Real-Time Volumetric Imaging In Interventional Radiology Suites: A Digital Phantom Study
Abstract
Purpose
In interventional radiology, the precise 3D localization of mobile targets, particularly within the thorax, is of utmost importance. While conventional modalities such as Cone Beam CT (CBCT) provide necessary depth information, they are unsuitable for real-time dynamic tracking of internal anatomy. This study evaluates Voxelmap, a deep learning framework, for real-time 3D motion tracking and volumetric imaging using the XCAT digital phantom.
Methods
Voxelmap predicts 3D deformation vector fields (DVFs) by learning patient-specific mappings between 2D X-ray projections and 3D volumetric anatomy. Eight unique patient anatomies were simulated using the XCAT phantom. To assess geometric robustness within clinically relevant limits, a neural network was evaluated across five C-arm tilt angles (-30° to +30°, with 15° increments). Reconstruction accuracy was quantified using Structural Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). Statistical significance was assessed via the Friedman test.
Results
Voxelmap demonstrated high volumetric fidelity across all configurations, with a population mean SSIM of 0.987 +/- 0.008 and PSNR of 36.7 +/- 6.2 dB Performance was highly stable across the full 360° gantry rotation arc, with minimal degradation even at high-attenuation lateral views. While C-arm tilt significantly influenced metrics (p < 0.001), accuracy remained clinically robust, peaking at a mean of 0.988 ± 0.003 for -30° tilt. Individual patient anatomy had a more pronounced impact on performance than tilt or gantry angle (p < 0.001), with mean SSIM ranging from 0.982 to 0.990 across the cohort. Accuracy was phase-dependent, peaking at peak-exhalation and declining towards peak-inhalation.
Conclusion
Voxelmap provides a robust, generalized framework for real-time volumetric imaging in interventional settings. The high reconstruction fidelity across extreme C-arm tilts and full gantry rotation confirms its flexibility for diverse clinical geometries without specialized hardware, potentially facilitating 3D guidance in the interventional suite.