Source-linked AI summary
Is Deformable Image Registration Ready for Brain Metastasis Reirradiation Dose Accumulation? A Longitudinal MRI Benchmark of Registration Accuracy
Hengjie Liu, Manju Sharma, Xinyi Fu, Di Xu, Ke Sheng
TL;DR
Dose accumulation for brain-metastasis reirradiation depends on reliable deformable registration, yet evidence in this clinical setting is limited. This study benchmarks learning-based and optimization-based methods on longitudinal MRI, finding that pair-specific adaptation improves pretrained models but optimization-based methods remain strongest overall. The results indicate that unmonitored DIR is not yet sufficiently reliable for dose accumulation and requires case-level quality control and uncertainty assessment.
Problem
Registration errors can propagate into accumulated dose estimates, while DIR performance may not transfer reliably from healthy-brain MRI benchmarks to recurrent-tumor scans.
Method
The study benchmarks learning-based and optimization-based DIR on longitudinal contrast-enhanced T1-weighted MRI, evaluating zero-shot inference, ISO, TSO, geometric accuracy, volume recovery, runtime, and memory.
Results
Optimization-based methods remained strongest overall, while ISO and TSO improved pretrained learning-based methods; modern optimization methods also offered practical runtime and memory tradeoffs.
Takeaways & Limitations
DIR is not yet sufficiently accurate and consistent for unmonitored use in brain-metastasis reirradiation dose-accumulation workflows.
Takeaways & Limitations
The study evaluates lesion correspondence but not downstream dose or outcome analyses, which remain future work.
Abstract
from arXiv · showhide
Dose accumulation is increasingly important in adaptive radiation therapy and reirradiation, but its clinical validity depends on the performance of deformable image registration (DIR). Reirradiation of brain metastases (BMs) with stereotactic radiosurgery (SRS) provides a controlled but clinically meaningful DIR test case: intra-subject brain deformation is usually limited after rigid alignment, yet recurrent lesions can undergo substantial local shape and volume changes that rigid registration cannot capture and can affect dose accumulation. We benchmarked a wide range of learning-based and optimization-based DIR methods on 87 manually screened longitudinal contrast-enhanced T1-weighted MRI lesion pairs from an institutional BM SRS retreatment cohort. Learning-based methods pretrained on healthy-brain MRI were evaluated zero-shot and after instance-specific optimization (ISO) or tumor-proximity target-specific optimization (TSO). Registration was assessed using lesion overlap (Dice), surface distance metrics (HD95 and sASD), target-volume recovery, and runtime and memory. Pretrained learning-based methods showed variable zero-shot performance, while ISO/TSO improved all tested learning-based families. However, optimization-based methods remained the best-performing approach while maintaining reasonable runtime. These findings suggest that even state-of-the-art DIR methods do not yet provide sufficiently accurate and consistent registration for unmonitored use in brain metastasis reirradiation dose accumulation. Because accurate registration is a prerequisite for deformable dose accumulation, clinical application will require case-level quality control and direct assessment of how registration uncertainty affects downstream dose metrics.
1 Introduction
DIR is important for dose accumulation in adaptive RT and reirradiation, but its clinical reliability remains difficult to establish. Brain-metastasis reirradiation offers a clinically relevant test setting, and this study benchmarks DIR methods on longitudinal MRI while examining whether healthy-brain pretraining transfers to recurrent tumors.
- Motivation: DIR errors can propagate into accumulated dose estimates because voxel-level ground truth is rarely available and registration is ill-posed.The paper therefore emphasizes evaluation in clinically relevant settings and assessment of registration uncertainty.
- Clinical setting: Combined dose estimation could support analysis of repeat SRS for brain metastases, but its validity depends directly on registration quality.This setting is clinically relevant because recurrent lesions and toxicity must be assessed across treatments.
- Research gap: Learning-based DIR methods can be fast and strong on established benchmarks, but models trained mainly on healthy-brain MRI may not generalize reliably to clinical scans with enhancing tumors.The paper contrasts benchmark performance with the domain shift represented by longitudinal contrast-enhanced MRI.
- Study objective: The study benchmarks pretrained learning-based and optimization-based DIR methods on longitudinal contrast-enhanced T1-weighted MRI using overlap, surface distance, volume recovery, runtime, and memory.Learning-based methods are evaluated with zero-shot inference, ISO, and TSO.
