Automated Deformable Registration and Three-dimensional Margin Assessment for Predicting Local Recurrence after Lung...

Purpose To develop a lung-specific deformable image registration algorithm optimized for lung thermal ablation and evaluate whether three-dimensional (3D) margin assessment predicts time to local recurrence. Materials and Methods This institutional review board-approved, single-institution retrospective study evaluated patients who underwent lung thermal ablation with available pre- and postprocedural CT scans suitable for deformable registration. Images were preprocessed with segmentation of tumor, ablation zone, and lung. A four-stage deformable image registration framework was applied: (a) affine registration, (b) deformable image registration to the cropped lung, (c) lung mask-guided deformable image registration, and (d) local deformable image registration focused on the neighborhood adjacent to the ablation zone. Registrations were performed using free-form B-spline transformations with cost function masking of the ablation zone. Registration accuracy was assessed using target registration error (TRE). The 3D ablation margins were quantified using a distance-transform-based analysis of the spatial relationship between the tumor surface and ablation zone boundary. Associations between margin size and time to local recurrence were evaluated using competing-risks regression, and time-dependent receiver operating characteristic analysis was performed. Results A total of 69 patients (median age, 59 years [IQR, 50-69 years]; 38 female) with 108 ablated lung tumors were included. Mean TRE ± SD was 0.4 mm ± 0.3 and mean ablation margin was 1.6 mm ± 2.1. Larger margins were associated with longer time to local recurrence (subdistribution hazard ratio, 0.5 per millimeter increase [95% CI: 0.4, 0.6]; P < .001) and remained independently associated on multivariate analysis. Using a 2-mm margin threshold, the 2-year local recurrence rate was 3% (95% CI: 1, 8). The area under the receiver operating characteristic curve for predicting 2-year local recurrence was 0.86. Conclusion The four-step lung-optimized deformable image registration framework enabled accurate automated 3D tumor ablation margin quantification, and margin size was associated with time to local recurrence. Keywords: Ablation Techniques, Interventional-Oncology, Percutaneous, Thorax, Lung, Computer Applications-3D, Computational Studies, CT © RSNA, 2026.