Deformable image registration, a way of matching corresponding anatomy between scans, is not yet reliable enough for unmonitored workflows that would accumulate radiation dose across repeat stereotactic radiosurgery for brain metastases, a benchmark study concludes. The strongest geometric results came from optimization-based methods, but the quality of the match still varied from case to case.
The benchmark tested whether an earlier contrast-enhanced T1-weighted MRI could be mapped into the space of a later scan and matched to the later lesion. It measured lesion overlap, surface distance, target-volume recovery, runtime and peak GPU memory.
A screened test from repeat-treatment scans
Researchers began with 97 patients and 171 recurrent lesions in an institutional cohort. Manual screening retained 55 patients and 87 lesion pairs for the benchmark. The retained pairs had MRI at both time points, sufficient anatomy and similar enhancement; cases with major lesion changes or ambiguous boundaries were excluded.
The learning-based group included VoxelMorph, TransMorph, VFA, SITReg and UniGradICON. The optimization-based group included Greedy, ANTs-SyN, FireANTs and SINR.
For each pair, the earlier lesion mask was warped into later-MRI space and compared with the later lesion mask. Dice captured overlap, while HD95, the 95th-percentile Hausdorff distance, and sASD, the symmetric average surface distance, captured separation between the lesion surfaces. The analysis compared recovered and observed log2 target-volume change using regression slope and Pearson correlation, and recorded runtime and peak GPU memory.
Adapting the learning models changed the picture
Used as-is, pretrained learning-based methods produced strongly model-dependent results. SIT performed best among the pretrained models, while UGI generalized poorly to this longitudinal brain-metastasis reirradiation cohort.
The researchers then tried pair-specific adaptation, optimizing network parameters separately for each image pair. Both instance-specific optimization, or ISO, and tumor-proximity target-specific optimization, or TSO, performed better than pretrained inference in the tested learning-based methods. TSO did not consistently outperform ISO.
The adaptation ran for 100 steps per pair using Adam at a learning rate of 10−4. TSO added a 5-mm expansion around the lesion contour, with a minimum spatial weight of 0.1 and a 5-mm distance scale.
Accuracy was not the same as volume recovery
Overall, optimization-based methods were strongest on the geometric measures, especially FireANTs-SyN and SINR.
In the 87-case comparison, FA-SyN-TW had a mean Dice score of 0.681, with a standard deviation of 0.139, and a mean HD95 of 2.38 mm, with a standard deviation of 1.22 mm. SINR reported a mean sASD of 0.66 mm, with a standard deviation of 0.36 mm.
Volume tracking told a slightly different story from outline matching. Pretrained methods often under-recovered observed lesion-volume change. ISO and TSO improved volume agreement, while SINR and target-weighted FireANTs showed the strongest volume-recovery trends.
Speed mattered too
Pretrained inference was fastest. VoxelMorph PT averaged 0.23 seconds, while the VoxelMorph ISO/TSO or optimization entry averaged 94 seconds. FireANTs Greedy averaged 13 seconds and 2.6 GB of peak GPU memory; FireANTs SyN averaged 20 seconds and 4.2 GB.
Pair-specific adaptation substantially increased runtime and GPU-memory use. FireANTs combined strong registration performance with relatively modest resource demands in the reported comparisons.
The evidence stops short of hands-off use
The authors' clinical conclusion was cautious. Current state-of-the-art DIR methods were not sufficiently accurate and consistent for unmonitored use in dose-accumulation workflows, and performance remained variable at the case level.
The evidence comes from a manually screened benchmark of 87 lesion pairs, with MRI anatomy and enhancement similarity used for inclusion and major lesion changes or ambiguous boundaries excluded. The findings therefore apply most directly to the tested cohort.
The study was funded by NIH R44CA183390, and the authors declared no competing interests relevant to the article's content. It is an arXiv preprint, version 1, dated 27 August 2026.
Paper data and sources
Original title: Is Deformable Image Registration Ready for Brain Metastasis Reirradiation Dose Accumulation? A Longitudinal MRI Benchmark of Registration Accuracy
Authors: Hengjie Liu, Manju Sharma, Xinyi Fu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-27
DOI: Not available
Original paper · Full text