Source-linked AI summary
Longitudinal tracking of multiple sclerosis lesions in the spinal cord: A validation study
Pierre-Louis Benveniste, Julian McGinnis, Shannon Kolind, Larry D. Lynd, Sarah A. Morrow, Jiwon Oh, Alexandre Prat, Alice Schabas, Penelope Smyth, Roger Tam, Anthony Traboulsee, Mark Mühlau, Herve Lombaert, Julien Cohen-Adad
TL;DR
The paper addresses the lack of consistent instance-level correspondence for tracking spinal-cord MS lesions through longitudinal MRI, where lesions may split, merge, or change representation. It compares five registration-based and registration-free tracking strategies and finds the registration-based overlap method performs best, while anatomical-coordinate classification remains competitive. The study supports flexible asymmetric matching but notes uncertainty from cohort size and segmentation noise.
Problem
Existing longitudinal MS segmentation frameworks do not consistently correspond lesion instances across time, despite the need to characterize individual lesion evolution.
Method
The study compares five automated tracking strategies using deformable registration or spinal-cord anatomical coordinates, including overlap matching, Hungarian assignment, gradient-boosted classification, and Siamese classification.
Results
Registration-based non-zero-overlap matching achieved the highest overall tracking performance, while the coordinate-based XGBoost strategy remained a competitive registration-free alternative.
Takeaways & Limitations
Flexible asymmetric matching is important for longitudinal lesions that undergo splitting or merging, with the preferred paradigm depending on registration quality and mapping complexity.
Takeaways & Limitations
The cohort size was borderline for non-parametric significance testing, and segmentation noise could cause tracking failures.
Abstract
from arXiv · showhide
Longitudinal characterization of multiple sclerosis (MS) lesions remains constrained by the lack of frameworks capable of establishing consistent instance-level correspondences across time. Conventional segmentation approaches produce semantic lesion masks at each visit and therefore fail to capture the complex instance temporal patterns associated with lesion appearance, disappearance, splitting, or merging. This study presents a comparative evaluation of five strategies for automated tracking of spinal cord MS lesions in longitudinal MRI data from a multi-site cohort. The investigated strategies rely either on deformable registration or on a spinal anatomical reference system, and encompass overlap-based matching, coordinate-based Hungarian algorithm, gradient-boosted classification, and Siamese model classification. Tracking accuracy is quantified using instance-level true positives, false positives, and false negatives, allowing to assess the presence of one-to-many and many-to-one associations. Results show best performance for the registration-based overlap method. This study provides the first systematic analysis of lesion-instance correspondence in the spinal cord and outlines the strengths and limitations of registration-based and registration-free paradigms for longitudinal MS assessment. The code is available at http://github.com/ivadomed/longitudinal-sc-ms-lesion-tracking .
1. Introduction
Longitudinal spinal-cord MS lesion assessment requires instance-level correspondences because lesions can split, merge, or be represented differently across visits. This study evaluates multiple automated tracking frameworks across multi-site longitudinal MRI data.
- Instance-level tracking complements lesion segmentation by enabling lesion-specific measurements and characterization of how individual lesions evolve over time.
- Existing longitudinal segmentation frameworks do not establish consistent lesion-instance correspondences, leaving no dedicated methodology for tracking MS lesions in the spinal cord.
- Lesion tracking must accommodate one-to-many and many-to-one evolution, as well as segmentation variability that can produce different longitudinal representations of the same pathology.
- Spinal-cord tracking is especially challenging because non-linear alignment, small lesions, partial-volume effects, and minimal misregistration can substantially alter lesion boundaries.
- The study validates five automated frameworks spanning registration-based and registration-free approaches, including Hungarian matching, anatomical coordinates, and deep learning, using multi-site clinical datasets.
2. Related Works
Prior longitudinal MS work has largely addressed segmentation rather than lesion-instance tracking, especially in the spinal cord. Existing tracking approaches from other domains often assume stable appearances and simple correspondences, whereas spinal cord lesions require methods that accommodate complex evolution.
