperforming previous methods.