Digital Image Processing Lab

Department of Radiology, University of Michigan

Quantifying Response to Radiation Therapy Using Registered 3D Data Sets

Background and Introduction

In patients undergoing radiation therapy for the control of cancerous tumors, it is of vital importance to gauge the effectiveness of the therapy. At the present time, this is typically done in a qualitative manner by visually comparing serial radiological exams of the patient over the course of therapy. The Department of Radiology's Digital Image Processing Lab has developed a unique and innovative system for automated registration or fusion of 3D data sets. This offers the unique opportunity to obtain quantitative measures of tumor size and shape changes by allowing precise comparisons of the serial exams. ("Registration" refers to an operation involving a pair of data sets in which one of the data sets is geometrically transformed so that the anatomical features of the pair of data sets are in the correct relation to one another. In general, radiologic exams are inherently "out of registration" and thus cannot easily be compared in a quantitative way.)

Our registration system has been used to register a variety of data set types (e.g., CT, MRI, PET, SPECT, and ultrasound). Control point pairs in the two data sets define the transformation of one data set onto the other. The quality of the registration is measured by mutual information (MI). An optimization loop systematically moves the control points until the MI is maximized, at which point the registration is optimal. Our software is known as "MIAMI Fuse" - Mutual Information for Automated Medical Image Fusion.

The present version of the registration system was developed under AVS/5 (Advanced Visual Systems, Inc.). The software will be ported to AVS/Express (the newest version of Advanced Visual Systems's visualization/development environment) which uses a standard Motif user interface. More importantly, after the port, the software will run on a multi-processor, high performance Intel platform to support extensive pre-clinical trials within Radiology where many clinical data sets may be routinely registered and quantitative lesion changes can be measured.

Examples

Here we present two cases that illustrate the operation of MIAMI Fuse.

This 3D rendering of a patient was performed using three coregistered data sets: MRI data, PET data, and SPECT data registered two at a time. The MRI was performed after the patient had received radiation therapy treatment for a metastatic tumor. A post-therapy MRI study raised concern that the cancer may be regrowing. A PET scan was performed to determine the metabolic rate of the suspected region, and a SPECT was performed to demonstrate the state of the blood-brain barrier. Side-by-side clinical comparison of the three scans left considerable doubt regarding the FDG activity in the suspect lesion due to uncertainty in location. Only after automatic registration was performed was the clinician certain that the lesion was simply necrotic.

In the rendering of the fused data sets shown here the green hue is driven by the MRI signal amplitude, the red hue is driven by the coregistered PET study, and the blue hue is driven by the coregistered SPECT study.

The PET and CT data sets shown in this study were acquired from a patient with lung cancer. For the CT exam the patient used shallow, free respiration with arms placed at the sides to mimic PET data acquisition conditions. Note the resulting accurate delineation of the cardiac and vascular structures, as well as that of the lesion. One pair of well-registered, prospectively placed, external sternal surface markers can also be seen, even though the marker played little to no role in the registration. The apparent registration accuracy using the full affine geometric model is a tribute to positioning the patient in a consistent scanning geometry for both modalities, i.e. arms down in both scanners, as well as the use of an accurate registration algorithm. Generally in cases where the geometry is less consistent, geometric warping is required to obtain accurate registrations.