Motion correction in Digital Subtraction Angiography


Digital Subtraction Angiography (DSA) is a powerful technique for the visualization of blood vessels in the human body. In DSA a sequence of X-ray projection images is taken to show the passage of a bolus of injected contrast material through one or more vessels of interest. The following images show, respectively, an image taken prior to injection (the mask image), and an image containing contrasted vessels (the live or contrast image):

Mask Image Live Image
Mask Image Live Image

The background structures are largely removed by subtracting the mask image from the live image:

Original DSA Image
Original DSA Image

The resulting image is usually displayed using contrast enhancement:

Original DSA Image
Contrast Enhanced DSA Image

The subtraction technique is based on the assumption that during exposure, tissues do not change in position or density. As can be observed from the DSA images shown above, this is not always the case. In practice, patient motion frequently occurs, which causes the subtraction images to show artifacts that may hamper proper diagnosis.

AngioRoom In order to reduce these so called motion artifacts, the misalignment of the successive images in the sequence needs to be determined and corrected for; operations often referred to as registration and motion correction . In clinical practice, this is partially accomplished by pixel-shifting, a rather cumbersome manual technique, which allows for correction of translational motion only.


An example of the limitations of the manual pixel- shifting technique is given in the image below, from which it is clear that if patient motion is more complex (in this case a rotation of the patient's head), a correction in one part of the image implies the introduction of new artifacts in other parts (see yellow arrows):

Motion Correction by Pixel-Shifting

This explains the need for a fully automatic elastic registration method. Although many studies have been carried out on this subject, they have not led to algorithms which are sufficiently fast so as to be acceptable for integration in a clinical setting. We have developed a new approach for registration of digital angiographic image sequences which is both effective, and computationally very efficient.

Registration Approach

In general, the registration of successive images in a sequence involves two operations: (i) Computation of the correspondence between successive pixels; (ii) Applying the correction based on this correspondence.

Since it is computationally too expensive to compute the correspondence explicitly for every pixel, we only consider a reduced set of so called control points. In our approach, the control points are selected in regions containing strong object edges, resulting in an irregular grid. This has three advantages over approaches that use regular grids of control points:

  1. Control points are chosen at those positions where artifacts can be expected to be largest;
  2. The reliability of subsequent displacement computations will be higher;
  3. The number of control points, and thus computation time, is reduced.

The following image shows the regions containing strong object edges in the original mask image (see above). Note that this image gives a very good indication of the regions containing motion artifacts in the subtraction image:

Edge Image

Based on this information the control points are selected, using a two-parameter algorithm which constrains the minimum and maximum distance between the points (white dots):

Control Points

The local displacements at the control points are computed using a template matching approach: using windows in the mask image of about 50x50 pixels, the optimally corresponding window in the live image is searched using a hill- climbing optimization algorithm. In our algorithm we use the energy of the histogram of grey-value differences as a measure of similarity.

Displacement Vectors

In order to correct for motion in the entire image, the computed local displacement vectors need to be interpolated so as to obtain a complete displacement vector field. In our algorithm we use the computationally cheapest approach to do this: linear interpolation. This requires a triangulation of the set of control points:

Triangulation of Control Points

The nice thing about this approach is that it can be mapped very efficiently on modern graphics hardware (programmed using OpenGL), resulting in a real-time implementation of the final warp operation.

O2By repeating the above described technique of displacement computation, interpolation, and warping of the mask image according to the computed displacement vector field, entire sequences can be corrected at a rate of about one image per second, even on relatively low-cost graphics workstations such as Silicon Graphics' O2.

The following two images show, respectively, an original DSA image, and the corrected image after applying the proposed approach:

DSA Sequences
Original DSA Sequence Corrected DSA Sequence

The following two animations show the complete sequence:

DSA Sequences
Original DSA Sequence Corrected DSA Sequence


This page describes work by Erik Meijering, Karel Zuiderveld, Wiro Niessen and others. More information can be found in the following publications:

[1] "Image registration for digital subtraction angiography.", E.H.W. Meijering, K.J. Zuiderveld, M.A. Viergever, International Journal of Computer Vision, 1999, vol. 31, p. 227-246

[2] "Reduction of patient motion artifacts in digital substraction angiography: Evaluation of a fast and fully automatic technique.", H.W. Meijering, W.J. Niessen, J. Bakker, A.J. van der Molen, G.A.P. de Kort, R.T.H. Lo, W.P.T.M. Mali, M.A. Viergever, Radiology, 2001, vol. 219, p. 288-293