TITLE:

OPTIMIZATION METHODS IN MEDICAL IMAGE RESTORATION (No. 70)

DATE:

Friday, February 11th, 2005

TIME:

3:30 PM

LOCATION:

GMCS 214

SPEAKER:

Astrid Franz, Philips Research Laboratories, Hamburg, Germany

ABSTRACT:

Registration of medical images means the integration of one or more images into a common geometrical system of reference so that corresponding image structures in the various images correctly align. The images may have been acquired with the same or different imaging modalities, simultaneously or at different times, from one or several patients. For instance, images taken at different times for the same patient, when correctly aligned, give information about tumor growth or treatment success. Another example is the combination of anatomical (Computed Tomography) with physiological (Positron Emmission Tomography) data, which has a dramatic impact on the early diagnosis and prognosis of cancer and inammatory diseases as well as on the planning and monitoring of therapy.

The first step to registration is to compensate for translations and rotations between the input images. This so-called rigid registration has become a standard technique and is deemed to be solved. However, the human body is not rigid. Hence an elastic mapping has to be found which optimally aligns the given images.

The elastic registration process consists of four fundamental constituents:

– pre-processing
summarizing all processing the input images undergo mainly to turn them more similar in view of the similarity measure (see below),

– a similarity measure
quantifying the comparison of the similarities between images, e.g. correlation or entropy based measures,

– a class of geometrical transformations
modelling displacements under the control of a number of parameters, e.g. various kinds of spline functions, and

– an optimization algorithm
determining that geometrical transformation that leads to optimal similarity between the images.

The similarity measure, over a high-dimensional transformation parameter space, in general features a lot of local optima. Hence a good starting point for local optimization strategies has to be known, or global optimization algorithms have to be applied. Performance measures for local as well as global stochastic optimization strategies are investigated for the registration of medical images. Test computations show the potency of global optimization for the elastic registration task, but also the challenges as reproducibility, identification of the solution that is of relevance for the clinical target application, and an acceptable computation time.

HOST:

Peter Salamon

DOWNLOAD: