Anomaly Detection through Registration - Robotics Institute Carnegie Mellon University

Anomaly Detection through Registration

Mei Chen, Takeo Kanade, Henry Rowley, and Dean Pomerleau
Tech. Report, CMU-RI-TR-97-41, Robotics Institute, Carnegie Mellon University, November, 1997

Abstract

The goal of this research is to perform automatic, fast and accurate registration of volumetric data in 3-D space, as well as anomaly detection based on the registration results and domain knowledge. One important application domain is medical image registration. Anatomical structures vary considerably in appearance across individuals or within one individual over time, and any pathology may aggravate these variations. We adopt an intensity-based approach for registration, which assumes no prior segmentation of the voxels in a volume. To address the appearance variations of anatomical structures, we have developed a three-level hierarchical deformable registration algorithm. First, a 3-D global transformation (rotation, uniform scaling, and translation) aligns the different image volumes. This procedure corrects for variations induced during the image acquisition process. Secondly, a smooth deformation, represented by the warping of a 3-D grid of control points, approximately matches the corresponding anatomical structures in the image volumes. This step partially adjusts for the inherent variations in the appearances of the anatomical structures across individuals. Several grid resolutions are used to progressively change the emphasis from global alignment to alignment of specific anatomical structures. Finally, a fine-tuning deformation, which allows each voxel to move independently, refines the matching of corresponding anatomical features. In this process each voxel determines the deformation instead of the control points alone, therefore, it is able to tune the finer scale variations overlooked by the smooth deformation. Interwoven into this three-level registration algorithm is a hierarchy of coarse-to-fine processing. An image pyramid, used in the global transformation and the smooth deformation, serves both to improve efficiency and to help the algorithm first focus on global patterns and gradually shift to specific details. Based on our algorithm, we have developed a prototype, ADORE (Anomaly Detection thrOugh REgistration), and conducted experiments on actual medical imaging data of different parts of the body (head, hip, and knee). This involves registration of volumetric data within different imaging modalities (MRI and CT), and images with and without pathologies. ADORE can match two 16MByte volumes in 12 minutes on an SGI workstation with four 194 MHz R10000 processors, with an accuracy comparable to manual registration. By matching an expert-segmented atlas to a patient's image data, ADORE can automatically build a customized atlas for different individuals. Pathology detection is critical, but the presence of pathologies is not normal and accurate registration cannot be achieved. Nevertheless, the fact that they cannot be registered may be sufficient evidence of their existence. By applying our hierarchical deformable registration algorithm, and using the knowledge of normal anatomy and pathologies, ADORE is able to detect certain pathologies in actual data that affect the anatomical morphology. In the paper we will discuss in detail the algorithm and implementation, the experiments and results, as well as the potential applications.

BibTeX

@techreport{Chen-1997-14528,
author = {Mei Chen and Takeo Kanade and Henry Rowley and Dean Pomerleau},
title = {Anomaly Detection through Registration},
year = {1997},
month = {November},
institute = {Carnegie Mellon University},
address = {Pittsburgh, PA},
number = {CMU-RI-TR-97-41},
}