- Main findings: Healthy-brain-pretrained models perform inconsistently without adaptation, whereas ISO and TSO improve tested learning-based families; optimization-based methods remain most accurate and robust overall.The authors conclude that variation across methods and cases prevents unchecked automatic use in dose accumulation.
2 Methods
The benchmark uses manually screened longitudinal MRI lesion pairs, aligns earlier scans to later scans, and evaluates both learning-based adaptation strategies and optimization-based registration. Accuracy is measured through lesion overlap, boundary agreement, and target-volume recovery, alongside computational cost.
- Data and preprocessing: Each lesion pair contains recurrent-lesion MRI from first and second SRS, with clinician-delineated enhancing tumor contours serving as reference masks.The earlier MRI is warped into the later MRI coordinate system after rigid alignment.
- Data and preprocessing: Pairs were manually screened for intensity-based DIR suitability and excluded when lesion structures or adjacent enhancing vessels changed substantially.Included pairs required sufficient similarity in normal brain anatomy and lesion enhancement.
- Registration formulation: DIR maps the earlier moving image into the later fixed-image space using a deformation parameterized by a displacement field.The deformation is estimated to align corresponding anatomy across time.
- Registration formulation: Optimization-based registration solves a pair-specific objective that balances image dissimilarity against deformation regularization through λ.D measures image dissimilarity, R regularizes the deformation field, and λ controls their tradeoff.
- Learning-based methods: Learning-based registration trains a network to predict displacement fields from image pairs by minimizing expected registration loss over a training distribution.Pretrained inference uses the learned network directly, making performance dependent on similarity between test and training distributions.
- Learning-based methods: Learning-based methods were evaluated in pretrained, instance-specific optimization, and tumor-proximity target-specific optimization modes.TSO emphasizes matching near the recurrent lesion while retaining contribution from the rest of the brain.
- Compared methods: The benchmark includes baseline and recent learning-based models plus conventional and GPU-based optimization methods, including FireANTs and SINR.FireANTs variants also use target-weighted objectives, denoted by the suffix “-TW.”
- Evaluation: Accuracy is assessed by warping the earlier lesion mask and measuring Dice, HD95, and sASD, while volume recovery compares observed and recovered log2 lesion-volume change.A slope near 1 with high correlation indicates better recovery of target-volume change.
3 Results and Discussion
Across the 87-case cohort, optimization-based methods achieved the strongest registration and volume-recovery performance, while learning-based methods benefited from pair-specific adaptation. Accuracy gains came with practical runtime and memory tradeoffs, and method performance varied across cases.
- Learning-based methods: Pretrained learning-based methods showed strongly model-dependent zero-shot performance, whereas ISO and TSO improved all tested learning-based method families.The largest gains occurred for weaker zero-shot models, and TSO did not uniformly outperform ISO.
- Evaluation: Table 1 reports Dice, HD95, sASD, regression slope, and Pearson correlation for registration accuracy and target-volume agreement on 87 cases.HD95 and sASD are lower-is-better, while volume-recovery slope is ranked by closeness to 1 and correlation is higher-is-better.
- Optimization-based methods: Optimization-based methods, particularly FireANTs-SyN and SINR, achieved the best overall registration accuracy.Target weighting further improved FireANTs, especially the Greedy variant.
- Target-volume recovery: Modern optimization-based methods, especially SINR and target-weighted FireANTs, showed the strongest target-volume recovery trends.Pretrained learning-based methods often under-recovered observed lesion volume change, while ISO and TSO improved volume agreement.
- Runtime and memory: Pretrained learning-based inference was fastest, but ISO and TSO substantially increased runtime and GPU memory use.FireANTs combined strong accuracy with relatively modest runtime and memory demand.
4 Conclusion
The benchmark supports evaluating deformable dose accumulation for brain metastasis reirradiation while showing that registration remains under investigation. Pair-specific adaptation improves pretrained learning-based models, but optimization-based methods provide the strongest overall geometric performance with practical runtimes; downstream dose validity remains untested.
- Pair-specific adaptation consistently improves pretrained learning-based models.
- Optimization-based methods, particularly FireANTs and SINR, provide the strongest overall geometric performance while maintaining practical runtimes.
- The study evaluates lesion correspondence only, leaving downstream dose and outcome analyses for future work.Planned extensions include case-level quality control, registration-acceptance criteria, and sensitivity analysis of accumulated-dose metrics to registration uncertainty.