Optimizing Brain Tissue Contrast with EPI: A Simulated Annealing Approach Vasiliki N. Ikonomidou, * Peter van Gelderen, Jacco A. de Zwart, Masaki Fukunaga, and Jeff H. Duyn A new magnetization preparation and image acquisition scheme was developed to obtain high-resolution brain images with optimal tissue contrast. The pulse sequence was derived from an optimization process using simulated annealing, with- out prior assumptions with regard to the number of radiofre- quency (RF) pulses and flip angles. The resulting scheme com- bined two inversion pulses with the acquisition of three images with varying contrast. The combination of the three images allowed separation of gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) based on T 1 , contrast. It also enabled the correction of small errors in the initial T 1 estimates in post- processing. The use of three-dimensional (3D) sensitivity-en- coded (SENSE) echo-planar imaging (EPI) for image acquisition made it possible to achieve a 1.15 3 mm 3 isotropic resolution within a scan time of 10 min 21 s. The cortical GM signal-to- noise ratio (SNR) in the calculated GM-only image varied be- tween 30 and 100. The novel technique was evaluated in com- bination with blood oxygen level-dependent (BOLD) functional magnetic resonance imaging (fMRI) on human subjects, and provided for excellent coregistration of anatomical and func- tional data. Magn Reson Med 54:373–385, 2005. Published 2005 Wiley-Liss, Inc. † Key words: brain tissue labeling; T 1 weighting; optimization (simulated annealing); functional mapping; double inversion re- covery The excellent soft-tissue contrast provided by MRI is one of the main advantages of this technique over X-ray CT, especially in imaging of the brain. It allows one to distin- guish between different tissue types, as well as to obtain detailed anatomical maps of human cortical architecture. Brain tissue can be classified into three broad categories: white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). Their relative volume and precise location convey potentially important information for the diagnosis and treatment of disease, as well as for the general under- standing of brain anatomy and function (1). Tissue char- acterization relies heavily on high-contrast MR images, which can be obtained by techniques such as T 1 -weighted magnetization-prepared rapid gradient-echo (MP-RAGE) (2,3). These techniques aim to maximize the intensity dif- ference between WM and GM, and at the same time main- tain a high enough signal-to-noise ratio (SNR) to safely distinguish those regions from the suppressed CSF signal. Typically, an image segmentation algorithm (1), which is usually intensity-based, complements this approach. Even though the three tissue types differ in terms of both proton density and relaxation times, which should result in well-defined distinct intensity values for each category, the segmentation problem is by no means trivial. The tissue classification process is often hampered by issues such as limited contrast-to-noise ratio (CNR), resolution, and partial volume effects. Furthermore, intensity inhomogeneities, even within the same type of tissue, severely hamper tissue separation. A significant part of the spatial variations can be attributed to inherent system limitations, such as radiofre- quency (RF) coil inhomogeneities and noise, as well as to normal variation in tissue homogeneity (4,5). Partial volume effects arise when a voxel is composed of a mixture of different tissue types. Since the signal is a weighted sum of different tissue types, the resulting inten- sities can fall in between the signals originating from these different tissue values, causing ill-defined edges and dis- persion of the expected intensity values. In this case, ei- ther multiple images or prior knowledge combined with modeling are needed in order to classify pixels as belong- ing to one or more tissue categories (6 – 8). A further difficulty comes into play in the context of functional mapping in applications such as BOLD fMRI. Functional data are typically acquired using echo-planar imaging (EPI), which is known to be susceptible to geo- metric distortions due to off-resonance effects. Anatomical images, on the other hand, are normally acquired with multishot acquisitions (e.g., fast low-angle shot (FLASH) or fast spin echo (FSE)), which have relatively little geo- metric distortion. This introduces a further level of com- plexity to the coregistration problem. Imaging of a single-tissue type in the brain by zeroing two tissue types with the use of a double inversion recov- ery (IR) sequence (9) was originally proposed by Redpath and Smith (10). Double IR is efficient in dealing with partial volume effects because the signal from the un- wanted tissue types is suppressed (ideally close to zero), and tissue suppression is not dependent on reception field inhomogeneities, thus bypassing one of the causes of prob- lems in traditional tissue segmentation. However, its effi- cacy depends on the accuracy of the a priori T 1 estimation and the uniformity of T 1 in the desired tissue types. In this paper, a new optimized sequence is presented that further elaborates on the concept of zeroing tissue signal based on T 1 relaxation. Our initial motivation was to derive a high-quality GM image that could serve as an anatomical reference for combination with BOLD fMRI data for the pur- pose of functional mapping. Given practical constraints, in terms of both minimizing subject motion and maximizing the time available for the functional experiment, the technique Advanced MRI Section, LFMI, NINDS, National Institutes of Health, Bethesda, Maryland, USA. *Correspondence to: Vasiliki N. Ikonomidou, Advanced MRI Section, LFMI, NINDS, National Institutes of Health, Bldg. 10, Rm. B1D-722, MSC 1065, 9000 Rockville Pike, Bethesda, MD 20892-1065. E-mail: viko@nih.gov Received 4 November 2004; revised 8 February 2005; accepted 4 March 2005. DOI 10.1002/mrm.20561 Published online in Wiley InterScience (www.interscience.wiley.com). Magnetic Resonance in Medicine 54:373–385 (2005) Published 2005 Wiley-Liss, Inc. † This article is a US Government work and, as such, is in the public domain in the United States of America. 373