Ultra-Sparse View Reconstruction of CBCT Using an Adversarial Implicit Neural Representation
Abstract
Purpose
In image-guided radiation therapy (IGRT), cone-beam CT (CBCT) reconstruction typically requires 300–600 projections to achieve acceptable image quality, which contributes to imaging dose and acquisition time. This study investigates the feasibility of reconstructing pelvic CBCT volumes from ultra-sparse projections using a patient-specific adversarial implicit neural representation (INR).
Methods
CBCT images (full rotation, pelvic protocol, ~600 projections) from three prostate cancer patients treated with radical IGRT were included. Ultra-sparse projection sets (5, 15, 25, 35, and 45 projections) were simulated from the reference CBCT volumes using known CBCT geometry. For each patient, an INR model was trained to map sparse 2D projections to a 3D CBCT volume. The INR utilized a multi-layer perceptron with sinusoidal (SIREN) activations to predict Hounsfield Unit (HU) values from voxel coordinates and local projection features. A PatchGAN discriminator was incorporated to improve anatomical realism and reduce over-smoothing. The model was optimized using a combination of adversarial loss, projection consistency loss, and total variation regularization. Reconstruction quality was evaluated using mean absolute error (MAE) within the body contour, structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR), and compared with conventional Feldkamp–Davis–Kress (FDK) reconstruction.
Results
The proposed adversarial-INR framework produced good-quality CBCT reconstructions from ultra-sparse projections. For projection counts of 25 and above, the average MAE was below 25 HU, SSIM exceeded 0.92, and PSNR was greater than 32 dB. In contrast, FDK reconstruction showed poor performance even with 45 projections (SSIM ≈ 0.60, PSNR ≈ 23 dB, MAE ≈ 60 HU).
Conclusion
This study demonstrates the feasibility of a patient-specific adversarial INR for ultra-sparse view CBCT reconstruction. The proposed approach substantially reduces the number of required projections while preserving anatomical accuracy, with potential applications in lowering imaging dose and improving IGRT workflows.