Cardiac Phase-resolved Blood Oxygen-Level-Dependent (CP-BOLD) MRI¹ is a new contrast and stress-free approach for detecting myocardial ischemia, that identifies the ischemic myocardium by examining changes in myocardial signal intensity as a function of cardiac phase. Visualizing and detecting these changes requires significant post-processing, including myocardial segmentation for isolating the myocardium and registration to establish correspondence among the cardiac images in this cine acquisition. Automated analysis approaches which can obtain pixel-level determination of ischemia are desirable, since they may lead to improved transmural assessment of the extent of ischemia. But, changes in myocardial intensity patterns and myocardial shape (due to the heart's motion) challenge automated standard CINE MR myocardial segmentation and registration techniques resulting in a significant drop of segmentation and registration accuracy. We hypothesize that using a multi-resolution dictionary based registration approach and iterative refinement can improve myocardial segmentation accuracy. In this work, myocardial segmentation accuracy is evaluated on CP-BOLD data at rest under baseline and stenosis conditions.Imaging Studies: Flow- and motion-compensated 2D short-axis CP-BOLD was acquired along the mid-ventricle in 10 canines at baseline and under severe LAD stenosis. Imaging studies were performed on a 1.5T Espree (Siemens Healthcare). TR/TE 6.2/3.1ms; spatial resolution=1.2x1.2x8 mm3, flip-angle=70°, ~25 frames per cardiac cycle. Image Processing: The main principle of the approach (Fig. 1) is to use an external initial database of pre-segmented images (and dictionaries for myocardium and background) to first come up with an initial segmentation (relying on classification when projecting on discriminatory dictionaries) for an unseen CP-BOLD dataset (cine acquisition).³ Then it registers images in a multi-resolution fashion across the cardiac cycle using a new registration algorithm⁴ that relies on the sparse coefficients. This step refines the obtained segmentation, which is used to update the dictionaries (adapting them to the dataset under consideration), and thus personalize the process. This process is repeated till convergence. To obtain an initial dictionary on prior available segmented data sets, at the coarsest resolution, first 9x9 patches are extracted for each pixel and concatenated with HOG and Gabor features for each patch and dictionaries D_B and D_M are learned. Multi-Resolution Registration Refinement: As highlighted in Fig. 2, for each pixel p when projected on the background or myocardium dictionaries sparse coefficients X_B and myocardium X_M are obtained and concatenated; X_p =[ X_B ;X_M ]. They are used to define a similarity metric between two images as: $$$ S(I_{t}(p),I_{t+1}(p+u))= \mid \mid X^{t}_{p}-X^{t+1}_{p+u} \mid \mid_{1} $$$ .⁴ Starting at the coarsest resolution, after all images in the cardiac cycle have been registered, individual per-image segmentations of the myocardium are obtained and propagated to the first image. Then labels are fused (via majority voting) to refine the segmentation of that image. The segmentation is upscaled and using the corresponding intensity image, new dictionaries D’_B and D’_M specific to this image resolution are obtained (Fig. 3). The process iterates using these new dictionaries and sparse representations. At the finest resolution all images I_t+j, j≠0 are registered to I_t, segmentations obtained via the dictionaries for I_t±j are propagated to I_t and fused with majority voting to obtain the final segmentation. Statistical analysis: To evaluate segmentation accuracy, delineations (ground truth) provided by experts are used. For a given unseen dataset, overlap of the segmentation of I_t with corresponding ground truth is measured using the Dice overlap metric, and averaged across all t. Multiple paired t-tests were applied to evaluate performance when compared to state-of-the-art registration methods, namely Diffeomorphic Demons⁵, Free Form Deformations (FFD)⁶, DRAMMS⁷ and MIND⁸.Dice accuracies are presented in Table 1, across our study population. Utilizing the proposed scheme increases myocardial segmentation accuracy under baseline (^#, p<0.001) and ischemia (^$, p<0.001) significantly. Exemplary results are shown in Fig. 4 highlighting the importance of multi-resolution and iterative refinement.The scheme herein improves myocardial segmentation accuracy in a statistically significant manner by refining and adapting the dictionary to the data using multiple resolutions. When only an external dictionary is used the resulting performance is not different to other methods. However, taking advantage of registration to refine the segmentation results leads to an improvement by 8%. Operating at multiple resolutions (capturing more context and spatial information and avoiding local minima) leads to even better improvement (15%).BOLD contrast affects the performance of image analysis algorithms and here a new scheme of iterative segmentation/registration is proposed. While the results are remarkable, further improvements are necessary to enable pixel-level assessment of ischemia. When combined with developments in 3D BOLD acquisitions a repeatable, truly non-invasive diagnosis of ischemic heart disease can be made possible.⁹