2458 - A Tumor Tracking Method with Deformation Image Registration-Enhanced Deep Learning Model in Surface-Guided Radiotherapy

G. Zhang¹, Z. Jiang¹, and L. Wang²; ¹Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, ²Shandong Cancer Hospital and Institute, Jinan, China

Purpose/Objective(s): Respiration-induced tumor deformation and motion lead to omission of prescription dose of partial target and increased additional irradiation risks to healthy tissue during radiotherapy delivery. Real-time tumor tracking can effectively compensate for the impact of respiratory motion. We propose a patient-specific external-internal correlation model driven by optical surface imaging (OSI) in surface-guided radiotherapy (SGRT) for respiration-induced tumor motion and deformation prediction. Materials/

Methods: A retrospective-prospective database was established, enrolling 276 lung cancer patients undergoing 4D-CT simulated localization. Retrospective patients were divided into cohorts for training/cross-validation (Cohort-T-CV) and testing (Cohort-Test), of which body surfaces were extracted from 4D-CT phases to compensate for the limited data volume of paired optical-CT images. Prospective patients consisted of paired optical-CT data for additional validation (Cohort-AddV). Respiration-induced tumor deformation and motion are predicted and concretized in the form of variable 3D mask volumes, with End-Expiratory (EE) and End-Inspiratory (EI) states selected as the starting and ending points of prediction tasks. Deformation image registration (DIR) between surface images was performed after morphological operations to obtain Jacobian determinant map as one of input channels to enhance voxel-wise deformation details for mask inference. Residual-blocks and spatial attention gates were integrated into U-net-based architecture to build the DIR-enhanced model 3D-U-RAD for respiration-induced external-internal correlation. Predictions of 3D-U-RAD and 3D-U-RA (simplified model without DIR-enhancement) were evaluated on different cohorts with absolute/relative deviations of centroid (DC/rDC), Dice similarity coefficient (DSC), 95% Hausdorff-Distance (HD₉₅), and absolute/relative volume changes (dV/rdV).
Results: Amplitude motion prediction errors of 3D-U-RAD are 0.61±0.46mm and 0.59±0.47mm on Cohort-Test and Cohort-AddV, respectively. In deformation prediction, DSC are respectively 0.80±0.04 and 0.81±0.03, HD₉₅ are 4.05±1.25mm and 3.90±1.52mm, and dV are 1.01±0.65cm³ and 1.12±0.64cm³ on the two cohorts, respectively. Except rDC in left-right direction, results of 3D-U-RAD are significantly superior to 3D-U-RA (p<0.05) in all other evaluation indicators.
Conclusion: Driven by OSI, the proposed model has feasibility to facilitate patient-specific accurate, non-radiative, and non-invasive tumor tracking for intra-fractional radiotherapy.

	A: 3D-U-RAD-Test	B: 3D-U-RA-Test	C: 3D-U-RAD-AddV	D: 3D-U-RAD-AddV	p (A vs. B)	p (C vs. D)	p (A vs. C)
Amplitude DC (mm)	0.61	3.2	0.59	3.14	<0.01	<0.01	0.86
Amplitude rDC (%)	0.34	1.07	0.76	1.63	<0.01	<0.01	0.13
DSC	0.80	0.61	0.81	0.65	<0.01	<0.01	0.32
HD95 (mm)	4.05	10.06	3.90	7.91	<0.01	<0.01	0.65
dV (mm3)	1.01	2.63	1.12	2.67	<0.01	<0.01	0.55
rdV (%)	0.34	0.72	0.24	0.53	<0.01	<0.01	0.32