1 of 4 Abstract—This paper describes a new method for the automated segmentation and extraction of cardiac MRI tagging lines. Our method is based on the novel use of a 2D Gabor filter bank. By convolving the tagged input image with our Gabor filters, the tagging lines are automatically enhanced and segmented out. We design the Gabor filter bank based on the input image’s spatial and frequency characteristics. The final result is a combination of each filter’s response in the Gabor filter bank. We demonstrate that compared to bandpass filter methods such as HARP, this method results in robust and accurate segmentation of the tagging lines. Keywords—Gabor filter bank, tagging line segmentation, tagged MRI images I. INTRODUCTION Tagged MRI is a non-invasive technique for the study of cardiac deformation. It generates an MRI-visible tag pattern within the cardiac tissue that deforms with the tissue during the cardiac cycle in vivo and gives motion information of the myocardium normal to the stripes (as shown in figure 1). A difficulty using this technique clinically is the lack of an efficient and robust post- processing method that can automatically segment and track over time the tagging lines. Figure 1: A tagged cardiac MRI HARP [1] is an example of a technique that has been developed for rapid segmentation and analysis of tagged MR images. It generates phase angle images that roughly resemble the original tag pattern. Tagged MR images have a quasi-regular tagging pattern, which leads to relatively isolated peaks in their spectral domain. HARP is basically a bandpass filter that selectively filters those isolated spectral peaks. Although it provides a good direction towards the automated tagline segmentation, HARP has its limitations. Even with the addition of a Gaussian rolloff outside [2], HARP’s bandpass filter is still a relatively global transform in the spatial domain (as shown in figure 2), i.e., HARP’s spatial local transform is affected by regions far away. Also, it is not obvious how to automatically design a bandpass filter that can simultaneously achieve good resolution in both the spatial and the frequency domains. When the first harmonic peak is not well concentrated, HARP has to increase the bandwidth of its bandpass filter. In this case, if the tagging lines deform a lot locally, it would not be robust to use such a wide bandpass filter, which cannot treat regions with and without tag deformations differently. Due to the phase-wrapping artifact [1], HARP is not suitable when large local deformations occur. Another limitation of HARP is that the synthetic tag lines obtained from the phase angle are an approximation to a tag line. Therefore it cannot represent its exact tagline shape, thickness and deformation. Segmenting Cardiac MRI Tagging Lines using Gabor Filter Banks Zhen Qian 1 , Albert Montillo 2 , Dimitris N. Metaxas 1 and Leon Axel 3 1 Department of Biomedical Engineering, Rutgers University, Piscataway, NJ 08854, USA 2 Department of Computer and Information Science, Univ. of Pennsylvania, Philadelphia, PA 19104, USA 3 Department of Radiology, New York University, NY, NY 10016, USA Figure 2: Simplified 1D model of HARP and Gabor filters in frequency domain (left) and spatial domain (right). Upper is HARP; lower is Gabor filter bank. A Gabor filter bank uses the combination of a group of Gabor filters to selectively cover the whole bandpass frequency range; each single filter can still get full constraints in its spatial domain. In this paper, we describe a new method for the segmentation and extraction of tagging lines based on 2D Gabor filters. Gabor filters have been widely used in image processing applications, such as texture segmentation [3, 4, 5] and edge detection [6]. A main advantage of Gabor filters due to their Gaussian envelopes is that they always achieve the minimum space-bandwidth product which is specified in the uncertainty principle [4]. This advantage helps Gabor filters to get full constraints in their spatial domains (as shown in figure 2) as well as in their frequency domain. However, a bandpass method like HARP cannot achieve this. Gabor filters are wavelet-like local filters in the spatial domain, which makes it possible to design adaptive filters with respect to different spatial patterns of different local regions. In this paper we design a bank of Gabor filters with In Proc of International Conference of the Engineering in Medicine and Biology Society, Cancun, Mexico, 2003, pp. 630-633. Acceptance rate: 589 of 1133, 52% for proceedings publication with oral presentation. Presentation by Albert Montillo