Optical surface imaging-driven tumor tracking with deformable image registration-enhanced deep learning model for surface-guided radiotherapy

Abstract

We propose a patient-specific external-internal correlation model driven by optical surface imaging (OSI) for intra-fractional respiration-induced tumor motion and deformation tracking. A retrospective-prospective database was established enrolling 276 lung cancer patients undergoing 4D-CT imaging. Retrospective patients were divided into cohorts for training/cross-validation (Cohort-T-CV) and testing (Cohort-Test), whose body surfaces were extracted from 4D-CT phases to compensate for limited data volume of paired OSI-CT images. Prospective patients consisted of paired data for additional validation (Cohort-Add-V). Respiration-induced tumor motion and deformation are predicted in form of deformable 3D masks, with different phases as starting and ending points of prediction task. Deformable image registration (DIR) was performed 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 DIR-enhanced model 3D-U-RAD for external-internal correlation. Predictions of 3D-U-RAD and 3D-U-RA (simplified without DIR-enhancement) were evaluated with absolute/relative deviations of centroid ($\mathrm{DC}$/$\mathrm{rDC}$), Dice similarity coefficient ($\mathrm{DSC}$), 95 % Hausdorff-Distance (${\mathrm{HD}}_{95}$), and absolute/relative volume changes ($\delta V$/$r\delta V$). Amplitude motion prediction errors of 3D-U-RAD are 0.61 ± 0.46 mm and 0.59 ± 0.47 mm on Cohort-Test and Cohort-Add-V, respectively. In deformation prediction,$\mathrm{DSC}$ are respectively 0.80 ± 0.04 and 0.81 ± 0.03, ${\mathrm{HD}}_{95}$ are 4.05 ± 1.25 mm and 3.90 ± 1.52 mm, and $\delta V$ are 1.01 ± 0.65 cm³ and 1.12 ± 0.64 cm³ on the two cohorts, respectively. Except $\mathrm{rDC}$ in left–right direction, results of 3D-U-RAD are significantly superior to 3D-U-RA ($p$ < 0.05) in all other evaluation indicators. Driven by OSI, the proposed framework has feasibility to facilitate patient-specific accurate, non-radiative, and non-invasive tumor tracking for intra-fractional radiotherapy.