Benchmarking and Optimizing Deformable Image Registration Methods for Brain Metastasis Retreatment Dose Accumulation
Abstract
Purpose
Stereotactic radiosurgery (SRS) reirradiation for locally recurrent brain metastases (BM) utilizes rigid image registration (RIR) for cumulative dose estimation to balance dose escalation against the risk of adverse radiation effects such as radionecrosis. This work aims to benchmark and optimize deformable image registration (DIR) methods to account for local brain deformation and its impact on retreatment dose accumulation.
Methods
Two optimization-based and five learning-based DIR methods (pretrained on the LUMIR brain MRI dataset) were evaluated on an institutional cohort of 29 recurrent lesions from 13 patients. To address domain shift (altered contrast and pathology), learning-based models were adapted using instance-specific optimization (ISO) and lesion-specific optimization (LSO). ISO finetunes on the entire image pair, while LSO focuses on lesion-specific ROI. Registration accuracy was quantified using lesion-based overlap/surface metrics and a proposed volumetric agreement metric. DIR- and RIR-based dose accumulation were compared via accumulated dose maps and DVH metrics, including V12.
Results
ISO/LSO facilitated the generalization of learning-based DIR to unseen data, with LSO yielding the most significant improvement. Compared with RIR, SITReg (LSO) achieved the best lesion mapping (Dice: RIR 0.456±0.162 → 0.694±0.135; HD95 (mm): RIR 4.28±2.42 → 2.15±1.30; sASD (mm): RIR 1.70±1.26 → 0.61±0.27) and volumetric agreement. Compared with RIR, DIR-based dose accumulation produced non-trivial differences in accumulated dose distributions. The downstream DVH impact was case-dependent. Eight out of twenty-nine lesions (28%) had |ΔV12| > 5% (range -12.0% to 21.6%).
Conclusion
Although RIR is commonly used for intracranial dose accumulation, DIR can produce meaningful, case-specific differences in accumulated dose and V12, indicating DIR can materially alter toxicity-relevant cumulative dose metrics in BM retreatment. Future work is planned to validate these findings in larger cohorts and test whether DIR-based dose accumulation improves prediction of toxicity endpoints and incorporating time-dependent effects.