- Existing longitudinal MS frameworks focus on segmentation and do not perform lesion tracking, with no prior work addressing spinal cord MS lesion tracking.
- Computer-vision tracking methods commonly assume 2D single-object trajectories, stable appearance, and minimal topological change.
- Only a few medical studies explicitly address asymmetric correspondences, including cancer-lesion methods using clustering, geometric features, and Hungarian matching.
- Spinal cord MS lesions pose challenges including small size, morphological variability, rater variability, and dependence on non-rigid registration.
- These gaps motivate tracking frameworks that explicitly model one-to-many and many-to-one lesion evolution, potentially using deep learning to improve temporal consistency.
3. Methods
The study evaluates five spinal cord lesion-tracking strategies spanning registration-based and registration-free designs, anatomical-coordinate matching, supervised classifiers, and overlap-based correspondence. Longitudinal lesion mappings are evaluated at the instance level against expert-defined ground truth that captures asymmetric lesion evolution.
- The cohort includes 34 participants scanned at baseline and 12-month follow-up across five hospitals using Philips, Siemens, or GE systems.
- Five strategies span registration-based versus registration-free tracking and one-to-one versus asymmetric matching schemes.
- Registration-free strategies represent lesions in cylindrical spinal-cord coordinates and use weighted Hungarian matching, XGBoost classification, or Siamese classification.
- Registration-based strategies warp follow-up segmentations into baseline space using disc-level pre-alignment followed by deep-learning registration, with visual quality control.
- The overlap strategy pairs lesions when warped masks have non-zero IoU, allowing asymmetric one-to-many and many-to-one correspondences.
- Performance is measured against expert longitudinal mappings using instance-level true positives, false positives, and false negatives, including split-lesion cases.
4. Results
Across test-set, manual-segmentation, and leave-one-out evaluations, Strategy #5, registration with IoU-based correspondence, performed best overall, while Strategy #2 remained competitive. The principal limitations were sensitivity to registration and segmentation errors, and limited statistical power for distinguishing the two leading strategies.
- In a high-lesion-burden example, Strategies #2 and #5 produced accurate correspondences, whereas Strategy #3 showed substantial mismatches.Strategies #1 and #4 were constrained by one-to-one Hungarian assignments and could not recover the illustrated many-to-one trajectory.
- Strategy #5 achieved the highest test-set performance, reaching an F1 score of 0.98 with registration-based IoU correspondence.
- On the independent test set, Strategy #5 did not significantly outperform Strategy #2 across any metric (p > 0.05).
- Strategy #5 remained superior on manual segmentations with F1 0.91, while Strategy #2 remained competitive at F1 0.81.The overall performance hierarchy remained unchanged despite a slight reduction in scores across strategies.
- In LOOCV, Strategy #5 ranked highest across all metrics and significantly outperformed other strategies in Precision and F1, except Strategy #2 for F1 (p = 0.062).The analysis used 34 subject-wise iterations with corrected pairwise Wilcoxon tests.
- Strategy #5 failed in representative cases because segmentation noise and small registration or partial-volume shifts disrupted overlap-based correspondence.The cohort size was also described as borderline for non-parametric significance testing.
5. Discussion
The registration-based IoU strategy achieved the strongest lesion-tracking performance, while registration-free geometric features offered a competitive alternative. The discussion identifies correspondence flexibility, registration and segmentation quality, anatomical references, and evaluation metrics as key determinants and limitations.
- The non-zero-overlap IoU criterion suits lesions undergoing large volumetric changes, splitting, merging, or diffuse confluence.Recall decreased from 0.96 to 0.42 as the IoU threshold increased from near-zero to 0.5, while Precision remained perfect.
- Strategy #5 remains dependent on segmentation and registration fidelity, so its performance partly reflects how accurately anatomy is aligned and lesions are delineated.Failure cases include segmentation-induced false negatives, interpolation-related boundary changes, and subtle spatial misregistration.
