587 - 3D Cerebrovascular Reconstruction from Biplane Projections: A CT Angiography-Based Deep Learning Simulation

Purpose: To reconstruct 3D cerebrovascular anatomy from biplane (anteroposterior and lateral) projection images using a deep learning back-projection algorithm and to evaluate reconstruction accuracy.

Materials and Methods: Sixteen CT angiography (CTA) datasets were divided into training (n=10), validation (n=3), and test (n=3) groups, with three cross-validation configurations. CTA-derived biplane projections served as DSA–like inputs. Different projection stack sizes were evaluated to assess reconstruction accuracy for the cerebral arteries, quantified by the Dice coefficient (DC) and intersection over union (IoU). Region-of-interest (ROI) magnified reconstructions of the left and right cerebral arteries were compared with non-magnified cropped images using DC, IoU, peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). The Wilcoxon signed-rank test was applied for statistical comparisons.

Results: For whole-brain cerebral arterial reconstructions, higher DC and IoU values were observed at stack sizes of 9, 11, and 13 with only marginal differences. Accordingly, a stack size of 9 was selected for subsequent evaluations, yielding a DC of 0.509 and an IoU of 0.344. In ROI-magnified reconstructions, DC (right/left: 0.783/0.765) and IoU (0.644/0.622) were higher than in non-magnified cropped images (0.566/0.554, 0.402/0.392), demonstrating improved accuracy in focal areas. PSNR improved from 26.6 dB/26.9 dB to 30.2 dB/30.2 dB, and SSIM from 0.682/0.680 to 0.892/0.883. Magnified reconstruction showed statistically significant improvement in all parameters (p < 0.01).

Conclusion: This proof-of-concept demonstrates the feasibility of 3D cerebrovascular reconstruction from biplane angiography-like input using deep learning. This approach may reduce radiation exposure and contrast load, and warrants validation in studies using actual high-resolution angiographic data with higher resolution.