Information of left atrial (LA) scar provides important information for treatment and management of atrial fibrillation (AF) patients.¹ In clinical practice, however, to extract the LA scar information from late gadolinium enhanced (LGE) MRI is highly challenging due to the complex LA and pulmonary vein (PV) geometry and the limited image contrast. Manual tracing in a slice-by-slice manner is not only time-consuming but also observer-dependent, practically limiting the utilization of LGE-MRI for AF treatment. The purpose of this study is to develop a fully automatic method to segment the LA and PV’s from LGE-MRI, which can facilitate objective LA scar assessment.

During the MR examination of AF patients, a contrast enhanced MR angiography (MRA) sequence is normally acquired prior to LGE-MRI for assessment of the patient anatomy. We propose to take advantage of the high blood contrast in MRA to improve the accuracy and robustness of LA and PV segmentation in LGE. The proposed segmentation method contains three steps: (1) Global segmentation of the MRA: Firstly, each of 10 MRA atlases, with previously obtained LA and PV segmentation, was registered to the given MRA image;² secondly, the known segmentations were propagated to the given image so that each voxel has 10 votes; finally, the majority-vote was made to provide the labeling of the LA and PV regions in the given image. (2) Inter-scan MRA-LGE registration and fusion: Mutual-information based registration was applied to align the MRA to the LGE sequence of the same subject, correcting for patient movement and respiratory motion. The registered MRA was multiplied with LGE to generate a fused volume with enhanced blood signal intensity. The MRA segmentation from the first step was also propagated to the fused volume. (3) Local refinement: With the propagated MRA segmentation as initialization, a 3D level-set approach was applied to the fused volume, refining the initial segmentation to patient-specific details.

The method was validated on a data set of 46 AF patients who underwent preprocedural MRI. During the MR examination, a contrast-enhanced MR angiography (MRA) sequence was acquired for anatomy assessment followed by a LGE sequence for LA scar assessment. An experienced observer manually annotated the LA and PV regions in all 46 LGE sequence in a slice-by-slice manner, as the reference to evaluate the automatic segmentation results. In addition, a second observer also manually annotated the LA and PV regions in 10 subjects for assessment of inter-observer variability. Evaluation of the segmentation performance was in terms of surface-to-surface distance and the Dice overlap index, in the LA and PV regions, respectively.

The LA and PV’s were automatically segmented in all 46 subjects. Compared with manual segmentation, the method yielded a surface-to-surface distance of 1.49±0.65 mm in the LA region, and 2.13±0.67 mm in the PV regions. The difference between the automatic and manual segmentation was comparable to the inter-observer difference, which was 1.36±0.38 mm in the LA region (P>0.05), and 2.12±0.57 mm in the PV regions (P>0.05). The Dice indices were 0.86±0.05 in the LA region between the automatic and manual segmentation, compared with 0.88±0.05 between two observers (P>0.05). The Dice indices were 0.36±0.08 in the PV regions between the automatic and manual segmentation, compared with 0.39±0.09 between two observers (P>0.05). Figure 2 shows the comparison between the automatic and manual results in 2D slice and in 3D volume.

The fully automatic method demonstrated comparable accuracy to the manual segmentation. The agreement was high in the LA region while relatively low in the PV regions, which can be explained by Figure 2. It is also shown that the automatic segmentation achieved better 3D spatial smoothness compared to the result of slice-by-slice manual contouring.

For clinical research, accurate LA and PV segmentation indicates that it is possible to further assess the LA scar from the LGE MRI without manual interference. Figure 3 shows the blood-normalized local LGE intensity,³ sampled on the profile line perpendicular to the surface mesh, mapped onto the segmented LA and PV geometry.

In conclusion, a fully automatic method has been developed to accurately segment LA and PV’s in LGE-MRI, combining a MRA sequence acquired during the same examination. The proposed method makes it possible to automatically generate a patient-specific LA and PV geometry model, and enables display of hyperenhanced tissue in LA wall for individual AF patients.