Longitudinal multi-parametric imaging is increasingly used in oncology applications for monitoring disease progression and treatment response. In longitudinal data, it is not uncommon to encounter large anatomical variations between imaging sessions due to differences in patient positioning, morphologic changes in tumor shape, volume or texture due to tumor growth or treatment intervention, and patient weight loss/gain. Typically, to establish spatial correspondence between image series at various time-points, rigid or non-rigid pair-wise registration is performed with one of the longitudinal time-points chosen as the reference. A common issue with pair-wise registration methods is the introduction of bias by the choice of the reference image, owing to the differences between the reference and other follow-up images. Furthermore, if the reference image exhibits extreme changes in anatomy, the other time-points will endure large deformations to align with the reference image, often leading to questionable registration results, potentially misleading interpretation of outcomes, while making the process computationally expensive. Registration when performed in a group-wise framework removes the need for choosing a reference, thereby offering the potential of consistent registration across time-points [1]. Previously, group-wise registration has been used for motion correction in dynamic contrast study at a single time-point only [3]. In this work, we investigate the utility of group-wise non-rigid registration in the context of longitudinal MRI data used for assessing breast tumor therapy response. We demonstrate that large deformations that are common in longitudinal data are better handled using a reference-free group-wise registration. The results of the group-wise registration approach are compared with those from a conventional pair-wise method (registration of individual time-point image series to baseline).

Patient Data: Two breast cancer patients were scanned at 3 time-points during chemotherapy on GE 3T Discovery MR750 scanner using HD Breast coil. An appropriate IRB approved the study.

Imaging: The imaging data was obtained from dynamic contrast enhancement (DCE) protocol. DCE data was acquired using VIBRANT-TRICKS method with TE/TR = 3.7/7ms, FA=12º, matrix size=512×512×112 (0.68mm×0.68mm×1.4mm resolution), axial orientation, 48 bolus phases with 9.3sec temporal resolution. DCE-MRI motion correction was performed individually on each DCE case by using the methodology as described in [3]. The last bolus phase was considered to be the representative longitudinal MRI dataset at each time point.

Group-wise registration Workflow: Group-wise registration is an iterative process performed as a succession of group-wise Affine and non-rigid B-Spline transformations with mutual information metric available in the Elastix toolkit [2]. All time-points are simultaneously registered to the evolving group-wise mean (acts as the reference image). The computation of the group-wise mean is done by averaging registered images from the previous step, thus iteratively refining it from a blurry image to one with greater clarity. In each iteration, the transform obtained from previous iteration is used as the initial transform to register the individual time-point image with the group-wise mean. Convergence for group-wise registration is deemed to be achieved after completing a set number of iterations (Affine = 5, Non Rigid B-Spline = 10). Since group-wise registration by its very nature moves the group-wise mean (fixed image) and the group (moving images) towards each other, leading to faster convergence. Transforms taken from all of the iterations are composed together to generate the final transform for each of the images.

Pair-wise registration: The mutual information metric was used to register the second and third time points to the reference first time-point in a pairwise manner using Elastix and compared with the group-wise registration approach. Pair-wise registration required 10x the number of iterations as group-wise to converge.

Figure 1 shows the un-registered MRI images at three time points TP#1, TP#2, and TP#3, which do not appear to be in alignment. Both pair-wise and group-wise registrations are able to align anatomical locations across the multiple longitudinal data (Figure 1). While group-wise and pair-wise registration methods provide results that are visually comparable (Figure 1), the deformed boundary of the breast after group-wise registration (Figure 2D) is closer to the original data (i.e TP#3, Figure 2B), compared to pair-wise registration result (Figure 2C). In pair-wise registration, the reference (TP#1, figure 2A) completely dictates the boundary deformation. Correspondingly, this may impact the deformation of internal tissue to match with the reference, potentially changing its morphological characteristics. We have introduced a group-wise registration scheme to synchronize multiple longitudinal time-points in breast exams. The group-wise registration scheme allows for a registration free of reference selection bias. The results of the registration are comparable to that attained using a pairwise registration scheme.