Two-point fat water separation methods are preferred in some fast applications such as dynamic imaging and abdominal imaging where fat needs to be suppressed homogeneously. The original two-point method proposed by Dixon¹ acquired two images at echo times (TEs) with fat and water in and out of phase. The two-point methods are then generalized to flexible TEs^2,3. Although these methods have been applied successfully to many cases, there still exist challenges with multiple disjoint regions or drastic change of B₀. In this paper, a novel two-point method using safest path region growing with self-feeding phasor estimation algorithm is introduced for robust fat water separation.Our method has three main features: multiresolution scheme, safest path region growing scheme and self-feeding mechanism. Firstly, the seed pixels identification for region growing is performed independently at different resolutions. The lower resolution source images are obtained by spatially filtering the original images with low pass filters and down-sampling at a certain rate [M_j,N_j], j=1,2,3, …,J. Then, the candidate phasor solutions for each pixel are calculated using the method proposed by Eggers et al². Seed pixels are identified according to two criteria: (1) adequate SNR; (2) uniqueness of the true solution. The advantage of multiresolution scheme is that more seed pixels can be found in lower resolution in assumption of field smoothness, as indicated in Figure 1 using simulation data. Second, for the region growing scheme, the safest path is maintained based on the SNR of the pixels, the number of neighboring seed pixels, and threshold for phasor differences between neighboring pixels, which makes the region growing path more robust in low SNR areas. Starting from the seed pixels, the region growing scheme is executed to obtain the phasor map in each resolution. The low resolution phasor maps are then propagated to the finest resolution directly. For each pixel at the finest resolution, the phasor closest to the phasor solution of the corresponding pixel at lower resolution is chosen. Since a group of phasor maps at different resolutions are obtained, the propagation step is also implemented independently between different resolutions, and a group of phasor maps at the finest resolution is consequently derived. Third, a self-feeding mechanism is introduced to merge the phasor maps into one phasor map. The pixels with the consistent phasor solution from all the resolutions are identified as seed pixels at the finest resolution, whereas the left pixels with inconsistent phasor solution are reset to underdetermined, as shown in Figure 2. The region growing scheme is applied again to obtain the final phasor map followed by smoothing. The “self-feeding” mechanism ensures the reliability of seed pixels selection at the finest resolution. Figure 3 shows the flowchart of the proposed method.

The proposed method was tested on c-spine and abdomen data. Two volunteers with informed consent (IRB approved) were recruited. The MRI scan in c-spine/abdomen was implemented on a 3T system (Siemens, TIM TRIO, Erlangen, Germany) with a FLASH sequence. The basic protocols for the study were (c-spine/abdomen): TR = 200/50ms, matrix = 320×240/256×152, slice thickness = 3/6mm, flip angle = 30°/25°, TE = 4.64/5.51ms.

The algorithm was implemented in MATLAB (Mathworks, NATICK, USA). Four resolutions with down-sampling rates [4 4], [6 6], [8 8] and [10 10] were used to obtain the coarser resolution images.

In all slices, fat/water images and phasor maps were successfully derived from two-point images. Figure 4 shows the separated fat/water images and the phasor maps of a sagittal slice of c-spine and a transversal slice of abdomen. The results show that the algorithm works well with the disjoint area (such as chin, liver and arm) and fast varying field area in the neck.Although the SNR of the reconstructed water/fat images by the proposed two-point method tends to be lower in comparison with three-point methods, the scan time is decreased as the proposed method is a flexible two-point method with no constraint on TEs. Different from traditional two-point methods, the phasor map in this paper is estimated by multiresolution region growing scheme where the seed pixels identification and region growing scheme is performed independently between coarser resolutions, avoiding the erroneous propagation between resolutions. Besides, the “self-feeding” mechanism when merging the phasor maps ensures the reliability of seed pixels selection at the finest resolution. A novel two-point fat water separation method using safest path region growing with self-feeding phasor estimation algorithm is proposed. It is shown that the algorithm is robust when applied to regions with fast varying field and disjoint areas.