1348017 - Deblurring 3D LA LGE Images obtained with Stack-of-Stars k-space Sampling using View Sharing and KWIC Filtering

Background: 3D left atrial (LA) late gadolinium enhancement (LGE) was developed for quantification of LA fibrosis[1] in patients with atrial fibrillation (AF). Manual segmentation is the gold standard method to segment the thin LA wall (1-2 mm); therefore, sharp image quality is necessary for accurate LA wall delineation. Our 3D LA LGE sequence is a free breathing sequence that is inherently affected by respiratory motion. To account for this, we split the data into respiratory state bins, perform XD-GRASP[2] reconstruction of multiple bins, and only use one bin for analysis. While using only one bin does improve temporal resolution (reduces motion blur), a lot of data is left unused. One way to exploit the extra data is to use view sharing (VS)[3] which uses portions of data from adjacent respiratory bins. The aim of this study was to investigate the impact of combining VS and KWIC[4] filtering to improve image quality of XD-GRASP reconstruction, which otherwise would suffer from blurring due to overregularization, for our 3D LA LGE protocol.

Methods: Retrospectively, we randomly selected 15 MRI exams from AF patients (age = 66±9 years, 8 males) from an ongoing, prospective study performed on 1.5T scanners (Siemens, AERA or AVANTO). Patients underwent previously described 3D LA LGE with b-SSFP and stack-of-stars k-space sampling[1]. Image reconstruction was performed offline with Matlab (Mathworks, Natick, MA) using XD-GRASP to create 6 equally distributed respiratory bins (Figure 1B). The data was also reconstructed with our proposed respiratory bin design with 50 VS lines and KWIC filter (Figure 1C). The imaging was evaluated using the blur metric (0[best] to 1[worst])[5] to determine relative image sharpness. Pairwise t-test was performed with significance at p< 0.05.  In addition, two board-certified radiologists analyzed randomized, side-by-side images for the reference and proposed binning, for noise, artifact, and conspicuity of LA myocardial wall on a 5-point Likert Scale (1–poor, 3–clinically acceptable, 5–excellent). Each rating was summed to calculate Summed Visual Score (SVS). The raters were also asked to select which image of the pair they prefer. Wilcoxon Signed-Rank test was performed on the average of the raters scores with significance at p< 0.05. 

Results: Blur metric was significantly reduced (p< 0.001) for the proposed images (0.25 ± 0.1) compared to the reference images (0.27 ± 0.1). For visual analysis by the raters, all metrics (noise, artifact, conspicuity, and SVS) were significantly (p< 0.001) improved with the proposed binning (Figure 2, Table 1). Finally, between the two raters, 90% of the time (15/15 for rater 1; 12/15 for rater 2), the proposed binning method imaging was preferred over the reference.

Conclusion: Image quality was significantly improved with the proposed binning technique of VS with KWIC. This sharper image quality might be helpful to improve LA fibrosis quantification our 3D LA LGE scan with both manual and newer automated segmentation methods.