Learning unbiased registration and joint segmentation: evaluation on longitudinal diffusion MRI.

Bo Li, Wiro J. Niessen, Stefan Klein, M. Arfan Ikram, Meike W. Vernooij, Esther E. Bron

Introduction

Longitudinal imaging studies are essential for monitoring disease progression and aging. However, aligning these images often introduces "processing bias." Most deep learning tools utilize a fixed-reference strategy, where a baseline scan is treated as the gold standard. This creates an asymmetry: the baseline remains untouched while follow-up scans are interpolated, potentially leading to spurious findings in anatomical changes.

The LUGR Framework

To overcome this, researchers developed LUGR (Learning-based Unbiased Group-wise Registration). This framework simultaneously aligns multiple longitudinal scans to an implicit "mean space" while performing automated segmentation.

The LUGR architecture features a constraint projection layer to ensure the reference frame is the geometric center of all scans.

The core of LUGR is a constraint projection layer. Instead of using a soft loss term, it enforces a hard mathematical constraint: the sum of the stationary velocity fields (SVFs) for all images in a group must equal zero. This guarantees that the common reference space is the true geometric center of the group, eliminating the asymmetry of fixed-reference methods. The framework also includes a segmentation network ($F_\Theta$) that predicts a single segmentation in the mean space, which is then warped back to each individual scan's native space to ensure perfect spatial correspondence.

Experimental Results

The framework was evaluated using 8,045 diffusion-weighted MRI (DWI) scans from the Rotterdam Study, focusing on the segmentation of the Forceps Minor (FMI). A subset of 97 individuals scanned twice within a month was used to test for processing bias, as no biological change is expected in such a short window.

Extracted table from page 6 of Learning unbiased registration and joint segmentation: evaluation on longitudinal diffusion MRI.

Key findings include:

  • Reduced Bias: In test-retest data, LUGR showed significantly lower processing bias in volume and diffusion measures (FA/MD) compared to fixed-reference approaches.
  • Superior Consistency: LUGR achieved a voxel-wise agreement (Cohen’s kappa) of 89.30%, compared to 83.21% for fixed-reference methods. This consistency is vital for longitudinal science, where reliable measurement of change is often more important than absolute overlap with manual labels.
  • Efficiency: LUGR provides the speed of deep learning while maintaining the unbiased nature of classical iterative group-wise registration. The mathematical proof of centrality was demonstrated by the norm of the average velocity fields being nearly zero ($2.98 \times 10^{-16}$).

Conclusion and Future Work

LUGR demonstrates that incorporating hard geometric constraints into deep learning architectures can effectively eliminate processing bias in longitudinal studies. While the current architecture is optimized for a fixed number of time-points, it sets a new standard for consistent and unbiased anatomical tracking. Future work will focus on adapting the model for variable numbers of scans and testing it on populations with more dramatic anatomical changes, such as those with advanced Alzheimer's disease.

Summary updated

Abstract

Analysis of longitudinal changes in imaging studies often involves both segmentation of structures of interest and registration of multiple timeframes. The accuracy of such analysis could benefit from a tailored framework that jointly optimizes both tasks to fully exploit the information available in the longitudinal data. Most learning-based registration algorithms, including joint optimization approaches, currently suffer from bias due to selection of a fixed reference frame and only support pairwise transformations. We here propose an analytical framework based on an unbiased learning strategy for group-wise registration that simultaneously registers images to the mean space of a group to obtain consistent segmentations. We evaluate the proposed method on longitudinal analysis of a white matter tract in a brain MRI dataset with 2-3 time-points for 3249 individuals, i.e., 8045 images in total. The reproducibility of the method is evaluated on test-retest data from 97 individuals. The results confirm that the implicit reference image is an average of the input image. In addition, the proposed framework leads to consistent segmentations and significantly lower processing bias than that of a pair-wise fixed-reference approach. This processing bias is even smaller than those obtained when translating segmentations by only one voxel, which can be attributed to subtle numerical instabilities and interpolation. Therefore, we postulate that the proposed mean-space learning strategy could be widely applied to learning-based registration tasks. In addition, this group-wise framework introduces a novel way for learning-based longitudinal studies by direct construction of an unbiased within-subject template and allowing reliable and efficient analysis of spatio-temporal imaging biomarkers.
BibTeX
@misc{li2020learning,
  author = {Bo Li and Wiro J. Niessen and Stefan Klein and M. Arfan Ikram and Meike W. Vernooij and Esther E. Bron},
  title = {{Learning unbiased registration and joint segmentation: evaluation on longitudinal diffusion MRI.}},
  year = {2020},
  eprint = {2011.01869},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.01869},
}
Figure