Improved motion correction using image registration based on variational synthetic image estimation: application to inline t1 mapping of myocardium

Abstract

Introduction: Myocardial T1 mapping has potential to quantify structural and pathological changes without use of contrast agent. Recently Modified Look-Locker Inversion Recovery (MOLLI) sequence has proven effective on cardiac T1 mapping [1]. Its clinical applicability is still limited by myocardial motion between frames mainly due to beat to beat variations and respiration. This undesired motion compromises the accuracy of pixel-by-pixel T1 estimation. Unfortunately, registration of MOLLI images is particularly difficult as image contrast changes significantly over time (Figure 1). A simple frame-to-frame registration often fails with unrealistic deformation. In this work, we propose a novel registration algorithm based on estimating motion-free synthetic images presenting similar contrast to original data by solving a variational energy minimization problem. Robust motion correction is achieved by registering synthetic images to corresponding MOLLI frames. Method(s): and materials 4 volunteers and 9 patients were scanned (Siemens MAGNETOM Avanto/Espree/Verio, 42/18 pre/postcontrast series). Given N frames of MOLLI images with inversion time TI, synthetic image is defined as a function to minimize the energy functional defined in Figure 2. To estimate the initial signal image, few MOLLI frames with similar contrast are selected and an initial registration and T1 fitting is performed. Final registration is performed between each synthetic image and corresponding MOLLI frame. This process of estimation and registration is empirically iterated twice to correct for residual motion. A fast variational non-rigid registration algorithm [2] is applied here with localized cross correlation as the cost function. All processing steps require no user interaction and are fully integrated into the scanners' reconstruction software. T1 maps are typically available in ~30s after the image acquisition. Validation and results Effectiveness was first evaluated by visual reading. All datasets were classified into two categories: 21 'no significant motion' and 39 'with significant motion'. A frame-to-frame registration between images with largely varying contrast can lead to unrealistic deformation (Figure 3), which was found in 40 cases among the whole cohort (67%), while proposed approach was robust against contrast changes. Quantitative validation was performed on all cases with discernible motion (reconstructed in-plane resolution: 1.67 ~ 2.08 mm2). For every case, two frames with motion were selected and their myocardium was manually delineated. Two measures are computed (Table 1): Dice ratio and MBE (minimal distance between endo/ epi contours of two frames). Typical performance is illustrated in Figure 4-5(Figure Presented).