3064 - Impact of Contour-Guided Deformable Image Registration on Normal Brain Dose Accumulation for Recurrent Brain Metastasis Reirradiation

Purpose/Objective(s):

Cumulative normal-brain V12 is commonly used for toxicity risk assessment after stereotactic radiosurgery (SRS) retreatment. Since high-dose regions are anchored to the target, rigid or intensity-only deformable registration can underestimate lesion-centric deformation and bias dose accumulation. We tested whether contour-guided deformable registration results in systematic changes in accumulated V12 relative to rigid and intensity-only deformable registration.

Materials/Methods:

We retrospectively analyzed 30 recurrent brain metastasis lesions (15 patients) with paired T1 contrast-enhanced MRI. Initial-to-retreatment registration was performed using rigid and 8 SOTA deformable methods: 3 optimization-based (ANTs, Greedy, GPU-accelerated FireANTs) and 5 deep learning (DL)-based (VoxelMorph, TransMorph, UniGradICON, VFA, SITReg). DL methods were pretrained on a large scale T1 MRI dataset (LUMIR) and finetuned (ft) with expanded lesion mask. For FireANTs, we further tested contour-guided (CG) DIR using (1-a)·Lsim + a·Lseg, where Lsim is similarity loss and Lseg enforces target contour agreement. a was carefully tuned to be 0.7 to balance accuracy and deformation plausibility.

Target mapping accuracy was assessed by Dice, HD95, and volumetric agreement (how accurately the deformation reflects target shrinkage/expansion; ideal=1). Initial dose was warped to retreatment space for dose accumulation. We report dV12% = (V12_DIR/V12_RIR-1), including its range and |dV12%|>10% rate.

Results:

Table 1 lists the best-performing methods. The best DL method, SITReg (ft), did not exceed the best optimization-based (Greedy/Fire-ANTs) in Dice/HD95. GPU-accelerated Fire-ANTs ran <5 s/pair (vs ~5 min conventional optimization, ~2 min DL finetuning). Fire-ANTs with contour guidance markedly improved contour and volumetric agreement and increased accumulated V12 differences vs rigid (dV12% range -21.8% to +64.8%; |dV12%|>10%: 40%). Contour guidance strengthened the dV12%–target volume-change association relative to intensity-only DIR. The exact relation needs validation in a larger cohort.

Conclusion:

Adding contour guidance matters for retreatment dose mapping as intensity-only registration can underestimate lesion-centric deformation and bias accumulated V12. We will expand the cohort and test whether proposed dose accumulation improves toxicity prediction. Pretrained DL-DIR (even with finetuning) did not outperform optimization-based methods, while FireANTs shows strong accuracy with seconds-level runtime, making it well suited for clinical development and validation.

	Dice	HD95	Volumetric agreement	range of dV12%	|dV12%| > 10%
Rigid	0.45±0.16	4.89±4.00	-	-	-
SITReg	0.64±0.16	3.32±4.07	0.55	-13.5%, 19.0%	16.7%
SITReg (ft)	0.69±0.14	2.68±3.17	0.68	-12.0%, 21.6%	10.0%
Greedy	0.69±0.15	2.40±1.85	0.64	-14.9%, 35.2%	20.0%
Fire-ANTs	0.69±0.16	2.67±2.35	0.69	-16.6%, 30.4%	23.3%
Fire-ANTs (CG)	0.89±0.04	1.01±0.27	0.96	-21.8%, 64.8%	40.0%