Image Registration for Radiation Therapy Planning


Image registration is the process of transforming different acquired images of the same content into the same spatial coordinate system.

Different types of transformations can be used: Rigid-body transformations which allow translations and rotations of the image content. Several non-rigid transformations can be applied which allow deformations of the image content. The result of an image registration process is a transformation, which frequently is represented by a vector field.



In radiation therapy different medical images need to be spatially related to facilitate the accuracy of cancer treatment with ionizing radiation. Prior to the delivery of the first fraction of the irradiation an accurate anatomical model of the patient is required for treatment planning. Here, delineation of volumes of interest (VOIs) – both target volumes and organs at risk - is a prerequisite. Computed tomography (CT) scans are used as the basis, but not all required volumes can be distinguished sufficiently, since some structures do not exhibit enough contrast in the CT scans. Different image modalities, like magnetic resonance imaging (MRI) scans, can complement the necessary information, if they can be spatially aligned with the planning CT scans.


Treatment plans, based on a 3D planning CT only, might not be sufficient to treat patients throughout an extended time period, since anatomical changes are known to occur during the treatment course. Different strategies were developed - and are still in development - to account for these changes. All of the compensating strategies have one in common: The occurred anatomical changes have to be quantified first. This necessary information can normally be extracted from the imaging scans throughout the treatment course. Finding the transformation of the patients’ anatomy in the images with respect to the planning CT scans is the task of image registration algorithms.



Image registration is a computationally intensive optimization problem with many degrees of freedom. Long computation time is therefore a significant limitation. Enhancements of the imaging techniques and the increasing frequency of the scan acquisition throughout the treatment course produce more and more medical image data challenging the computational performance of registration algorithms. Fast and accurate registration methods are needed to handle the massive amounts of data.
We explore and develop methods to increase the efficiency of registration algorithms including hardware parallelization using multi-CPU systems as well as graphical processing units (GPUs).

 A drawback of the purely intensity based image registration methods is the lack of bio-mechanical consistency of the resulting transformations. Since the material properties of the represented tissues are not incorporated in the CT scans itself, additional modeling of the tissue properties is necessary to impress physically meaningful transformations into the registration process. Modeling of the bio-mechanical properties promises more accurate and more reliable transformations, but also implicates higher computational costs.
 → We explore different bio-mechanical modeling methods to describe anatomical changes and investigate how to incorporate tissue-specific properties in registration algorithms with only moderate increase of computational costs.

Reliable software-based multi-modal deformable registration is an important prerequisite to merge all time dependent anatomical as well as functional information gathered along the treatment course of each individual patient. Software-based reliable multi-modal registration methods need to be developed to enable a flexible utilization of functional imaging inside radiotherapy treatment.
We investigate the usage of multi-modality enabled similarity measures in deformable image registration algorithms.

Accuracy estimation and validation of deformable image registration results is an important issue, whenever the resulting transformations are used in consecutive calculations. The quantification of the uncertainty of the resulting transformation is not straightforward, since the true transformation of the anatomy of the patient is generally not known.
We investigate the utilization of diverse methodologies to estimate the uncertainties associated with rigid and deformable image registration methods.

People involved:
Rolf Bendl
Gernot Wurst
Kristina Giske
Markus Stoll 

Urban Malsch
Pan Li
Nicola Leucht

You want to know more? …. Get in touch with Prof. Dr. Rolf Bendl

to top