Vessel enhancement in digital X-ray angiographic sequences by temporal statistical learning Andra ´s Lasso ´ a, * , Emanuele Trucco b a Department of Control Engineering and Information Technology, Budapest University of Technology and Economics, Magyar tudo ´sok ko ¨ru ´tja 2, Budapest 1117, Hungary b Department of Electrical, Electronic and Computer Engineering, School of Engineering and Physical Sciences, Heriot Watt University, Riccarton, Edinburgh EH14 4AS, UK Received 20 October 2004; revised 28 February 2005; accepted 28 February 2005 Abstract In this paper, we present a vessel enhancement method, SVM temporal filtering (STF), for X-ray angiographic (XA) images using Support Vector Machine (SVM). We show that the linear SVM applied to vessel enhancement can be regarded as a matched linear filter optimizing the contrast-to-noise ratio in XA images. We propose a non-linear kernel function for the SVM leading to good enhancement with noisy, varying grey-level dynamics at vessel pixels. One key advantage over the matched filters is that an optimal filter is learnt from images, not estimated at design stage. Results on clinical XA images show that learning-based enhancement achieves better results compared to simple subtraction and other image stacking methods. q 2005 Elsevier Ltd. All rights reserved. Keywords: X-ray angiography; Image enhancement; Matched filters; Learning systems; Support vector machines 1. Introduction This paper presents a method for vessel enhancement in X-ray angiographic (XA) image sequences. A Support Vector Machine (SVM) learns the function giving grey- level at vessel pixels in time, due to the flow of a contrast medium. Such function is then used for SVM classification of pixels as vessel or non-vessel. Morphological or anatomical information about the vascular system is essential for enhancement of vascular diseases and planning of surgical procedures or catheter interventions. Vessels, organs of interest and tools (e.g. the tip of the catheter, balloon, stent) have to be displayed with the highest spatial and temporal resolution and fidelity so that their size, relative position and temporal changes can be estimated accurately, while irrelevant background structures (bones, muscles, etc.) should be suppressed. Despite developments in vascular imaging (e.g. MR, PET, US), X-ray [1,2] remains an important diagnostic and therapeutic tool. In X-ray angiography, a radio-opaque contrast agent (also called contrast medium, indicator or bolus) is usually injected into the vessels of interest by means of a catheter. The contrast agent travels through the vessels, making them visible in the X-ray, and is eventually washed out by the blood stream. An X-ray sequence capturing the passage of the agent through a given vessel region gives information about the vessel morphology and the dynamics of the blood flow, but, in general, not all vessels of interest are opacified equally and simultaneously in any frame. Current methods computing morphological information from a XA image sequence tend to utilize only a fraction of the spatio-temporal information available. The use of temporal information (i.e. the evolution of the grey-level values in time) is generally limited to simple averaging and subtraction operations on selected frames. The prevalent technique in clinical practice is Digital Subtraction Angiography (DSA), surveyed by Meijering et al. in [3]; here, we summarize the points relevant for our work. In DSA, a mask frame is acquired before the appearance of the contrast agent, then subtracted from subsequent frames containing opacified vessels (live or Computerized Medical Imaging and Graphics 29 (2005) 343–355 www.elsevier.com/locate/compmedimag 0895-6111/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.compmedimag.2005.02.002 * Corresponding author. Tel.: C36 20 4351999. E-mail addresses: lasso@topcat.iit.bme.hu (A. Lasso ´), e.trucco@hw. ac.uk (E. Trucco).