SinoSynth: A Physics-based Domain Randomization Approach for Generalizable CBCT Image Enhancement

TL;DR

This paper tackles the challenge of restoring CBCT image quality across diverse scanners and protocols, where artifacts and noise hinder diagnostic reliability. It introduces SinoSynth, a physics-based CBCT degradation model that converts high-quality CT images into synthetic CBCT data by simulating artifacts in the sinogram domain with randomized cone-beam geometry, enabling unlimited aligned CBCT-CT training pairs. The authors couple SinoSynth with multiple generative frameworks and introduce structure-guided losses to preserve anatomy, achieving superior generalization on multi-institutional CBCT datasets compared with models trained on actual CBCT-CT pairs. This approach demonstrates that CT-only data, when processed through realistic, domain-randomized CBCT simulations, can yield robust CBCT enhancement, potentially reducing data-collection burdens and improving clinical reliability across settings.

Abstract

Cone Beam Computed Tomography (CBCT) finds diverse applications in medicine. Ensuring high image quality in CBCT scans is essential for accurate diagnosis and treatment delivery. Yet, the susceptibility of CBCT images to noise and artifacts undermines both their usefulness and reliability. Existing methods typically address CBCT artifacts through image-to-image translation approaches. These methods, however, are limited by the artifact types present in the training data, which may not cover the complete spectrum of CBCT degradations stemming from variations in imaging protocols. Gathering additional data to encompass all possible scenarios can often pose a challenge. To address this, we present SinoSynth, a physics-based degradation model that simulates various CBCT-specific artifacts to generate a diverse set of synthetic CBCT images from high-quality CT images without requiring pre-aligned data. Through extensive experiments, we demonstrate that several different generative networks trained on our synthesized data achieve remarkable results on heterogeneous multi-institutional datasets, outperforming even the same networks trained on actual data. We further show that our degradation model conveniently provides an avenue to enforce anatomical constraints in conditional generative models, yielding high-quality and structure-preserving synthetic CT images.

Figures (6)

• Figure 1: (a) Comparisons between ours and other CBCT data augmentation methods. (b) Our method significantly improves the denoising performance of CycleGAN without requiring pre-aligned paired datasets.
• Figure 1: Quantitative results comparing networks trained with actual CBCT-CT data and our simulated CBCT-CT data. Our method significantly improves the denoising and CBCT image enhancement performance of the networks.
• Figure 2: Overview of the proposed framework. (a) A planning CT volume is first transformed into the 3D sinogram. (b) Simulated CBCT-specific artifacts are randomly applied to the sinogram, which are then transformed back to a CBCT-quality volume using the FDK algorithm. (c) The simulated CBCT is fed into the generative network trained with the proposed sinogram and structure consistency constraints. (d) During testing, the network takes the actual CBCT as input and outputs a denoised synthetic CT image.
• Figure 2: Quantitative evaluations of different data augmentation methods applied to CycleGAN. Our method significantly improves the denoising and CBCT image enhancement performance of the base model.
• Figure 3: Qualitative comparisons between networks trained with actual CBCT-CT data and our simulated CBCT-CT data.