APIII - Advancing Practice, Instruction & Innovation Through Informatics

A Novel Image Registration Pipeline for 3-D Reconstruction from Microscopy Images

Kun Huang, PhD (khuang@bmi.usc.edu); Ashish Sharma, PhD; Lee Cooper, MS;
Tony Pan, MS; Metin Gurcan, PhD; Joel Saltz, MD, PhDDepartment of Biomedical Informatics Ohio State University, Columbus, OH

Context: Registration is the key step for 3-D reconstruction of microanatomical structures from large number of microscopy images of biomedical samples. In such experiments, biological samples are sectioned and digitized into hundreds or thousands of microscopy images. Registration is to align these images in 3-D space. Currently, most approaches are focused on optimally align consecutive images without incorporating the structural constraints on the anatomical structures (eg, smoothness of blood vessels). Here we present a novel pipeline for automatically registering stack of microscopy images with a smooth constraint on the reconstructed anatomical structures taken into consideration.

Technology: Histological sections are prepared with standard hematoxylin and eosin protocols and are digitized using Aperio ScanScope slide digitizer.  The pipeline uses segmentation and registration algorithms developed in Matlab.

Design: The pipeline has two major components: rigid transformation (rotation and translation) that aligns consecutive images globally and non-rigid transformation that matches fine structures locally. The first component has been developed in our previous works in which landmarks (eg, blood vessels) are extracted and matched. The rigid transformation is non-iteratively determined from the matched landmarks. For the second component, the key assumption is that anatomical structures such as blood vessels are smooth. This component includes the following steps:

• automatically track the landmarks across the stack of rigidly registered images using maximal cross correlation;
• for each track of landmarks, apply the smooth constraint by applying a median filter to its trajectory;
• in each image, compute the new locations of the landmarks derived from the previous step;
• for each image, compute the non-rigid transformation by applying thin-plate spline (non-rigid) algorithm which matches the original locations of the landmarks to their new locations.

 
Results: The pipeline has been implemented and applied to the reconstruction from two sets of microscopy images: mouse placenta and mouse mammary ducts (volumetric rendering of the registered images and the ducts using VolSuite Software are shown in the figure).

Conclusions: The constraint of smoothness of continuous anatomical structures is incorporated into our automatic registration pipeline. The transformation for each image is derived from the structural information of the entire stack of images instead of just its neighbor images.