Author: Chloe Min Seo Choi, Joseph O. Deasy, Jue Jiang, Sudharsan Madhavan, Nikhil Mankuzhy, Nishant Nadkarni, Andreas Rimner, Maria Thor, Harini Veeraraghavan, Abraham Wu 👨🔬
Affiliation: University of Freibrug, Department of Medical Physics, Memorial Sloan Kettering Cancer Center, Memorial Sloan Kettering Cancer Center 🌍
Purpose: Voxel-based analysis (VBA) requires accurate topology-preserving inter-patient deformable image registration (DIR). This study assessed whether guiding a DIR method with geometric priors of tumors in moving and reference images improves tumor and dose preservation in lung cancer patients.
Methods: Two trained deep learning (DL) methods—1) Tumor-Aware Recurrent Registration (TRACER), incorporating tumor guidance, and 2) Patient-Specific Context and Shape (PACS)—along with an iterative intensity-based B-spline method (SyN), were tested on 240 patients with stage III non-small cell lung cancer. Geometric accuracy was assessed using Dice Similarity Score (DSC). Tumor shrinkage before and after DIR was measured for each model. Dosimetric impact on organs-at-risk (OARs) was evaluated by comparing mean doses pre- and post-DIR. Tumor dose preservation was quantified using the dose preservation ratio (DPR = V60_moving_afterDIR / V60_moving), where DPR ≈ 1 indicates good preservation of the high-dose region. Correlations between registration accuracy, body mass index (BMI), and gender were analyzed using Pearson correlation.
Results: TRACER outperformed PACS and SyN in preserving OAR and tumor doses, with significantly smaller mean dose differences in 10 of 12 OARs and lower tumor volume differences after DIR (1.39 ± 1.97%, p<0.001) compared to PACS (26.83%) and SyN (29.04%). TRACER showed the most stable tumor dose preservation with a DPR of 0.97, closer to 1 than PACS (1.08, p<0.001) or SyN (0.78, p<0.001). PACS demonstrated higher organ geometry matching for several organs compared to TRACER and SyN. None of the methods showed gender-specific deviations. Geometric accuracy for tubular organs is weakly to moderately correlated with BMI for TRACER and SyN.
Conclusion: TRACER preserved tumor volume and minimized OAR dose impact, balancing accuracy and tumor integrity. The tradeoff between geometric accuracy and tumor preservation underscores the importance of careful method selection for inter-patient registration applications.