- Flexible asymmetric matching is needed because MS lesions commonly undergo one-to-many and many-to-one evolution, unlike Hungarian methods’ enforced one-to-one mappings.The proposed pairwise correspondence framework can extend to more than two scans by sequentially composing links.
- Strategy #5 achieved the highest tracking performance, with F1=0.97 in LOOCV and F1=0.98 in the reported study comparison.It significantly outperformed the other strategies in Precision and F1 in LOOCV, except for the F1 comparison with Strategy #2.
- Strategy #2 was the strongest registration-free alternative, reaching F1=0.91 and showing that geometric descriptors can encode longitudinal information without deformable registration.Its higher false-positive rate may reflect over-confident predictions associated with early-stage lesion merging.
- The coordinate-based strategies depend on intervertebral-disc labeling and a disc-based reference system that is sensitive to positioning and cervical-curvature differences.No disc-labeling failures occurred in the current cohort, but such failures would invalidate subsequent tracking.
- Isotropic resampling or super-resolution is proposed as a way to reduce failures caused by slice-direction misalignment in anisotropic native-resolution MRI.The cited native resolution is 0.7 × 0.7 × 3 mm, which can make sub-voxel slice shifts eliminate voxel overlap.
- Existing segmentation and multi-object tracking metrics do not adequately evaluate longitudinal lesion correspondence with complex temporal mappings.The study therefore emphasizes dedicated link-based evaluation for associations that are not strictly one-to-one.
6. Conclusion
The study presents the first spinal-cord investigation of longitudinal instance-level MS-lesion tracking and finds registration-based overlap performs best overall.
- Registration-based overlap achieved the highest overall performance for longitudinal spinal-cord MS-lesion correspondence.The result indicates spatial alignment is powerful but imperfect for associating lesions across time.
- The study formulates lesion tracking as a correspondence problem rather than solely a segmentation task.This addresses one-to-many and many-to-one associations in evolving lesions.
- The competitive spinal-cord-coordinate XGBoost model suggests that meaningful longitudinal information can be extracted without deformable registration.The authors identify registration-free approaches as an opportunity for future work.
Ethical Standards
The manuscript states that the research followed applicable ethical standards, laws, and regulations for studies involving animals or human subjects.
- The research and manuscript followed appropriate ethical standards.
- The work followed applicable laws and regulations concerning animals or human subjects.
- The ethical statement covers both research conduct and manuscript writing.
Conflicts of Interest
The authors disclose industry honoraria, advisory roles, and research funding involving several pharmaceutical companies.
- Penelope Smyth reports speaker honoraria and advisory-board participation for EMD Serono Canada, Novartis, Roche, and Amgen Canada Pharmaceuticals.
- Mark Mühlau reports receiving speaker honoraria from Merck.
- Jiwon Oh reports research funding from Biogen-Idec and Roche relevant to the work.
Appendix A. Results on all the data splits
Appendix A presents longitudinal lesion-tracking performance across five strategies using predicted lesion segmentations.
- Table 4 covers longitudinal lesion-tracking performance across five strategies.
- The reported evaluation uses predicted lesion segmentations.
- The table concerns longitudinal lesion tracking rather than standalone lesion segmentation.
Appendix B. Per-site results of Strategy #5
This appendix reports per-site lesion-tracking performance for Strategy #5.
- Per-site lesion-tracking performance is reported for Strategy #5.
Appendix C. Per-participant results of Strategy #5
The appendix provides table-based analyses of Strategy #5, including per-participant performance, cross-strategy tracking results, IoU-threshold sensitivity, and registration-direction effects.
- Table 6 reports per-participant average lesion-tracking performance for Strategy #5.
- Table 7 compares longitudinal lesion-tracking performance across the five evaluated strategies.
- Tables 8 and 9 examine Strategy #5 sensitivity to IoU threshold and the effect of registration direction.