Small-animal PET registration method with intrinsic validation designed for large datasets

Abstract

We present a procedure to validate the results of small animal Positron Emission Tomography (PET) image registration by means of consistency measures. Small animal 2-Deoxy-2-[F-18]fluoro-D-glucose (FDG) PET studies do not show the same intensity distribution even for images acquired in similar conditions, as the resulting image is influenced by several variables which are not always completely under control. Because of these difficulties, the results from automatic registration methods have to be visually inspected to detect failures. We propose a method to automate this validation process. Two reference images from the dataset are selected by an expert user avoiding images with poor contrast, animal movement or low quality, and both are co-registered using anatomical landmarks. All the remaining images in the dataset are then registered to every reference with an automatic two-step algorithm based on Mutual Information. The known transformation relating both references allows measuring the registration consistency, which is a good estimator of the accuracy of the alignment process, for every image in the dataset. This value can be used to assess the quality of the registration and therefore detect the incorrect results. We have applied this validation process on a large dataset of 120 FDG-PET rat brain images obtained with a rotating PET scanner. The registration consistency was calculated for every image in the dataset and values below 1.65 mm (PET image resolution) were considered as successful registrations. 116 images were correctly registered with an average error of 0.839 mm, while in four images the proposed method detected a registration failure. Two of these failures were due to very low image quality and these studies were discarded from the study, while the other two were correctly aligned after applying a manual pre-alignment step. Our approach requires minimal user interaction and provides automatic assessment of the registration error, making it unnecessary to visually inspect and check every registration result.