A high accuracy, landmark-based, sub-pixel level image registration method

Many medical applications require high-accuracy registration of images. In this paper, we present a landmark-based method for registering images that allows for scaling, translation, and rotation. The method is applied to register high-resolution (HR) Visible Human Male cryosection images with their corresponding registered low-resolution (LR) images. The difference in resolution presents difficulty in landmark selection. The solution for this requires equalizing pixel sizes of images from both sets. Registering HR images to LR images, already registered to each other, results in a registered set of HR images. The accuracy of the registration algorithm is evaluated by registering an image transformed with preset factors and comparing it to the original image. The accuracy of registration is calculated as 10−3 HR pixels for scaling, 10−3 degrees for rotation and 10−7 HR pixels for translation. Once the HR set is registered and segmented it is possible to generate 3D virtual models of anatomical structures for other applications.