Traditionally, cardiovascular interventions have been performed under X-ray fluoroscopy guidance. Recently, the use of MRI as an alternative interventional guidance tool^1-3 has been investigated due to its superior soft tissue imaging contrast. However, guidance with real-time MRI is limited by spatial resolution, whereas high-resolution prior roadmaps do not reflect patient respiratory motion during the intervention. In this study, we propose a novel dynamic motion-modeling technique to correct for this respiratory motion.

Cardiac images were acquired from 8 subjects with a 1.5T GE MRI scanner. Two separate image acquisition protocols were used. For each subject, a high-resolution scan was performed to acquire a 3D prior roadmap image, followed by a low-resolution real-time scan. The prior roadmap consists of a multi-slice stack of images covering the left ventricle (LV). The GE FIESTA sequence was used with resolution=1.2×1.2×8 mm³, and FOV=30 cm. The acquired prior images were gated to end expiration and mid-diastole. For the real-time images, a fast spiral balanced SSFP scan was acquired with in-plane resolution=1.8×1.8×8 mm³, and an effective temporal resolution=67 ms. During the real-time scan, the subjects were able to breath freely, and the cardiac and respiratory physiology signals were acquired through the ECG leads and respiratory bellows respectively.

To correct for the respiratory motion induced misalignment between the real-time and prior images, we adopted the general motion-modeling framework illustrated in Fig. 1. Each prior volume was registered to the ECG-gated (mid-diastole) real-time free-breathing images using a GPU accelerated version of our previously reported multiscale registration algorithm⁴. The individual rigid-body motion parameters were extracted and fitted as a function of the respiratory physiology data (Fig. 1). The derived motion model ϕ can then be used to estimate the motion correction transform M=ϕ(s) based on the given respiratory physiology data s during the intervention. Unfortunately, the limitation of this approach is that the model is built based on images acquired at the beginning of the procedure. This is problematic for respiratory motion compensation during the procedure, as the subject’s breathing pattern may change over time. Therefore, we propose an extension to the above framework, which is to continuously update the model in real-time. The proposed approach is illustrated in Fig. 2. Specifically, for each heartbeat, we performed motion estimation and regenerated a motion model based on the K=10 most recent real-time images and their corresponding respiratory surrogate values by using the previously described framework (Fig. 1). The updated motion model was then used to predict the motion correction estimate for the subsequent input image.

With the GPU accelerated image registration algorithm, each individual motion estimation could be complete in 176.9 ms, while regeneration of the updated dynamic motion model requires approximately 5.3 ms. The image registration alone is not fast enough for correction within the diastolic window of the same cardiac cycle, but it is sufficiently fast for updating the dynamic motion model once every heartbeat. The accuracy of the model based motion correction was evaluated via the Dice similarity coefficient (DSC), which computed the overlap between the manual segmentation contours of the LV endocardium belonging to the corresponding prior and real-time images respectively (Fig. 3). The average perpendicular distance (APD) between the LV endocardial contours was also computed. In retrospective analysis, we generated two subject-specific motion models to correct for the respiratory motion induced misalignment. The first model was built based on the framework illustrated in Fig. 1, where only the real-time images and the corresponding respiratory surrogate values acquired in the initial K heart cycles were used to generate the static model ϕ(s). The second model was built based on the framework illustrated in Fig. 2. This dynamic model ϕ_t(s) was continuously updated after each cardiac cycle. Both models were used for motion correction, and compared to the reference image registration approach, which lags behind by 1 heart cycle. The highest mean DSC value (93.0 ± 3.3%) and the lowest mean APD (1.78 ± 0.84 mm) were achieved using the dynamic model correction method. These results were statistically significant compared to the reference and static model correction approaches (Fig. 4). A comparison of the different motion correction approaches is also demonstrated in a subject where large breathing variations occurred during the real-time scan (Fig. 5). We have presented a dynamic motion-modeling based alignment method that can potentially improve the feasibility of MRI-guided cardiac interventions. Specifically, the proposed correction framework can reduce respiratory induced motion errors by aligning the prior roadmap to the dynamically updated real-time images. This will facilitate improved interventional guidance accuracy in corresponding cardiac procedures.