Recent advances in myocardial MOLLI T1 mapping¹ allows quantitative estimation of the extracellular volume fraction (ECV)² of myocardium. Accurate T1 estimation and deformable image registration are both crucial for ECV mapping. This study improved the procedure of T1 fitting and developed an image registration based on segmentations of the left-ventricle (LV) walls to improve the accuracy of ECV mapping.Nine participants underwent the study performed on a 3T MRI system (Skyra, Siemens) after providing informed consents. Two sets of MOLLI images were obtained before and after contrast administrations using the parameters (TR/TE: 295.26/1.12 ms, matrix: 256×152, flip angle: 35°, MOLLI acquisition scheme: 5-3, 3 slices: basal, mid-cavity, apical, waiting time for postcontrast imaging: 10-20 min). The level of hematocrit (Hct) was obtained using a blood sample. The image processing was performed in the Python software environment. Figure 1 displays the flow diagram of the procedures. We first performed a deformable registration (software package: SimpleITK, transformation: spline, similarity measurement: mutual information) to register the pre- and postcontrast data sets using the images with the longest inversion time (TI) as reference images. The registration method was similar to the previously presented method². We then applied pixel-by-pixel T1 fitting on the registered images with multiple initial T1 values (100, 200, 500, 800, 1000, 1500, 2000, 2500 ms). The T1 value calculated with the least error was selected. This T1-mapping procedure with the above registration method was termed the method 1 (M1). In addition, we implemented a registration procedure based on segmentation of the LV walls. The segmentation procedures will be detailed in the abstract entitled “Fully automatic bullseye analysis on short-axis MOLLI mapping: LV segmentation and AHA 17 parcellation”. In brief, we first synthesized images corresponding to different inversion-recovery times (TI: 0 to 4250 ms) according to the estimated T1 recovery model and calculated the null point of the T1 recovery curve. Utilizing the T1 characteristics and applying morphological processing, we calculated a fat-liver mask, a heart ROI, and a mask of the LV blood pool region. We developed a layer-growing method to obtain a mask of LV walls, registered the masks using the demons registration³, applied the transformations to the pre- and postcontrast images and calculated the corresponding T1 maps. This procedure to was termed the method 2 (M2). We used the T1 maps obtained using M1 and M2 to produce ECV maps [ECV = (1−Hct) × (ΔR1myocardium/ΔR1blood)]. In order to evaluate the accuracy of the registration results, we manually outlined the ROIs of the LV walls on four types of MOLLI images [pre-contrast; original postcontrast (Org); postcontrast images registered with M1 and M2] to calculate an overlap rate (the number of the overlapped pixels of pre- and postcontrast masks divided by the total pixel number of the pre-contrast mask).Figure 2 displays the error maps obtained using multiple (Fig. 2a) and single (Fig. 2b) initial values. The fitting errors achieved with multiple initial values were prominently reduced. Figure 3 displays the comparisons of M1 and M2. The regions of the LV walls show aligned better with the M2 mask than with the other two masks (Org and M1). The ECV map obtained using the M2 method appears more homogeneous than those obtained with the other two data sets. Group analysis reveals that the overlap rates (Org: 54.7 ± 22.8%, M1: 71.9 ± 9.6%, M2: 86.3 ± 6.7%) obtained using the M2 data set are significantly higher than those obtained using the org or the M1 data sets (p < 0.01, paired t test, N = 9 participants × 3 slices = 27).For the results of T1 estimation, the errors were prominently reduced using the multiple initial values. For the image registration, this study proposed a deformable registration method in cooperated with the automatic LV segmentation. The results supported that the M2 method significantly improved the overlap rate. The improved results are most likely due to the clear boundary between the LV walls and the blood pool region in the LV mask, which increase the performance of the deformable image registration. The obtained LV masks are potentially useful to automatically segment the myocardium into bullseye diagram such as the American Heart Association (AHA) 17 segments. It warrants further investigations. In conclusion, we developed a method to improve the image registrations in the short-axis ECV data sets. This method could be a practical method to improve the accuracy of quantitative ECV analysis in the myocardium.