1029 - Integrating Laser Surface Mapping with CBCT for Enhanced Decision Support from ML Guided Resimulation Predictions

S. Shang¹, G. Evans¹, M. Kligerman¹, M. Shang², T. R. Williams¹, and C. Y. Shang¹; ¹South Florida Proton Therapy Institute, Delray Beach, FL, ²Orlando Health Cancer Institute, Orlando, FL

Purpose/Objective(s): The studys primary objective is to evaluate the comparability of C-RAD Surface Mapping Images (SMI) scores with traditional CT Surface Mapping (CTSM) scores in prostate proton therapy. It aims to determine if SMI scores can be effectively integrated into machine learning models to enhance the predictive value for resimulation decisions, thereby improving the precision of dose delivery and the overall decision-making process in Intensity-Modulated Proton Therapy (IMPT). Materials/

Methods: A retrospective analysis was conducted on 38 prostate cancer patients undergoing IMPT. The dataset comprised 1767 CTSM score sets from 2606 beams in patients requiring resimulation and 4212 beams from 109 patients without resimulation needs. C-RAD surface point clouds, sourced from the C-RAD software when available, were reoriented to beam eye view coordinates, and underwent poisson surface reconstruction similarly to CBCT and Simulation CT point clouds. Large SMI surface deviations and artifacts were excluded before correlation studies with CTSM Scores. Patient and treatment session Metadata were obtained manually or via SQL from the Varian database. Feature preprocessing and engineering involved Autoencoder, and custom python scripts. Machine learning algorithms, particularly Random Forest, were trained on 70% of the data with the remaining 30% used for downstream analysis.
Results: The investigation revealed a median geometric discrepancy of 1.83 mm between C-RAD Surface Mapping Image (SMI) and Cone-Beam Computed Tomography (CBCT) SM reconstructed surfaces. SMI demonstrated a closer alignment with CBCT surfaces compared to simulation CT surfaces, with a Pearson correlation coefficient of 0.961 (p-value < 0.01). Following median correction, a robust correlation was observed between SMI scores and CTSM scores across all evaluated percentile scores (p<0.001, Table 1), suggesting that SMI scores, with appropriate correction may complement CTSM scores in machine learning models. Integration of CRAD SMI scores into the CTSM and metadata machine learning framework enhanced model performance over using CTSM scores alone, achieving an area under the curve of 0.93 in resimulation prediction.
Conclusion: Daily laser surface guidance using C-RAD SMI and CBCT in the initial alignment phase of IMPT correlate well with each other, potentiating objective estimation of patient body changes before CBCT acquisition. Supported by machine learning, this approach provides an objective method for integrating laser surface guidance with CBCT images to make informed resimulation decisions, potentially enhancing patient care within proton therapy. Table 1. Representative CTSM vs SMI Score Correlation Statistics

CTSM (vs SimulationCT) Score Percentile	SMI (vs SimulationCT) Score Percentile	Correlation Coefficient(r)	p-Value (Pearson Correlation)	R-Squared Value
75%	75%	0.817	p<0.001	0.414
70%	75%	0.82	p<0.001	0.339
90%	85%	0.792	p<0.001	0.446
90%	90%	0.783	p<0.001	0.28